Micro_Activity_Without_Masking_Classifier.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "1jPihmLbUvoa"
   },
   "source": [
    "# Function Definitions (unchanged)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "NWzF5-ugUvoc"
   },
   "outputs": [],
   "source": [
    "import glob\n",
    "import os\n",
    "import numpy as np\n",
    "import csv\n",
    "import cv2\n",
    "import io\n",
    "from sklearn.preprocessing import StandardScaler\n",
    "import shutil\n",
    "\n",
    "os.environ[\"CUDA_DEVICE_ORDER\"] = \"PCI_BUS_ID\"\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Options to set for GPU machine"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# from keras.backend.tensorflow_backend import set_session\n",
    "# import tensorflow as tf\n",
    "# config = tf.ConfigProto()\n",
    "# config.gpu_options.allow_growth = True  # dynamically grow the memory used on the GPU\n",
    "# config.log_device_placement = True  # to log device placement (on which device the operation ran)\n",
    "# sess = tf.Session(config=config)\n",
    "# set_session(sess)  # set this TensorFlow session as the default"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "q87cb-fJUvoo"
   },
   "outputs": [],
   "source": [
    "sub_dirs = ['left_hip','right_arm','right_wrist'] #three sensors, 9 channels total"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "ho3e4qQQUvor"
   },
   "outputs": [],
   "source": [
    "def parse_IMU_files(parent_dir, sub_dirs, startTime, endTime, file_name, window_length):\n",
    "    \n",
    "    data = []\n",
    "    \n",
    "    data_count = 0\n",
    "    \n",
    "    for sub_dir in sub_dirs:\n",
    "        channel=[]\n",
    "        \n",
    "        for fn in glob.glob(os.path.join(parent_dir,sub_dir, file_name)):\n",
    "            file = open(fn, newline='')\n",
    "            reader = csv.reader(file)\n",
    "            first = True\n",
    "            count = 0\n",
    "            for row in reader:\n",
    "                \n",
    "                if first:\n",
    "                    first = False\n",
    "                    continue\n",
    "                \n",
    "                timestamp=float(row[3]) #4th column is timestamp\n",
    "                if timestamp >=startTime and timestamp <=endTime and count<window_length:\n",
    "                    \n",
    "                    channel.append([float(row[0]),float(row[1]),float(row[2])])\n",
    "                    count = count + 1 \n",
    "                    data_count = data_count+1\n",
    "                    \n",
    "        data.append(channel)         \n",
    "    return data, data_count\n",
    "\n",
    "import random\n",
    "\n",
    "def parse_IMU_files_2(parent_dir, sub_dirs, startTime, endTime, file_name, window_length):\n",
    "    \n",
    "    data = []\n",
    "    \n",
    "    data_count = 0\n",
    "    for sub_dir in sub_dirs:\n",
    "        channel=[]\n",
    "        \n",
    "        \n",
    "        for fn in glob.glob(os.path.join(parent_dir,sub_dir, file_name)):\n",
    "            file = open(fn, newline='')\n",
    "            reader = csv.reader(file)\n",
    "            first = True\n",
    "            count = 0\n",
    "            for row in reader:\n",
    "                \n",
    "                if first:\n",
    "                    first = False\n",
    "                    continue\n",
    "                \n",
    "                timestamp=float(row[3]) #4th column is timestamp\n",
    "                \n",
    "                window_jitter1 = random.randint(-150,150)\n",
    "                \n",
    "                window_jitter2 = random.randint(-150,150)\n",
    "        \n",
    "                if timestamp >=(startTime+window_jitter1) and timestamp <=(endTime+window_jitter2) and count<window_length:\n",
    "                    \n",
    "                    channel.append([float(row[0]),float(row[1]),float(row[2])])\n",
    "                    count = count + 1  \n",
    "                    \n",
    "                    data_count = data_count+1\n",
    "                    \n",
    "        data.append(channel)         \n",
    "    return data, data_count\n",
    "\n",
    "def parse_IMU_files_3(parent_dir, sub_dirs, startTime, endTime, file_name, window_length):\n",
    "    \n",
    "    data = []\n",
    "    \n",
    "    data_count = 0\n",
    "    for sub_dir in sub_dirs:\n",
    "        channel=[]\n",
    "                \n",
    "        window_jitter1 = random.randint(-1500,1500)\n",
    "        window_jitter2 = random.randint(-1500,1500)\n",
    "            \n",
    "        for fn in glob.glob(os.path.join(parent_dir,sub_dir, file_name)):\n",
    "            file = open(fn, newline='')\n",
    "            reader = csv.reader(file)\n",
    "            first = True\n",
    "            count = 0\n",
    "            \n",
    "            for row in reader:\n",
    "                \n",
    "                if first:\n",
    "                    first = False\n",
    "                    continue\n",
    "                \n",
    "                timestamp=float(row[3]) #4th column is timestamp\n",
    "    \n",
    "                if timestamp >=(startTime+window_jitter1) and timestamp <=(endTime+window_jitter2) and count<window_length:\n",
    "                    \n",
    "                    channel.append([float(row[0]),float(row[1]),float(row[2])])\n",
    "                    count = count + 1  \n",
    "                    \n",
    "                    data_count = data_count+1\n",
    "                    \n",
    "        data.append(channel)         \n",
    "    return data, data_count"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_min_count = 100 # per sample, valid points"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "DLINn0siUvou"
   },
   "outputs": [],
   "source": [
    "def get_train_data(data_dir, sub_dirs, pos_lab, neg_lab, labeltable):\n",
    "    files = os.listdir(data_dir+'/left_hip')\n",
    "    number_of_samples = 500\n",
    "    \n",
    "    labels_macro =dict()\n",
    "    \n",
    "    labels_macro[pos_lab] = 0\n",
    "    labels_macro[neg_lab] = 1\n",
    "    \n",
    "    \n",
    "    #read the labels\n",
    "    labels_loc = labeltable\n",
    "    file_label = open(labels_loc, newline='')\n",
    "    label_reader = csv.reader(file_label)\n",
    "    file_label_mapping = dict()\n",
    "    \n",
    "    first = True\n",
    "    for row in label_reader:\n",
    "        \n",
    "        if first:\n",
    "            first = False\n",
    "            continue\n",
    "        \n",
    "        file_label_mapping[row[0]+'.csv'] = labels_macro[row[2]]\n",
    "        \n",
    "    all_data = []\n",
    "    all_labels = []\n",
    "    all_filename = []\n",
    "    for f in files:\n",
    "        \n",
    "        st_index = 0\n",
    "        end_index = 30000\n",
    "        step = 1000 #overlapping window, step: 1000. \n",
    "        window_index = 10000 #6 second window\n",
    "        \n",
    "        #print('reading file:',f)\n",
    "        f_name = f\n",
    "        \n",
    "        if f_name == '.DS_Store':\n",
    "            continue\n",
    "        \n",
    "        curr_label_file = file_label_mapping[f_name]\n",
    "        \n",
    "        while st_index+step < end_index:\n",
    "        \n",
    "            data, data_count = parse_IMU_files(data_dir, sub_dirs, st_index, st_index+window_index,  f, number_of_samples)\n",
    "            st_index = st_index+step\n",
    "            \n",
    "            if data_count<data_min_count:\n",
    "                continue\n",
    "            \n",
    "            train_data_sample  = np.zeros((9, number_of_samples))\n",
    "            train_data_label   = curr_label_file\n",
    "            for i in range(len(data)):\n",
    "                for j in range(len(data[i])):\n",
    "                    train_data_sample[i*3,j]=data[i][j][0]\n",
    "                    train_data_sample[i*3+1,j]=data[i][j][1]\n",
    "                    train_data_sample[i*3+2,j]=data[i][j][2]\n",
    "            all_filename.append(f_name)\n",
    "            all_data.append(train_data_sample)\n",
    "            all_labels.append(train_data_label)\n",
    "            \n",
    "            \n",
    "            \n",
    "    return all_data, all_labels, all_filename\n",
    "\n",
    "\n",
    "def get_train_data_2(data_dir, sub_dirs,pos_lab, neg_lab, labeltable):\n",
    "    files = os.listdir(data_dir+'/left_hip')\n",
    "    number_of_samples = 500\n",
    "    \n",
    "    labels_macro =dict()\n",
    "    \n",
    "    labels_macro[pos_lab] = 0\n",
    "    labels_macro[neg_lab] = 1\n",
    "    \n",
    "    \n",
    "    #read the labels\n",
    "    labels_loc = labeltable\n",
    "    file_label = open(labels_loc, newline='')\n",
    "    label_reader = csv.reader(file_label)\n",
    "    file_label_mapping = dict()\n",
    "    \n",
    "    first = True\n",
    "    for row in label_reader:\n",
    "        \n",
    "        if first:\n",
    "            first = False\n",
    "            continue\n",
    "        \n",
    "        file_label_mapping[row[0]+'.csv'] = labels_macro[row[2]]\n",
    "        \n",
    "    all_data = []\n",
    "    all_labels = []\n",
    "    \n",
    "    all_counts = []\n",
    "    \n",
    "    for f in files:\n",
    "        \n",
    "        st_index = 0\n",
    "        end_index = 30000\n",
    "        step = 1000 #overlapping window, step \n",
    "        window_index = 10000 #6 second window\n",
    "        \n",
    "        #print('reading file:',f)\n",
    "        f_name = f\n",
    "        \n",
    "        if f_name == '.DS_Store':\n",
    "            continue\n",
    "        \n",
    "        curr_label_file = file_label_mapping[f_name]\n",
    "        \n",
    "        while st_index+step < end_index:\n",
    "        \n",
    "            data, data_count = parse_IMU_files_2(data_dir, sub_dirs, st_index, st_index+window_index,  f, number_of_samples)\n",
    "            \n",
    "            st_index = st_index+step\n",
    "            \n",
    "            if data_count<data_min_count:\n",
    "                continue\n",
    "            \n",
    "            train_data_sample  = np.zeros((9, number_of_samples))\n",
    "            train_data_label   = curr_label_file\n",
    "            for i in range(len(data)):\n",
    "                for j in range(len(data[i])):\n",
    "                    train_data_sample[i*3,j]=data[i][j][0]\n",
    "                    train_data_sample[i*3+1,j]=data[i][j][1]\n",
    "                    train_data_sample[i*3+2,j]=data[i][j][2]\n",
    "            \n",
    "            all_data.append(train_data_sample)\n",
    "            all_labels.append(train_data_label)\n",
    "            \n",
    "            all_counts.append(data_count)\n",
    "            \n",
    "            \n",
    "            \n",
    "    return all_data, all_labels, all_counts\n",
    "\n",
    "\n",
    "def get_train_data_3(data_dir, sub_dirs,pos_lab, neg_lab, labeltable):\n",
    "    files = os.listdir(data_dir+'/left_hip')\n",
    "    number_of_samples = 500\n",
    "    \n",
    "    labels_macro =dict()\n",
    "    \n",
    "    labels_macro[pos_lab] = 0\n",
    "    labels_macro[neg_lab] = 1\n",
    "    \n",
    "    \n",
    "    #read the labels\n",
    "    labels_loc = labeltable\n",
    "    file_label = open(labels_loc, newline='')\n",
    "    label_reader = csv.reader(file_label)\n",
    "    file_label_mapping = dict()\n",
    "    \n",
    "    first = True\n",
    "    for row in label_reader:\n",
    "        \n",
    "        if first:\n",
    "            first = False\n",
    "            continue\n",
    "        \n",
    "        file_label_mapping[row[0]+'.csv'] = labels_macro[row[2]]\n",
    "        \n",
    "    all_data = []\n",
    "    all_labels = []\n",
    "    \n",
    "    all_counts = []\n",
    "    for f in files:\n",
    "        \n",
    "        st_index = 0\n",
    "        end_index = 30000\n",
    "        step = 1000 #overlapping window, step: 1000. \n",
    "        window_index = 10000 #6 second window\n",
    "        \n",
    "        #print('reading file:',f)\n",
    "        f_name = f\n",
    "        \n",
    "        if f_name == '.DS_Store':\n",
    "            continue\n",
    "        \n",
    "        curr_label_file = file_label_mapping[f_name]\n",
    "        \n",
    "        while st_index+step < end_index:\n",
    "                \n",
    "            \n",
    "            data, data_count = parse_IMU_files_3(data_dir, sub_dirs, st_index, st_index+window_index,  f, number_of_samples)\n",
    "            \n",
    "            st_index = st_index+step\n",
    "            \n",
    "            if data_count<data_min_count:\n",
    "                continue\n",
    "            \n",
    "            train_data_sample  = np.zeros((9, number_of_samples))\n",
    "            train_data_label   = curr_label_file\n",
    "            for i in range(len(data)):\n",
    "                for j in range(len(data[i])):\n",
    "                    train_data_sample[i*3,j]=data[i][j][0]\n",
    "                    train_data_sample[i*3+1,j]=data[i][j][1]\n",
    "                    train_data_sample[i*3+2,j]=data[i][j][2]\n",
    "            \n",
    "\n",
    "            all_data.append(train_data_sample)\n",
    "            all_labels.append(train_data_label)\n",
    "            \n",
    "            all_counts.append(data_count)\n",
    "            \n",
    "            \n",
    "            \n",
    "    return all_data, all_labels, all_counts"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_prediction_accuracy(data_dir, sub_dirs,pos_lab,neg_lab,labeltable,model):\n",
    "\n",
    "    files = os.listdir(data_dir+'/left_hip')\n",
    "    number_of_samples = 500\n",
    "\n",
    "    labels_macro =dict()\n",
    "\n",
    "    labels_macro[pos_lab] = 0\n",
    "    labels_macro[neg_lab] = 1\n",
    "\n",
    "\n",
    "    #read the labels\n",
    "    labels_loc = labeltable\n",
    "    file_label = open(labels_loc, newline='')\n",
    "    label_reader = csv.reader(file_label)\n",
    "    file_label_mapping = dict()\n",
    "\n",
    "    first = True\n",
    "    for row in label_reader:\n",
    "\n",
    "        if first:\n",
    "            first = False\n",
    "            continue\n",
    "\n",
    "        file_label_mapping[row[0]+'.csv'] = labels_macro[row[2]]\n",
    "\n",
    "\n",
    "    total_count = 0\n",
    "    correct_count = 0\n",
    "\n",
    "    for f in files:\n",
    "\n",
    "        file_data = []\n",
    "        file_label = []\n",
    "\n",
    "        st_index = 0\n",
    "        end_index = 30000\n",
    "        step = 1000 #overlapping window, step: 1000. \n",
    "        window_index = 10000 #6 second window\n",
    "\n",
    "        #print('reading file:',f)\n",
    "        f_name = f\n",
    "\n",
    "        if f_name == '.DS_Store':\n",
    "            continue\n",
    "\n",
    "        total_count = total_count+1\n",
    "\n",
    "        curr_label_file = file_label_mapping[f_name]\n",
    "\n",
    "        while st_index+step < end_index:\n",
    "\n",
    "            data, data_count = parse_IMU_files(data_dir, sub_dirs, st_index, st_index+window_index,  f, number_of_samples)\n",
    "            st_index = st_index+step\n",
    "\n",
    "            if data_count<data_min_count:\n",
    "                continue\n",
    "\n",
    "            train_data_sample  = np.zeros((9, number_of_samples))\n",
    "            train_data_label   = curr_label_file\n",
    "            for i in range(len(data)):\n",
    "                for j in range(len(data[i])):\n",
    "                    train_data_sample[i*3,j]=data[i][j][0]\n",
    "                    train_data_sample[i*3+1,j]=data[i][j][1]\n",
    "                    train_data_sample[i*3+2,j]=data[i][j][2]\n",
    "\n",
    "            file_data.append(train_data_sample)\n",
    "            #file_label.append(train_data_label)\n",
    "\n",
    "        file_data = np.array(file_data)\n",
    "        file_label = curr_label_file\n",
    "\n",
    "        file_data = file_data.reshape((-1, 9,500, 1))\n",
    "\n",
    "        y_pred = np.argmax(model.predict(file_data), axis=1) \n",
    "        counts = np.bincount(y_pred)\n",
    "        prediction = np.argmax(counts) #max occuring value in windows\n",
    "        \n",
    "        #correct prediction\n",
    "        if int(prediction)==int(file_label):\n",
    "            correct_count = correct_count+1\n",
    "        \n",
    "    \n",
    "    return total_count, correct_count"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import keras\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Conv2D, MaxPooling2D, LSTM, Dense, Dropout, Flatten, Bidirectional\n",
    "from keras.layers.normalization import BatchNormalization\n",
    "from keras.layers.core import Permute, Reshape\n",
    "from keras import backend as K\n",
    "from keras.callbacks import ModelCheckpoint"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def _data_reshaping(X_tr, X_va, X_ts, network_type):\n",
    "    _, win_len, dim = X_tr.shape\n",
    "    print(network_type)\n",
    "    if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "\n",
    "        X_tr = np.swapaxes(X_tr,1,2)\n",
    "        X_va = np.swapaxes(X_va,1,2)\n",
    "        X_ts = np.swapaxes(X_ts,1,2)\n",
    "\n",
    "        X_tr = np.reshape(X_tr, (-1, dim, win_len, 1))\n",
    "        X_va = np.reshape(X_va, (-1, dim, win_len, 1))\n",
    "        X_ts = np.reshape(X_ts, (-1, dim, win_len, 1))\n",
    "    \n",
    "    return X_tr, X_va, X_ts"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import regularizers\n",
    "\n",
    "def model_variant(model, num_feat_map, dim, network_type,p):\n",
    "    print(network_type)\n",
    "    if network_type == 'ConvLSTM':\n",
    "        model.add(Permute((2, 1, 3))) \n",
    "        model.add(Reshape((-1,num_feat_map*dim)))\n",
    "        model.add(Bidirectional(LSTM(128, return_sequences=False, stateful=False)))\n",
    "    if network_type == 'CNN':\n",
    "        \n",
    "        model.add(Flatten())\n",
    "        model.add(Dense(64, activation='relu'))\n",
    "        model.add(BatchNormalization()) \n",
    "        model.add(Dropout(p))\n",
    "\n",
    "        \n",
    "def model_conv(model, num_feat_map,p,b):\n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10),    \n",
    "                 activation='relu',\n",
    "                 input_shape=(dim, win_len, 1),\n",
    "                 padding='same'))\n",
    "    \n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10), activation='relu',padding='same'))\n",
    "    \n",
    "    if (b==1):\n",
    "        model.add(BatchNormalization()) \n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10), activation='relu',padding='same'))\n",
    "    \n",
    "    if (b==1):\n",
    "        model.add(BatchNormalization()) \n",
    "    model.add(MaxPooling2D(pool_size=(1, 3)))\n",
    "    \n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10), activation='relu',padding='same')) \n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10), activation='relu',padding='same'))\n",
    "    if (b==1):\n",
    "        model.add(BatchNormalization()) \n",
    "    model.add(MaxPooling2D(pool_size=(1, 2)))\n",
    "    model.add(Dropout(p))\n",
    "    \n",
    "    model.add(Conv2D(num_feat_map, kernel_size=(1, 10), activation='relu',padding='same'))  \n",
    "    if (b==1):\n",
    "        model.add(BatchNormalization()) \n",
    "    model.add(MaxPooling2D(pool_size=(1, 2)))\n",
    "    \n",
    "    model.add(Dropout(p))\n",
    "    \n",
    "def model_LSTM(model,p):\n",
    "    model.add(LSTM(num_hidden_lstm, \n",
    "               input_shape=(win_len,dim), \n",
    "               return_sequences=True))\n",
    "    model.add(Dropout(p))\n",
    "    model.add(LSTM(num_hidden_lstm, return_sequences=False))\n",
    "    model.add(Dropout(p))\n",
    "    \n",
    "def model_output(model):\n",
    "    model.add(Dense(num_classes, activation='softmax'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.models import load_model\n",
    "from sklearn.metrics import confusion_matrix\n",
    "from sklearn.metrics import f1_score, accuracy_score\n",
    "import matplotlib.pyplot as plt\n",
    "import itertools\n",
    "\n",
    "class monitor_Training(keras.callbacks.Callback):\n",
    "    def on_train_begin(self, logs={}):\n",
    "        return\n",
    " \n",
    "    def on_train_end(self, logs={}):\n",
    "        return\n",
    " \n",
    "    def on_epoch_begin(self, epoch, logs={}):\n",
    "        return\n",
    " \n",
    "    def on_epoch_end(self, epoch, logs={}):\n",
    "        y_pred = np.argmax(self.model.predict(X_test), axis=1)\n",
    "        accuracy = accuracy_score(y_true, y_pred)\n",
    "        print('test accuracy is:', accuracy)\n",
    "        return\n",
    " \n",
    "    def on_batch_begin(self, batch, logs={}):\n",
    "        return\n",
    " \n",
    "    def on_batch_end(self, batch, logs={}):\n",
    "        return"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training/Testing:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Add"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_add_TTV/train'\n",
    "data_dir_val= 'data_m_add_TTV/val'\n",
    "data_dir_test='data_m_add_TTV/test'\n",
    "pos_lab = 'Add'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelAdd.csv'\n",
    "model_name = \"DCBL_micro_ADD_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5053, 9, 500)\n",
      "(5053,)\n",
      "(5053, 9, 500)\n",
      "(5053,)\n",
      "(5054, 9, 500)\n",
      "(5054,)\n",
      "(1382, 9, 500)\n",
      "(1382,)\n",
      "(1378, 9, 500)\n",
      "(1378,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15160, 9, 500)\n",
      "(15160,)\n",
      "(1382, 9, 500)\n",
      "(1382,)\n",
      "(1378, 9, 500)\n",
      "(1378,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15160, 9, 500)\n",
      "(15160, 2)\n",
      "(1382, 9, 500)\n",
      "(1382, 2)\n",
      "(1378, 9, 500)\n",
      "(1378,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:68: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:504: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3828: The name tf.random_uniform is deprecated. Please use tf.random.uniform instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:166: The name tf.get_default_session is deprecated. Please use tf.compat.v1.get_default_session instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:180: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:1794: The name tf.nn.fused_batch_norm is deprecated. Please use tf.compat.v1.nn.fused_batch_norm instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3652: The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.\n",
      "\n",
      "WARNING:tensorflow:From /home/nesl/.conda/envs/tf-gpu/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3135: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.\n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_1 (Conv2D)            (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_2 (Conv2D)            (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_1 (Batch (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_3 (Conv2D)            (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_2 (Batch (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_1 (MaxPooling2 (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_4 (Conv2D)            (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_5 (Conv2D)            (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_3 (Batch (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_2 (MaxPooling2 (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_6 (Conv2D)            (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_4 (Batch (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_3 (MaxPooling2 (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_2 (Dropout)          (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_1 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_1 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_1 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15160, 9, 500, 1) (1382, 9, 500, 1) (1378, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "3\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.9659913169319826\n",
      "Training accuracy of best model:  0.9892480211396331\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  13   71]\n",
      " [   6 1288]]\n",
      "the mean-f1 score: 0.61\n",
      "accuracy is: 0.94\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 47\n",
      "Percentage accuracy: 94.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 186\n",
      "Percentage accuracy: 98.93617021276596\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 49\n",
      "Percentage accuracy: 98.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Cut"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_cut_TTV/train'\n",
    "data_dir_val= 'data_m_cut_TTV/val'\n",
    "data_dir_test='data_m_cut_TTV/test'\n",
    "pos_lab = 'Cut'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelCut.csv'\n",
    "model_name = \"DCBL_micro_Cut_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5072, 9, 500)\n",
      "(5072,)\n",
      "(5076, 9, 500)\n",
      "(5076,)\n",
      "(5071, 9, 500)\n",
      "(5071,)\n",
      "(1416, 9, 500)\n",
      "(1416,)\n",
      "(1327, 9, 500)\n",
      "(1327,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15219, 9, 500)\n",
      "(15219,)\n",
      "(1416, 9, 500)\n",
      "(1416,)\n",
      "(1327, 9, 500)\n",
      "(1327,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15219, 9, 500)\n",
      "(15219, 2)\n",
      "(1416, 9, 500)\n",
      "(1416, 2)\n",
      "(1327, 9, 500)\n",
      "(1327,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_7 (Conv2D)            (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_8 (Conv2D)            (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_5 (Batch (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_9 (Conv2D)            (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_6 (Batch (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_4 (MaxPooling2 (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_10 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_11 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_7 (Batch (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_5 (MaxPooling2 (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_3 (Dropout)          (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_12 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_8 (Batch (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2 (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_4 (Dropout)          (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_2 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_2 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_2 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15219, 9, 500, 1) (1416, 9, 500, 1) (1327, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.7838983050847458\n",
      "Training accuracy of best model:  0.8607004402548406\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[264 171]\n",
      " [ 76 816]]\n",
      "the mean-f1 score: 0.78\n",
      "accuracy is: 0.81\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 44\n",
      "Percentage accuracy: 88.0\n",
      "total_count of train files: 187\n",
      "correctly predicted train files: 185\n",
      "Percentage accuracy: 98.93048128342245\n",
      "total_count of val files: 51\n",
      "correctly predicted val files: 42\n",
      "Percentage accuracy: 82.35294117647058\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Wash"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_wash_TTV/train'\n",
    "data_dir_val= 'data_m_wash_TTV/val'\n",
    "data_dir_test='data_m_wash_TTV/test'\n",
    "pos_lab = 'Wash'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelWash.csv'\n",
    "model_name = \"DCBL_micro_Wash_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5073, 9, 500)\n",
      "(5073,)\n",
      "(5068, 9, 500)\n",
      "(5068,)\n",
      "(5072, 9, 500)\n",
      "(5072,)\n",
      "(1411, 9, 500)\n",
      "(1411,)\n",
      "(1331, 9, 500)\n",
      "(1331,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15213, 9, 500)\n",
      "(15213,)\n",
      "(1411, 9, 500)\n",
      "(1411,)\n",
      "(1331, 9, 500)\n",
      "(1331,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15213, 9, 500)\n",
      "(15213, 2)\n",
      "(1411, 9, 500)\n",
      "(1411, 2)\n",
      "(1331, 9, 500)\n",
      "(1331,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_13 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_14 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_9 (Batch (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_15 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_10 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_7 (MaxPooling2 (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_16 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_17 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_11 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_8 (MaxPooling2 (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_5 (Dropout)          (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_18 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_12 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_9 (MaxPooling2 (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_6 (Dropout)          (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_3 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_3 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_3 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15213, 9, 500, 1) (1411, 9, 500, 1) (1331, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "6\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.9114103472714387\n",
      "Training accuracy of best model:  0.9994741339643726\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 61,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  85  147]\n",
      " [   5 1094]]\n",
      "the mean-f1 score: 0.74\n",
      "accuracy is: 0.89\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 62,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 45\n",
      "Percentage accuracy: 90.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 188\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 46\n",
      "Percentage accuracy: 92.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 63,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Take"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_take_TTV/train'\n",
    "data_dir_val= 'data_m_take_TTV/val'\n",
    "data_dir_test='data_m_take_TTV/test'\n",
    "pos_lab = 'Take'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelTake.csv'\n",
    "model_name = \"DCBL_micro_Take_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5095, 9, 500)\n",
      "(5095,)\n",
      "(5092, 9, 500)\n",
      "(5092,)\n",
      "(5096, 9, 500)\n",
      "(5096,)\n",
      "(1382, 9, 500)\n",
      "(1382,)\n",
      "(1336, 9, 500)\n",
      "(1336,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15283, 9, 500)\n",
      "(15283,)\n",
      "(1382, 9, 500)\n",
      "(1382,)\n",
      "(1336, 9, 500)\n",
      "(1336,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15283, 9, 500)\n",
      "(15283, 2)\n",
      "(1382, 9, 500)\n",
      "(1382, 2)\n",
      "(1336, 9, 500)\n",
      "(1336,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_19 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_20 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_13 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_21 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_14 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_10 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_22 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_23 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_15 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_11 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_7 (Dropout)          (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_24 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_16 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_12 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_8 (Dropout)          (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_4 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_4 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_4 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 72,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15283, 9, 500, 1) (1382, 9, 500, 1) (1336, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "6\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.8241678730364992\n",
      "Training accuracy of best model:  0.9919518419158542\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[477 128]\n",
      " [117 614]]\n",
      "the mean-f1 score: 0.82\n",
      "accuracy is: 0.82\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 42\n",
      "Percentage accuracy: 84.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 188\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 45\n",
      "Percentage accuracy: 90.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 80,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Put"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 81,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_put_TTV/train'\n",
    "data_dir_val= 'data_m_put_TTV/val'\n",
    "data_dir_test='data_m_put_TTV/test'\n",
    "pos_lab = 'Put'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelPut.csv'\n",
    "model_name = \"DCBL_micro_Put_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 82,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5116, 9, 500)\n",
      "(5116,)\n",
      "(5109, 9, 500)\n",
      "(5109,)\n",
      "(5117, 9, 500)\n",
      "(5117,)\n",
      "(1344, 9, 500)\n",
      "(1344,)\n",
      "(1353, 9, 500)\n",
      "(1353,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15342, 9, 500)\n",
      "(15342,)\n",
      "(1344, 9, 500)\n",
      "(1344,)\n",
      "(1353, 9, 500)\n",
      "(1353,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15342, 9, 500)\n",
      "(15342, 2)\n",
      "(1344, 9, 500)\n",
      "(1344, 2)\n",
      "(1353, 9, 500)\n",
      "(1353,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_25 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_26 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_17 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_27 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_18 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_13 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_28 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_29 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_19 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_14 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_9 (Dropout)          (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_30 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_20 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_15 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_10 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_5 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_5 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_5 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_5 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15342, 9, 500, 1) (1344, 9, 500, 1) (1353, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 91,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 94,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "35\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.7924107142857143\n",
      "Training accuracy of best model:  0.9915917090340243\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 95,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[377  77]\n",
      " [152 747]]\n",
      "the mean-f1 score: 0.82\n",
      "accuracy is: 0.83\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 96,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 45\n",
      "Percentage accuracy: 90.0\n",
      "total_count of train files: 189\n",
      "correctly predicted train files: 189\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 49\n",
      "correctly predicted val files: 41\n",
      "Percentage accuracy: 83.6734693877551\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 97,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Peel"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_peel_TTV/train'\n",
    "data_dir_val= 'data_m_peel_TTV/val'\n",
    "data_dir_test='data_m_peel_TTV/test'\n",
    "pos_lab = 'Peel'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelPeel.csv'\n",
    "model_name = \"DCBL_micro_Peel_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 99,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 100,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5161, 9, 500)\n",
      "(5161,)\n",
      "(5162, 9, 500)\n",
      "(5162,)\n",
      "(5163, 9, 500)\n",
      "(5163,)\n",
      "(1280, 9, 500)\n",
      "(1280,)\n",
      "(1371, 9, 500)\n",
      "(1371,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15486, 9, 500)\n",
      "(15486,)\n",
      "(1280, 9, 500)\n",
      "(1280,)\n",
      "(1371, 9, 500)\n",
      "(1371,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 102,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15486, 9, 500)\n",
      "(15486, 2)\n",
      "(1280, 9, 500)\n",
      "(1280, 2)\n",
      "(1371, 9, 500)\n",
      "(1371,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_31 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_32 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_21 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_33 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_22 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_16 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_34 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_35 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_23 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_17 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_11 (Dropout)         (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_36 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_24 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_18 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_12 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_6 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_6 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_6 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 107,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15486, 9, 500, 1) (1280, 9, 500, 1) (1371, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 110,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 111,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "4\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.8515625\n",
      "Training accuracy of best model:  0.9806276637578247\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 112,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[316  90]\n",
      " [ 85 880]]\n",
      "the mean-f1 score: 0.84\n",
      "accuracy is: 0.87\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 113,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 49\n",
      "Percentage accuracy: 98.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 188\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 48\n",
      "Percentage accuracy: 96.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 114,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Mix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 115,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_mix_TTV/train'\n",
    "data_dir_val= 'data_m_mix_TTV/val'\n",
    "data_dir_test='data_m_mix_TTV/test'\n",
    "pos_lab = 'Mix'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelMix.csv'\n",
    "model_name = \"DCBL_micro_Mix_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 117,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5042, 9, 500)\n",
      "(5042,)\n",
      "(5034, 9, 500)\n",
      "(5034,)\n",
      "(5044, 9, 500)\n",
      "(5044,)\n",
      "(1407, 9, 500)\n",
      "(1407,)\n",
      "(1363, 9, 500)\n",
      "(1363,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 118,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15120, 9, 500)\n",
      "(15120,)\n",
      "(1407, 9, 500)\n",
      "(1407,)\n",
      "(1363, 9, 500)\n",
      "(1363,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 119,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15120, 9, 500)\n",
      "(15120, 2)\n",
      "(1407, 9, 500)\n",
      "(1407, 2)\n",
      "(1363, 9, 500)\n",
      "(1363,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 121,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 122,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_37 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_38 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_25 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_39 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_26 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_19 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_40 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_41 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_27 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_20 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_13 (Dropout)         (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_42 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_28 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_21 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_14 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_7 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_7 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_7 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_7 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 123,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15120, 9, 500, 1) (1407, 9, 500, 1) (1363, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 125,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "14\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.9388770433546553\n",
      "Training accuracy of best model:  0.9998677248677249\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 129,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  12   91]\n",
      " [   3 1257]]\n",
      "the mean-f1 score: 0.58\n",
      "accuracy is: 0.93\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 46\n",
      "Percentage accuracy: 92.0\n",
      "total_count of train files: 187\n",
      "correctly predicted train files: 187\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 51\n",
      "correctly predicted val files: 47\n",
      "Percentage accuracy: 92.15686274509804\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 131,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Pour"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 132,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_pour_TTV/train'\n",
    "data_dir_val= 'data_m_pour_TTV/val'\n",
    "data_dir_test='data_m_pour_TTV/test'\n",
    "pos_lab = 'Pour'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelPour.csv'\n",
    "model_name = \"DCBL_micro_Pour_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 133,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 134,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5206, 9, 500)\n",
      "(5206,)\n",
      "(5187, 9, 500)\n",
      "(5187,)\n",
      "(5205, 9, 500)\n",
      "(5205,)\n",
      "(1274, 9, 500)\n",
      "(1274,)\n",
      "(1335, 9, 500)\n",
      "(1335,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 135,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15598, 9, 500)\n",
      "(15598,)\n",
      "(1274, 9, 500)\n",
      "(1274,)\n",
      "(1335, 9, 500)\n",
      "(1335,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 136,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15598, 9, 500)\n",
      "(15598, 2)\n",
      "(1274, 9, 500)\n",
      "(1274, 2)\n",
      "(1335, 9, 500)\n",
      "(1335,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 137,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 138,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 139,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_43 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_44 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_29 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_45 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_30 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_22 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_46 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_47 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_31 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_23 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_15 (Dropout)         (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_48 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_32 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_24 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_16 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_8 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_8 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_8 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_8 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 140,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 141,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15598, 9, 500, 1) (1274, 9, 500, 1) (1335, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 142,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 144,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 145,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "9\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.9246467817896389\n",
      "Training accuracy of best model:  0.9992306705987947\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 146,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  37  110]\n",
      " [   4 1184]]\n",
      "the mean-f1 score: 0.67\n",
      "accuracy is: 0.91\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 147,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 45\n",
      "Percentage accuracy: 90.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 188\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 46\n",
      "Percentage accuracy: 92.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 148,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Open"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 149,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_open_TTV/train'\n",
    "data_dir_val= 'data_m_open_TTV/val'\n",
    "data_dir_test='data_m_open_TTV/test'\n",
    "pos_lab = 'Open'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'LabelOpen.csv'\n",
    "model_name = \"DCBL_micro_Open_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 150,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 151,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5046, 9, 500)\n",
      "(5046,)\n",
      "(5033, 9, 500)\n",
      "(5033,)\n",
      "(5046, 9, 500)\n",
      "(5046,)\n",
      "(1428, 9, 500)\n",
      "(1428,)\n",
      "(1340, 9, 500)\n",
      "(1340,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 152,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15125, 9, 500)\n",
      "(15125,)\n",
      "(1428, 9, 500)\n",
      "(1428,)\n",
      "(1340, 9, 500)\n",
      "(1340,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 153,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15125, 9, 500)\n",
      "(15125, 2)\n",
      "(1428, 9, 500)\n",
      "(1428, 2)\n",
      "(1340, 9, 500)\n",
      "(1340,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 154,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 155,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 156,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_49 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_50 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_33 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_51 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_34 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_25 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_52 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_53 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_35 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_26 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_17 (Dropout)         (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_54 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_36 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_27 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_18 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_9 (Permute)          (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_9 (Reshape)          (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_9 (Bidirection (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_9 (Dense)              (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 157,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 158,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15125, 9, 500, 1) (1428, 9, 500, 1) (1340, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 159,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 161,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 162,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "29\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.9292717085165136\n",
      "Training accuracy of best model:  0.9999338842975206\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 163,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  77   39]\n",
      " [  17 1207]]\n",
      "the mean-f1 score: 0.85\n",
      "accuracy is: 0.96\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 164,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 47\n",
      "Percentage accuracy: 94.0\n",
      "total_count of train files: 187\n",
      "correctly predicted train files: 187\n",
      "Percentage accuracy: 100.0\n",
      "total_count of val files: 51\n",
      "correctly predicted val files: 47\n",
      "Percentage accuracy: 92.15686274509804\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 165,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "other"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 166,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_train='data_m_other_TTV/train'\n",
    "data_dir_val= 'data_m_other_TTV/val'\n",
    "data_dir_test='data_m_other_TTV/test'\n",
    "pos_lab = 'other'\n",
    "neg_lab = 'No'\n",
    "labelfile = 'Labelother.csv'\n",
    "model_name = \"DCBL_micro_other_jitter.hdf5\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 167,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FEYY86nqUvoy",
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "#receive windowed training and validation data\n",
    "train_x, train_y, train_files = get_train_data(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "val_x, val_y, val_files = get_train_data(data_dir_val,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "test_x, test_y, test_files = get_train_data(data_dir_test,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x2, train_y2, all_counts = get_train_data_2(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)\n",
    "train_x3, train_y3, all_counts = get_train_data_3(data_dir_train,sub_dirs,pos_lab,neg_lab,labelfile)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 168,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "u2w_y9dmUvo1"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5107, 9, 500)\n",
      "(5107,)\n",
      "(5103, 9, 500)\n",
      "(5103,)\n",
      "(5107, 9, 500)\n",
      "(5107,)\n",
      "(1373, 9, 500)\n",
      "(1373,)\n",
      "(1334, 9, 500)\n",
      "(1334,)\n"
     ]
    }
   ],
   "source": [
    "train_samples_2 = np.array(train_x2) \n",
    "train_labels2_2 = np.array(train_y2)\n",
    "print(train_samples_2.shape)\n",
    "print(train_labels2_2.shape)\n",
    "train_samples_3 = np.array(train_x3) \n",
    "train_labels2_3 = np.array(train_y3)\n",
    "print(train_samples_3.shape)\n",
    "print(train_labels2_3.shape)\n",
    "train_samples = np.array(train_x) \n",
    "train_labels2 = np.array(train_y)\n",
    "val_samples = np.array(val_x) \n",
    "val_labels2 = np.array(val_y)\n",
    "test_samples = np.array(test_x) \n",
    "test_labels2 = np.array(test_y)\n",
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)\n",
    "train_samples = np.vstack((train_samples,train_samples_2, train_samples_3))\n",
    "train_labels2 = np.hstack((train_labels2, train_labels2_2, train_labels2_3))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 169,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15317, 9, 500)\n",
      "(15317,)\n",
      "(1373, 9, 500)\n",
      "(1373,)\n",
      "(1334, 9, 500)\n",
      "(1334,)\n"
     ]
    }
   ],
   "source": [
    "print(train_samples.shape)\n",
    "print(train_labels2.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels2.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 170,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15317, 9, 500)\n",
      "(15317, 2)\n",
      "(1373, 9, 500)\n",
      "(1373, 2)\n",
      "(1334, 9, 500)\n",
      "(1334,)\n"
     ]
    }
   ],
   "source": [
    "#convert to one hot encoding\n",
    "from keras.utils  import to_categorical\n",
    "train_labels = to_categorical(train_labels2)\n",
    "train_labels.shape\n",
    "val_labels = to_categorical(val_labels2)\n",
    "val_labels.shape\n",
    "test_labels = to_categorical(test_labels2)\n",
    "test_labels.shape\n",
    "print(train_samples.shape)\n",
    "print(train_labels.shape)\n",
    "print(val_samples.shape)\n",
    "print(val_labels.shape)\n",
    "print(test_samples.shape)\n",
    "print(test_labels2.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 171,
   "metadata": {},
   "outputs": [],
   "source": [
    "#For CNN and BidirLSTM:\n",
    "number_of_samples = 500\n",
    "train_samples = train_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "val_samples = val_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "test_samples = test_samples.reshape((-1, 9,number_of_samples, 1))\n",
    "#For LSTM, uncomment these lines and comment above two lines\n",
    "#train_samples = train_samples.reshape((-1, number_of_samples, 9))\n",
    "#val_samples = val_samples.reshape((-1, number_of_samples, 9))\n",
    "\n",
    "#9 refers to 9 accelerometer channels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 172,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wf8LuDxlUvpH"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "9\n"
     ]
    }
   ],
   "source": [
    "batch_size = 64\n",
    "num_feat_map = 128\n",
    "num_hidden_lstm = 128\n",
    "num_classes = 2\n",
    "\n",
    "#Uncomment the classifier you are working with:\n",
    "#network_type = 'CNN'\n",
    "network_type = 'ConvLSTM'\n",
    "#network_type = 'LSTM'\n",
    "#for CNN/DCBL\n",
    "_, dim, win_len,_ = train_samples.shape\n",
    "#for LSTM uncomment the following line and comment above line\n",
    "#_, dim, win_len = train_samples.shape\n",
    "\n",
    "print(win_len)\n",
    "print(dim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 173,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "fAQRf1DNUvpK"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "building the model ... \n",
      "ConvLSTM\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_55 (Conv2D)           (None, 9, 500, 128)       1408      \n",
      "_________________________________________________________________\n",
      "conv2d_56 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_37 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "conv2d_57 (Conv2D)           (None, 9, 500, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_38 (Batc (None, 9, 500, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_28 (MaxPooling (None, 9, 166, 128)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_58 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "conv2d_59 (Conv2D)           (None, 9, 166, 128)       163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_39 (Batc (None, 9, 166, 128)       512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_29 (MaxPooling (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_19 (Dropout)         (None, 9, 83, 128)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_60 (Conv2D)           (None, 9, 83, 128)        163968    \n",
      "_________________________________________________________________\n",
      "batch_normalization_40 (Batc (None, 9, 83, 128)        512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_30 (MaxPooling (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "dropout_20 (Dropout)         (None, 9, 41, 128)        0         \n",
      "_________________________________________________________________\n",
      "permute_10 (Permute)         (None, 41, 9, 128)        0         \n",
      "_________________________________________________________________\n",
      "reshape_10 (Reshape)         (None, 41, 1152)          0         \n",
      "_________________________________________________________________\n",
      "bidirectional_10 (Bidirectio (None, 256)               1311744   \n",
      "_________________________________________________________________\n",
      "dense_10 (Dense)             (None, 2)                 514       \n",
      "=================================================================\n",
      "Total params: 2,135,554\n",
      "Trainable params: 2,134,530\n",
      "Non-trainable params: 1,024\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "p=0.5 #Dropout\n",
    "b = 1 #BatchNorm\n",
    "print('building the model ... ')\n",
    "model = Sequential()\n",
    "\n",
    "if network_type=='CNN' or network_type=='ConvLSTM':\n",
    "    model_conv(model, num_feat_map,p,b)\n",
    "    model_variant(model, num_feat_map, dim, network_type,p)\n",
    "if network_type=='LSTM':\n",
    "    model_LSTM(model,p)\n",
    "       \n",
    "model_output(model)    \n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 174,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "wrNCEAD8UvpL"
   },
   "outputs": [],
   "source": [
    "X_train = train_samples\n",
    "y_train = train_labels\n",
    "X_val = val_samples\n",
    "y_val = val_labels\n",
    "X_test = test_samples\n",
    "y_test = test_labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 175,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tNsDOBNrUvpN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(15317, 9, 500, 1) (1373, 9, 500, 1) (1334, 9, 500, 1)\n"
     ]
    }
   ],
   "source": [
    "print(X_train.shape, X_val.shape, X_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 176,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = np.argmax(y_test, axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "YpsJRPEfUvpP",
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "epochs = 40\n",
    "\n",
    "model.compile(loss=keras.losses.categorical_crossentropy,\n",
    "              optimizer='adam',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# checkpoint\n",
    "filepath=model_name\n",
    "#checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True)\n",
    "\n",
    "checkpoint2 = monitor_Training()\n",
    "\n",
    "callbacks_list = [checkpoint, checkpoint2]\n",
    "\n",
    "H = model.fit(train_samples, train_labels,\n",
    "            batch_size=batch_size,\n",
    "            epochs=epochs,\n",
    "            verbose=1,\n",
    "            shuffle=True,\n",
    "            validation_data=(X_val, y_val),\n",
    "             callbacks=callbacks_list\n",
    "             )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 178,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "_ASUmtFSUvpQ"
   },
   "outputs": [],
   "source": [
    "history = H"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 179,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "C6yi5aaJUvpS"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "7\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYgAAAEWCAYAAAB8LwAVAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4xLjMsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy+AADFEAAAgAElEQVR4nO3dd3zU9f3A8dc7lx1CIIMZIEwRAVHiFnHviloXra2z1P5crfXXauuvjtZWu9RWq0VFcVIctVhRVBT3YCOyQRICRDKAQHZy798fn2/gCJfkEnK5jPfz8bhH7r7r3ndJvu/vZ35FVTHGGGPqi4p0AMYYY9onSxDGGGOCsgRhjDEmKEsQxhhjgrIEYYwxJihLEMYYY4KyBGG6PBHJEhEVkegQtr1SRD5ui7iMiTRLEKZDEZGNIlIlIun1li/2TvJZkYnMmM7HEoTpiL4BJte9EJExQGLkwmkfQikBGdMcliBMR/Qs8MOA11cAzwRuICIpIvKMiBSISI6I3CEiUd46n4j8WUQKRWQDcE6QfZ8Uka0isllEficivlACE5GXRCRfRHaKyIcickjAugQR+YsXz04R+VhEErx1x4vIpyKyQ0Q2iciV3vJ5InJtwDH2qeLySk3Xi8haYK237CHvGCUislBEJgRs7xORX4nIehHZ5a0fICKPiMhf6n2WWSLys1A+t+mcLEGYjuhzoLuIHOyduC8Dnqu3zd+BFGAIMBGXUK7y1v0IOBc4DMgGLqq379NADTDM2+Z04FpC8yYwHOgFLAKeD1j3Z2A8cCyQCvwC8IvIIG+/vwMZwDhgSYjvB3A+cBQwyns93ztGKvAC8JKIxHvrbsGVvs4GugNXA2XAdGByQBJNB0719jddlarawx4d5gFsxJ247gD+AJwJvANEAwpkAT6gChgVsN+PgXne8/eA6wLWne7tGw30BiqBhID1k4H3vedXAh+HGGsP77gpuIuxcuDQINvdDvy7gWPMA64NeL3P+3vHP7mJOLbXvS+wGpjUwHYrgdO85zcAsyP9+7ZHZB9WZ2k6qmeBD4HB1KteAtKBGCAnYFkO0N973g/YVG9dnUHevltFpG5ZVL3tg/JKM/cCF+NKAv6AeOKAeGB9kF0HNLA8VPvEJiK3AtfgPqfiSgp1jfqNvdd04HJcwr0ceOgAYjKdgFUxmQ5JVXNwjdVnA6/WW10IVONO9nUGApu951txJ8rAdXU24UoQ6araw3t0V9VDaNr3gEm4Ek4KrjQDIF5MFcDQIPttamA5QCn7NsD3CbLNnimZvfaGXwCXAD1VtQew04uhqfd6DpgkIocCBwOvNbCd6SIsQZiO7Bpc9Upp4EJVrQVmAveKSLJXx38Le9spZgI3iUimiPQEbgvYdyvwNvAXEekuIlEiMlREJoYQTzIuuRThTuq/DziuH5gG/FVE+nmNxceISByuneJUEblERKJFJE1Exnm7LgEuFJFEERnmfeamYqgBCoBoEfkNrgRR5wngtyIyXJyxIpLmxZiHa794FnhFVctD+MymE7MEYTosVV2vqgsaWH0j7up7A/AxrrF1mrfucWAOsBTXkFy/BPJDIBZYgau/fxnoG0JIz+CqqzZ7+35eb/2twFe4k3AxcD8Qpaq5uJLQz73lS4BDvX0ewLWnfIurAnqexs0B3gLWeLFUsG8V1F9xCfJtoAR4EkgIWD8dGINLEqaLE1W7YZAxxhGRE3AlrUFqJ4cuz0oQxhgARCQGuBl4wpKDAUsQxhhARA4GduCq0h6McDimnbAqJmOMMUFZCcIYY0xQnWagXHp6umZlZUU6DGOM6VAWLlxYqKoZwdZ1mgSRlZXFggUN9Xg0xhgTjIjkNLTOqpiMMcYEZQnCGGNMUJYgjDHGBNVp2iCCqa6uJi8vj4qKikiH0mbi4+PJzMwkJiYm0qEYYzq4Tp0g8vLySE5OJisri4CpmzstVaWoqIi8vDwGDx4c6XCMMR1c2KqYRGSaiGwTkeUNrBcR+ZuIrBORZSJyeMC6K0Rkrfe4oqUxVFRUkJaW1iWSA4CIkJaW1qVKTMaY8AlnG8TTuLt9NeQs3K0ZhwNTgEcBRCQVuBN3C8UjgTu9KZlbpKskhzpd7fMaY8InbFVMqvqhiGQ1sskk4BlvUrDPRaSHiPQFTgTeUdViABF5B5doXgxXrMZ0FdW1fgSI9nWs/illVTV8W1JJaWUNlTV+Kmtq3c/qvc/9fuW4YekMSE1s+oBhoKqUVdWys7yaHWXV7CyvZmd5FTvLqykpr6HGr/j33M4T/AqK4legiSmPYnxR9OoeR6/keDKS4+jdPZ60pFiiosJ7QRjJNoj+7DtPfZ63rKHl+xGRKbjSBwMHDgy2SUQVFRVxyimnAJCfn4/P5yMjww1Y/PLLL4mNjW3yGFdddRW33XYbBx10UFhj7Wr8fkWkY5W4dpZVs2JrCXnby0hNiiUjOY70bu4RG73vCV9V+bakkpX5JazauotV+SWszt/Fum27iY/xcfSQNI4flsbxw9MZmtHtgL8HVWVTcTnrC3cT64siPsZHfEwUCTE+77mPhBgfIlBeVUtZdS3lVTWUVdVSVlXrllXVUlxWxbc7K8gvqeDbkgryvee7KmpCjuW4YWlckj2AMw7pQ3yM74A+V2N2VVTzyboiPlxbwCfrCtm8vZwaf8vmtmvq6w+WP3xRQka3OHp1j+OwAT24e9LoFr13Yzp0I7WqTgWmAmRnZ7e7WQfT0tJYsmQJAHfddRfdunXj1ltv3WebupuDR0UFv6J76qmnwh5nS3yyrpDHPlhPt7hoxmSmMLZ/D8b0TyElMbK9p3ZX1rB8806Wb95Jwe5KdnpXcnuv6Nxjd6U74cT4hOioKKKjhGifEO2LIiZK8PkEv3dHab/WXfl5V30hTHA5tFc3Lhqfydlj+tItrnn/Zn6/sml7GSu2lLByawkrtpawcusuNu9o+AZvPRJjyPCShV+V1d/uYkdZ9Z71/XskcFCfZE4a2Yud5dV8sq6Qd1d+C0Dv7nEcNyyd44elc9ywdHp3j28yxqLdlSzN28GSTTtZumkHy/J2sD3g/Q5ElEBGchx9usczOD2JY4am0bt7PL27x9M9Ppq4GB9x0VHew0dcjHteWePnjWVbmblgEzfPWEJyfDSTxvXjkuwBjOmfcsBJ0O9Xlm/ZyYdrCvhwTSELc7dT61eSYn0cMzSds8f0JSUhhh4JMaTUPRLdz+4JMcR6pbYoEaK8i5OoEC9SKmtqKdhVybZdlWwrqWDbrkq+LalgW4lbVlUbntNfJBPEZva9L3Cmt2wzrpopcPm8NouqDaxbt47zzjuPww47jMWLF/POO+9w9913s2jRIsrLy7n00kv5zW9+A8Dxxx/Pww8/zOjRo0lPT+e6667jzTffJDExkf/85z/06tWrTWNfnb+LP7y5knmrC+ibEk9sdBRvLs/fsz4rLZExmT0Y2z+FsZkpjO6fQlIzT5ChKq+qZcXWnSzL28lXeTtZmreDDYWle662Yn1Re/5BeyTE0DclnpF9k0lJiCE5PgZUqfYrNbV+qmuVWr9S43fP/X4FCfhnRoiKcv/MQuNXfH6Fz9cX8YuXl3HXrK85a3RfLhqfyVGDU4NWCdT6lRVbSvhsQyGfrS9iwcbt7PISmC9KGJKeRHZWT37QdxCj+nZnYGoiO8qrKdhVSeHuSgp2BTx2V6KqnDW6Lwf3TWZkn+4c1Ds5aOLeVFzGx+sK+XhdIe+v2sari9wtu2Ojo0iM9ZEY4yMh1kdibLR7HevDFyWsyt9F3naXrKIERvRO5vRRfTh0QA9G9O5GrV+pqPFTXlVLZY0rHZRX11JR7afW799zvMBjJ3jH75EQS3q32BZXgd10ynBuOGkYn28oYuaCTby0II/nPs9lZJ9kJo3rz6C0RNKSYknrFkdGtzi6J0Tvd4LeXVnDpuIy99hevuf54k07KC6tAmB0/+78+IQhnDAig8MH9tyvBNfa4qJ9ZPZMJLNn21afRTJBzAJuEJEZuAbpnaq6VUTmAL8PaJg+Hbj9QN/s7te/ZsWWkgM9zD5G9evOnd8J5V72+1u1ahXPPPMM2dnZANx3332kpqZSU1PDSSedxEUXXcSoUaP22Wfnzp1MnDiR++67j1tuuYVp06Zx22237bONqlJV4+dPc1bx7opt5BaXMXFEBmeN6cPJI3u5E2MLfFtSwV/fXsNLCzeRFBfN7WeN5Ipjs4iP8bGjrIrlm0tYmreDr/J2sihnO68v3QK4E8jwXskcOiCFsZk9GDegBwf1SSYmxBNAZU0tm4rLyS0uJaeojJyiMnKLy8gpKmVjURm1XpE+IzmOQzNTOO/Q/owdkMKY/imkJcVGrApJVVmUu4OXF27i9aVbeWVRHgNSE/ju4ZlceFgmuyqr+XxDMZ+tL+LLb4oo8apQhmQk8Z1x/RjbP4VR/bozondy2KpJBqQmMvnIgUw+ciB+v7JiawmfbyiiYHflniof99NVBRWVVlFV4+fQAT244pissF8AtFRUlHDssHSOHZbO3eXVvL50CzMXbOL+t1btt210lJDWLZa0pDiifULe9vI9SaBOUqyPAamJTByRwQkj0jl+WAYZyXFt9XEiKmy/WRF5EVcSSBeRPFzPpBgAVX0MmI27D+86oAy4yltXLCK/xd23F+CeugbrzmTo0KF7kgPAiy++yJNPPklNTQ1btmxhxYoV+yWIhIQEzjrrLADGjx/PRx99BLgr0N2VNewqr6akooZtuyp57IN8jsjqyWED+/Heqm289XU+sdFRnDA8g7NG9+HUUb1JSWg6WeyurGHqB+t5/KNvqPH7ufLYwdx48jB6Ju1tP+mRGMvxw9M5fnj6nmUFuyr5avMOlm5yV/bvrPiWmQvyAIiLjuKQft0Z3iuZWlWvsbF2T+NjVY2fyho/20ur2FpSsU/9a2Ksj4GpiQzr1Y2zRvdlbGYKhw7oEVK1SFsSEcYP6sn4QT35zbmHMOfrfF5auImH5q7lwXfX7tkuKy2Rs8f05ZihaRw9JC1inyMqShjd353wO5OUhBguP3oQlx89aE+Jq3B3JUW7q9zP0iqKvNdVtX4O6ZfCgNQEBqYmMqBnIgNSE+mZGNOh2qpaUzh7MU1uYr0C1zewbhp7bzDfKlp6pR8uSUlJe56vXbuWhx56iC+//JIePXpw+eWXBx3LEBsbu+cEWlbtp6Ssgm8KSymtrMGvik+E5PhoUpNiWHTHaXuqFfx+ZVHudmZ/lc+by7fy7spvifEJxw1LZ/xAV1BT6uraAe9nVa2fVxflUbi7inPH9uUXZ4xkYFpoRdyM5DhOHtmbk0f2dsdXJW97OUs27WDpph0szdvB3FXb9tQlx0ZHubplXxRJcdGkJkVxUO9kBqYlMigtkYGpiQxMTSK9W+RKBS2VEOvj/MP6c/5h/cnbXsZby/NJTYrl6CFp9OuREOnwuoyM5Lguc+XfWtpX2bALUlWKt++gW7dkfHGJrNmQy1tvzeHYiaeQv7Ocqho/ecVlROeXuMbH/F0AFJdWUVHtp7rGT2pSLN3jo0mMiyZKhNJt0fvUOUdFCdlZqWRnpXLHOQezNG8Hby7PZ/ZXW5m3umC/mETw6tmF7EE9efyHIzlsYIuHonjHFAakuiuy7xza74CO1ZFl9kzk2glDIh2GMSGxBBFB1TV+vikqJb7vMPoNHsboQ0bRL3MAY8cfyY7yKgp2VVHjV2pREmN8CEJmzwRifT7690igZ2IMI/okN+s9o6KEwwb25LCBPbn9rJFU17runlEBja8d7QrdGBMeneae1NnZ2Vr/hkErV67k4IMPjlBEjauq8bOhcDe1tUqv7nFE+6LwRQnRUYKv7iHSopN1e/7cxpj2RUQWqmp2sHVWgoiAwOSQlZ7U7nqBGGMMWIJoc1U1tWwoLKXWrwzOSCIx1n4Fxpj2yc5ObaiqppYNBaXUqjI43ZKDMaZ9szNUG6msqeUbSw7GmA7EzlJtoNIrOfhVGZKeRIIlB2NMB2BnqjCrrK715gey5GCM6VjsbBVGGzfnc+bppwHC9qJtRLdgum+AadOmcfbZZ9OnT58wRmuMMfuyBBEGqkrBrkpKNJ5Z733KoNRE/nDvb4NO9x2KadOmcfjhh1uCMMa0KUsQrazW72dTcTklFdX0SIilf88EfEGmeJ4+fTqPPPIIVVVVHHvssTz88MP4/X6uuuoqlixZgqoyZcoUevfuzZIlS7j00ktJSEhoVsnDGGMORNdJEG/eBvlfte4x+4yBs+7b87KiupacojKqavz0S0kgrYGJ5ZYvX86///1vPv30U6Kjo5kyZQozZsxg6NChFBYW8tVXLs4dO3bQo0cP/v73v/Pwww8zbty41o3fGGMa0XUSRJjtKKsib3s5USIMyWh8dPS7777L/Pnz90z3XV5ezoABAzjjjDNYvXo1N910E+eccw6nn356W4VvjDH76ToJIuBKvzWpKvk7KyjYXUlibDSD0hKbvBmOqnL11Vfz29/+dr91y5Yt48033+SRRx7hlVdeYerUqWGJ2xhjmhLe++R1Adu82zymdYtjSEZSSHdKO/XUU5k5cyaFhYUAFBUVkZubS0FBAarKxRdfzD333MOiRYsASE5OZteuXWH9HMYYU1/XKUGEgapSXFpFt7ho+jfjxi9jxozhzjvv5NRTT8Xv9xMTE8Njjz2Gz+fjmmuuQVUREe6//34ArrrqKq699lprpDbGtCmb7vsAlFRUs7GwlEGpiaQktp+Ttk33bYwJVWPTfVsV0wHYXlpFdFQUySHc29kYYzoaSxAtVF3rp6S8hp5JMUTZHdiMMZ1Qp08Q4apC215WhaL0bEdVSxC+z2uM6Xo6dYKIj4+nqKio1U+aqsr20mqSYqOJj/G16rEPhKpSVFREfHx8pEMxxnQCnboXU2ZmJnl5eRQUFLTqcStrainYVUVqUgwri9rXVxgfH09mZmakwzDGdALt6+zWymJiYhg8eHCrH/dn/1rCuysL+PJXp5IQ235KEMYY05o6dRVTOOwsq2b2V1s5f1x/Sw7GmE7NEkQzvbZkM5U1fi49YkCkQzHGmLCyBNEMqsqLX+Yypn8Ko/unRDocY4wJK0sQzbAsbyer8ndZ6cEY0yWENUGIyJkislpE1onIbUHWDxKRuSKyTETmiUhmwLpaEVniPWaFM85QzZifS0KMj0nj+kU6FGOMCbuw9WISER/wCHAakAfMF5FZqroiYLM/A8+o6nQRORn4A/ADb125qrabO+SUVtYwa8kWzhnbl+R4m1rDGNP5hbMEcSSwTlU3qGoVMAOYVG+bUcB73vP3g6xvN95YtpXSqlous+olY0wXEc4E0R/YFPA6z1sWaClwoff8AiBZRNK81/EiskBEPheR84O9gYhM8bZZ0NqD4ep7cX4uw3p1Y/ygnmF9H2OMaS8i3Uh9KzBRRBYDE4HNQK23bpA3Be33gAdFZGj9nVV1qqpmq2p2RkZG2IJcnb+Lxbk7uOyIAUHvMW2MMZ1ROEdSbwYC62MyvWV7qOoWvBKEiHQDvquqO7x1m72fG0RkHnAYsD6M8TZoxvxcYnzChYfbFBbGmK4jnCWI+cBwERksIrHAZcA+vZFEJF1E6mK4HZjmLe8pInF12wDHAYGN222morqWfy/ezOmH9CE1qX3N3GqMMeEUtgShqjXADcAcYCUwU1W/FpF7ROQ8b7MTgdUisgboDdzrLT8YWCAiS3GN1/fV6/3UZr74ppgdZdVcNN5KD8aYriWsk/Wp6mxgdr1lvwl4/jLwcpD9PgXGhDO2UG0o2A3AGBs5bYzpYiLdSN3u5RSVkRTrI82ql4wxXYwliCbkFpcxMC3Jei8ZY7ocSxBN2FhUSlZaYqTDMMaYNmcJohG1fiWvuJyBliCMMV2QJYhG5JdUUFXrZ1BqUqRDMcaYNmcJohE5haUAVsVkjOmSLEE0Iqe4DMCqmIwxXZIliEbkFJUR4xP6piREOhRjjGlzliAakVNUyoDURHxR1sXVGNP1WIJoRE5RGYNSrXrJGNM1WYJogKqSW1zGoDTrwWSM6ZosQTSgqLSK3ZU1DLIGamNMF2UJogE5Ra4HkyUIY0xXZQmiAbnFbgzEQBskZ4zpoixBNGBjYRkiMCDVurgaY7omSxANyC0uo19KAnHRvkiHYowxEWEJogE5RaUMtC6uxpguzBJEA3KKyshKtwRhjOm6LEEEsbuyhqLSKmugNsZ0aZYggsgpcj2YrIurMaYrswQRhI2BMMYYSxBB7U0QVsVkjOm6LEEEkVtcSlpSLN3ioiMdijHGRIwliCA2FpZZ9ZIxpsuzBBGEzeJqjDGWIPZTWVPLlp3lNkjOGNPlWYKoZ1NxOarYIDljTJdnCaIem8XVGGOcsCYIETlTRFaLyDoRuS3I+kEiMldElonIPBHJDFh3hYis9R5XhDPOQDYGwhhjnLAlCBHxAY8AZwGjgMkiMqreZn8GnlHVscA9wB+8fVOBO4GjgCOBO0WkZ7hiDZRTVEa3uGjSkmLb4u2MMabdCmcJ4khgnapuUNUqYAYwqd42o4D3vOfvB6w/A3hHVYtVdTvwDnBmGGPdo24WVxFpi7czxph2K5wJoj+wKeB1nrcs0FLgQu/5BUCyiKSFuC8iMkVEFojIgoKCglYJOqfYxkAYYwxEvpH6VmCiiCwGJgKbgdpQd1bVqaqararZGRkZBxxMrV/ZZGMgjDEGgHDOJbEZGBDwOtNbtoeqbsErQYhIN+C7qrpDRDYDJ9bbd14YYwVg685yqmvVShDGGEN4SxDzgeEiMlhEYoHLgFmBG4hIuojUxXA7MM17Pgc4XUR6eo3Tp3vLwiq3rgeTDZIzxpjwJQhVrQFuwJ3YVwIzVfVrEblHRM7zNjsRWC0ia4DewL3evsXAb3FJZj5wj7csrDbWJYh0q2Iyxpgmq5hE5EbgOa83UbOo6mxgdr1lvwl4/jLwcgP7TmNviaJN5BSXEuuLok/3+LZ8W2OMaZdCKUH0BuaLyExv4Fun7f+ZW1RGZmoCvqhO+xGNMSZkTSYIVb0DGA48CVwJrBWR34vI0DDH1uY2FpWRZT2YjDEGCLENQlUVyPceNUBP4GUR+WMYY2tTqkquN0jOGGNMaG0QNwM/BAqBJ4D/VdVqr/fRWuAX4Q2xbRSVVlFaVWtdXI0xxhPKOIhU4EJVzQlcqKp+ETk3PGG1vZwiN4urVTEZY4wTShXTm8CeLqYi0l1EjgJQ1ZXhCqyt1c3iOtBKEMYYA4SWIB4Fdge83u0t61RyisoQgcyeCZEOxRhj2oVQEoR4jdSAq1oivFN0REROUSn9UhKIi/ZFOhRjjGkXQkkQG0TkJhGJ8R43AxvCHVhbs1lcjTFmX6EkiOuAY3ET7eXhbuIzJZxBRUJukSUIY4wJ1GRVkapuw02012ntqqimqLTKpvk2xpgAoYyDiAeuAQ4B9kxSpKpXhzGuNpVjs7gaY8x+Qqliehbog7sN6Ae4ezPsCmdQbS232Lq4GmNMfaEkiGGq+n9AqapOB87BtUN0Ghu9QXJWxWSMMXuFkiCqvZ87RGQ0kAL0Cl9IbS+3qIz0brF0i+t0vXeNMabFQjkjTvXu6nYH7o5w3YD/C2tUbSynqMwm6TPGmHoaTRDehHwl3s2CPgSGtElUbSynqJSjh6RFOgxjjGlXGq1i8kZNd4rZWhtSWVPL1pIKa6A2xph6QmmDeFdEbhWRASKSWvcIe2RtpKS8hqMGpzKqb/dIh2KMMe2KBEyzFHwDkW+CLFZVbVfVTdnZ2bpgwYJIh2GMMR2KiCxU1exg60IZST249UMyxhjT3oUykvqHwZar6jOtH44xxpj2IpRurkcEPI8HTgEWAZYgjDGmEwuliunGwNci0gOYEbaIjDHGtAuh9GKqrxSwdgljjOnkQmmDeB2o6+oUBYwCZoYzKGOMMZEXShvEnwOe1wA5qpoXpniMMca0E6EkiFxgq6pWAIhIgohkqerGsEZmjDEmokJpg3gJ8Ae8rvWWNUlEzhSR1SKyTkRuC7J+oIi8LyKLRWSZiJztLc8SkXIRWeI9Hgvl/YwxxrSeUEoQ0apaVfdCVatEJLapnUTEBzwCnIa7l/V8EZmlqisCNrsDmKmqj4rIKGA2kOWtW6+q40L8HMYYY1pZKCWIAhE5r+6FiEwCCkPY70hgnapu8BLMDGBSvW0UqJsEKQXYEsJxjTHGtIFQShDXAc+LyMPe6zwg6OjqevoDmwJe57H/nejuAt4WkRuBJODUgHWDRWQxUALcoaof1X8DEZkCTAEYOHBgCCEZY4wJVSgD5dYDR4tIN+/17lZ8/8nA06r6FxE5BnjWu2vdVmCgqhaJyHjgNRE5RFVL6sU2FZgKbrK+VozLGGO6vCarmETk9yLSQ1V3q+puEekpIr8L4dibgQEBrzO9ZYGuwRtToaqf4abySFfVSlUt8pYvBNYDI0J4T2OMMa0klDaIs1R1R90L7+5yZ4ew33xguIgM9hq1L8PdsjRQLm5uJ0TkYFyCKBCRDK+RGxEZAgwHNoTwnsYYY1pJKG0QPhGJU9VKcOMggLimdlLVGhG5AZgD+IBpqvq1iNwDLFDVWcDPgcdF5Ge4BusrVVVF5ATgHhGpxnWxvU5Vi1v0CY0xxrRIKAnieWCuiDwFCHAlMD2Ug6vqbFzX1cBlvwl4vgI4Lsh+rwCvhPIexhhjwiOURur7RWQproeR4koEg8IdmDHGmMgKdTbXb3HJ4WLgZGBl2CIyxhjTLjRYghCREbhuqJNxA+P+hbuH9UltFJsxxpgIaqyKaRXwEXCuqq4D8BqTjTHGdAGNVTFdiBuw9r6IPC4ip+AaqY0xxnQBDSYIVX1NVS8DRgLvAz8FeonIoyJyelsFaIwxJjKabKRW1VJVfUFVv4MbDb0Y+GXYIzPGGBNRzbontapuV9WpqnpKuAIyxhjTPjQrQRhjjOk6LEEYY4wJyhKEMcaYoCxBGGOMCcoShDHGmKAsQRhjjAnKEoQxxpigLEEYY4wJyhKEMcaYoCxBGGOMCcoShDHGmKAsQRhjjAnKEoQxxpigLEEYY4wJyhKEMcaYoCxBGGOMCcoShDHGmKAsQRhjjAnKEoQxxpigLEEYY4wJKqwJQkTOFJHVIrJORG4Lsn6giLwvIotFZBSbKq0AABy6SURBVJmInB2w7nZvv9UickY44zTGGLO/6HAdWER8wCPAaUAeMF9EZqnqioDN7gBmquqjIjIKmA1kec8vAw4B+gHvisgIVa0NV7zGGGP2Fc4SxJHAOlXdoKpVwAxgUr1tFOjuPU8BtnjPJwEzVLVSVb8B1nnHM8YY00bCmSD6A5sCXud5ywLdBVwuInm40sONzdgXEZkiIgtEZEFBQUFrxd15bPwYXpwMtdWRjsQY0wFFupF6MvC0qmYCZwPPikjIManqVFXNVtXsjIyMsAXZYc1/ElbPhi2LIx2JMaYDCmeC2AwMCHid6S0LdA0wE0BVPwPigfQQ9zWNqa2B9XPd828+jGwsxnR01eVQVRbpKNpcOBPEfGC4iAwWkVhco/OsetvkAqcAiMjBuARR4G13mYjEichgYDjwZRhj7XzyvoSKnSBRsPGjSEdjTMelCs9eAM+c5553IWFLEKpaA9wAzAFW4norfS0i94jIed5mPwd+JCJLgReBK9X5GleyWAG8BVxvPZiaae3bEBUNh06G3C+gpjLSERnTMa2bC7mfQd5897wLEe0kGTE7O1sXLFgQ6TDaj0ePg4SecNR18K/vw1VvwqBjIx2VMR2LKjxxKuz+1j1PyYSr3wKRSEfWakRkoapmB1sX6UZqEw47N8O3y2H46ZB1HCDwjVUzGdNs6+bC5gUw4edw3M2w6XPI+STSUbUZSxCd0bp33M/hp7tSRJ8x1g5hTHOpwrw/QMoAGPd9OPwHkNQLPvxTpCNrM5YgOqM1b0PKQMg4yL0efAJs+sL1xDDGhCaw9BAdCzEJcOwNsGEe5HWN6mxLEOG2+i34/LG2e7+aSvcHPPy0vfWkWROgtgo2WUcwY0KiCh/ct7f0UCf7alcq//DPkYutDVmCCCd/LbxxC7z1S1jwVNu8Z86nUF3qqpfqDDoWxGfVTMaEav1c12uprvRQJy4Zjv4fWPMm5H8VufjaiCWIcNowD0o2u6uQ2f/rupuG29p3wBcHgyfsXRbfHfqNs4ZqY0KhCvOClB7qHPkjiE2Gj/7S9rG1MUsQABUl4Tnu4udccfRH70GPATDzB1Cypen9DsTat11yiE3ad3nWBNi8EKpKw/v+xnR0DZUe6iT0dEni69egYE3bx9eGLEEUb4CHs2HRM6173LJiWPVfGHMJdOsFl73gTs7/uhyqK1r3veoUb4CitftWL9UZPAH81ZD7eXje25jOoKnSQ51jrofoePj4gbaLLQIsQXTPhN6j4fWbYdXs1jvu8ldcw/Bhl7vXvQ6GCx5zV/Gzfx6eIftrve6tw07df92Ao93IamuHMKZhTZUe6iSlQ/ZVsOxfsH1jm4XX1ixBRMfCJc9A33Hw8lWtd4W9+Fk3/qDv2L3LDv4OTPylq3qa/0TrvE+gtW9D2jBIG7r/urhu0H+8tUMY05BQSw91jr0RonzwyUPhjy1CLEGAO3l+/yU3jP6FS2DbygM7Xv5XsHUpHPaD/ddNvA1GnAVv3ebu19BaqsrcyT9Y9VKdrAlu6u/KXa33vsZ0FqGWHup07+dqCBY/F/62xQixBFEnKR0ufxWiE+C578LOvJYfa/Hz4IuFMRfvvy4qCi6cCqlDYOYPYcem/bdpiY0fQW2lG//QkMETQGsh57PWeU9jOovmlh7qHHez687+6d/DF1sEWYII1HMQXP6Ku8J+9kLX0NxcNVWuXvKgsyExNfg28d1do3VtNcz4XuvMM79mDsQkwaDjGt5mwFEucX3zwYG/nzGdSXNLD3V6ZsHYS904p9LCsIUXKZYg6uszGia/6BqeXrik+d1C17wJ5cXBq5cCpQ+H7z7hqqPe+12LwwXc1c/ad2DIiRAd1/B2MQmQeUTrNFTnLYC/HgIvXAqb5h/48YyJpM/+Acn9mld6qDPhFqipcP8L69/vVPeMsAQRTNbx7uS9eSG8dFXz7um8+Dn3hzb0pKa3HXEGHHIBLH3BlTxaqmA17MxtvHqpTtYE2LoMyre3/P3WzIHp33HPN30BT54K089zbSCd6J/DdBHbN8L692D8Fc0rPdRJHw6THnGDYp89H544BVa/2Sn+FyxBNGTUeXDOX2DtHJh1U2i/7JKtsO5dGDfZ9W4IxdhL3cl6/Xstj3Xt2+5nKAli8ARA3ZQcLbHoWXhxMqSPgCnvw0+Xw2m/dQ3708+FaWe60kwn+OcwXcTC6W7esqZK/Y057Ptw81I49wEoLYAXL4PHJsDyV10bRUvVVsP7f4DnLoLib1p+nBayBNGY7KvhxF+5K/y3bmv6pLf0RVB/84qpQ092IzO/mtnyONe+Db0Ocb2wmpJ5hBvg09zurqrwwZ9g1g0wZCJc+V83ADCuGxx3E/x0GZz9Z9e4//xFMHUirHqjZZ+nqyvZGukIwqOqrP2NGaitdqX+EWdCSv8DO1Z0nDtn3LgIzn/MVTu9fBU8chQseaF5NREA21a50sgH97lq4X+2/f+UJYimTPwFHH09fPEYvHtXw0lC1f2hDTw2+DiEhkTHumqmVbOhcnfz46vY6W6HGErpAdwf8YCjmtcO4a+FN34O7//OlXgm/8tNWhYoJsFNP3DTYjjvYdfQP+N7rk7WhO6zf8BfR7b+yP5IK94Aj58EDx0KT58LX/+7+SfM8h0HdjUezKo3oHQbjL+q9Y7pi3G1CNd/ARc/7f7nXvsJ/P1wWDCt6dv/+v3u7+CfJ7gLrkueheu/hLQh7n9qzq+b/921kCWIpojAGfdC9jXwyYPwwf3Bt9v0BRSv3ztyujnGXAw15S27OtgwD/w1rj0jVIMnuDvOlRY1vW11Obx0BSx40nXpO/+xxutpo2PdjVV+8pnrVbXy9dDj6uqWzYQ5t0NUDLx3b8t6t+V+4Up6u7e1fnwt9c2H8PjJsCsfjvspbM+Bl66EB0bD+79veAxBdYWrep3za/jHMXD/IJdktq1qvdgWPuW6tg47pfWOWSfK5y7+rvsYJs9wNxv6789ckvz80eC/3x2b4NlJ7u9g6Enu/2jUea6H5dVz4Mgp8NnD8NTZB9YVP9SPEPZ36AxEXPXJuMvdHaaCzb+y+FmI7QajJjX/+AOOdn+kX73U/H3Xvg1xKZB5ZOj7ZJ3gfuY0MVCvfDs8ewGs/C+ceR+cdo8bxxGKmHj3B772bWuPCMW6ue4qc9Dx8P2ZsDvflVqbo3K3q9J4/3fw4Bh485fu9rORtGCa+xtK6uUmrTztbrh5CXxvpptl4IM/ukTxr8vdxc62VfDZI66b+f2D3L5fToWkDNcFdedmd2X92SPuSvtAFK1373n4FaG3GbaECBx0Flz7LvzgNTfbwVu3ud/RR391k4WqwtIZ8OixsHkRnPd3l1SSe+89TnQcnP0nuOgp1+b32ARY+2744gaiw3r0ziQqCs77m6tXfPcuV49/9E/cusrdbmbHQ853dfItOfbo77rBNrsLoFtGaPvVdW8ddjL4mvGr7H+4u7r/5qOGE9qGee5qZ2ceXPSki6+5hp/uJizctgJ6H9L8/buKzQvhXz+AjINh8gsQnwLDz4CPH4TxVzY8nqa+D//oetJc+Dhs+MBN5zL/SRj3PTj+Z5A6OKwfYx+1Ne4q+MupMOw09zcUn+LWRflciXfEGa7hdeFTrvNDYGkzfYT77ENPcfdVr5ud+Kjr3Lxpc37legqd/w/oMbBlMS582t0npSWl/pYQcRdNQ09yg1U/+jPMvdvVTPQZ66p9Bx4D5z/a+O9q9IVu+5eucO19E34OJ97evHNAqFS1UzzGjx+vbaKmWnXG91Xv7K46/0m3bNFz7nXOZy0/bv5yd4wvpoa+T+4Xbp/Fzzf//Z65QPXhI/dfvmub6is/csd98FDVjZ80/9h1dm5xx/nwzy0/RmdXsFb1/sGqD4xWLdm6d3n+ctU7U1Tn/Dq042xbpXp3qupr/7N3WfFG1dd/pnpPuupdPVVfmeK2C7eyYtXp57nf/Vu/Uq2taXqfqnLVpTNVFzytuj2n8W39ftVFz6re21/195nu79/vb16M1RXue3/xe83br7VtXuRi+MNA1Y8eCO27qlNVpvra9e57nj5Jtba2RSEAC7SB86poJyn+Z2dn64IFbXSf2JoqVyReO8dl+0XPuoauGxbsvc1nS/zjGFdNde07TW+r6sYifPu1K7LXXZ2F6uMHXEno52tcMdbvh8XPwDt3usGBx//UXZnEJLToo+zxz4mutHXNnAM7Tme0Kx+ePM1931e/DenD9l3/7+tcN8mbFjXeQ63ubyH/K7hxoZs2JlDJVldvvWCaa1M67iZXXRgOBWtcF88dufCdB8N7db49x1XL5XwCI8+F7zy0/2dvyFcvwyvXuJkTgs1+3JEsecFVBx9zfYt2F5GFqpodbJ1VMbVE3QywL14K/7nedW095c4DSw4AYy6Cufe4YndT1QFr5rgi6Vl/an5ygL3tEBs/gl6jXHXSps/dVB3nPgAZBzX/mMGMOAM+/JNrEE9Ka51jdgYVO92cX6VFcOXr+ycHgJN+5aaNn/cHNxCrIctfcb/Hc/4a/ATZva/raHH8z+Dt/3Ozj/Y73FWJNkfJVph9K5Q10rnh26/ddC5XvA6Djmne8Zur5yC44r/w+SPu/+YfR7sedpnjm9534dPQYxAMOTm8MbaFcd8L26GtkbqlYuLhshddnaEvFg6dfODHHH2R+7n85ca3q62Gd/7PNXZlt7B7Xt9D3W0T378X/jkBCtfApH/AlW+0XnIAlyDU7wYQtheF6+C/t7TeRInNVV0BL34PClbBpc+6adiD6TEQjviRu0JsqOdORYnr5dPvMFdn35ikdNeO1u9weP2m5n3+mkp3R8T177lunA09hp7kGqPDnRzqREW5abenfODaKZ6/CArXNr5P4VqXUMdfEXqniy7KShAHIjYRfvBvV1XQve+BH6/nINejadlLMOHWhkski6a7E/plL7h/ypbwRcPgE2D1G6531mn3hOcKv+9hrgfLmrfg0Etb//jNtfxVmHUjVO12U59f/Vbj81c11658Nx6msQFhhWtct+gLn2i6e+WEn7secnPvcQ3Y9c27D3Z/69aF0hPHF+MajB+bAK/+yF2BN9W4qerGweTNh4unN7/k0RZ6j3L/i0+e7no+XfO2m447mIVPu5tnjWujxukOzNLngYqOcyf21jL2Yihc7cYpBFNR4obeDzrezRh7IM59AK77BM5/JHzVP1FRrjfT+rltNrgnqJpKmP2/rhtor1Gu2/KWRfD2HQd+bFV3b4+XroQHDoH3futKTOvmBn+UbHHTuIwNMh18fUlprs1g9RtujEOgb792XWHHX9FwKSSY1CGuOirX60nTlAVPuiQ14db2mRzqpA6B77/s6uOf+64bWFdfdQUseR5GnrNvF1ITlJUg2ptRF7j+68tmujvS1ffxX6GsEE7/7YG3eST3bpt/khFnwJLn3FVz1vHhf7/6tm90J+8ti+GYG+DUu9yV9PaNrvF24NEt68ZbUeKmdp//hKsuiu/humFmX9280fRNOfp/4MvH4d074ao33e9dFd641bU/nXJn84956KUuaX9wPwye2HCVUM6n7u9x+BmuTaS96zcOLnvezV304mT4wav7drRYOcslkNYcOd2JhbUEISJnishqEVknIrcFWf+AiCzxHmtEZEfAutqAdbPCGWe7kpTm+n4vf2X/gUA7Nrkh+GMvdWMZOoqhJ7nRwWveavv3XjXbDawq2gCXPu8aa+uq5U69y007MuumpuutA23f6Br1/zLSNdpGx7tG5FtWuuO3ZnIAV7c+8Rfuin+N1xts6QzI/dR9hlDHSdR39p9dQ+0r1waf3XdnnrupVc8sd5OrcA4ma01DTnTx5n7mPlvg9BwLnoKeg11SNE0KW4IQER/wCHAWMAqYLCKjArdR1Z+p6jhVHQf8HXg1YHV53TpVPS9ccbZLYy52A55y6824Ovced/V48v9FJq6Wikt2g53WvN1271lb7XrszJjsTnA//gAOPnffbXwxblRqdJw7EYYytcWK/8Cjx7u7Bh5yvmuQ/fEHrjtnbGJYPgrgRvumDnFdk8uKXSeF/tkHNgNpfHfXHrE7f/8Zi6vLYcb3XZXMZS9AQo8D/ghtavSFcNb9bqDmG7e4z7ZtlfufGn+lNU6HKJzf0pHAOlXdoKpVwAygsXkoJgMvhjGejmPk2RCTuO/UG5sXuhlfj/4f6DEgcrG11IgzXdtK8YbwvYeqq5f/5G9uFsxP/wZHXOvGGDTUbTilvxt5vG2lKw00pKYK3rzNJZKMg+DGBW4Ub3Pq/g+EL8ZdGBSsdPPwlBW5dowDPdH1H++Ou3KW6/wA7nv8789g6xJ3Jd6avdra0lE/du0mC592XYUXPu1Ksi25KVAXFc42iP5AYD+6POCoYBuKyCBgMBB4U4R4EVkA1AD3qeprQfabAkwBGDiwhcPt26PYJNeI9vVrbpyDL8ZdDSdluL7sHdGIM9z8M2vehqOva73jlha6aUHWzXVdMHfnu+UZB8N3n3RjS5oy7BRXhfPB/a7b8uH1rsp35Lo2jM0LXYI+9e6W3VjmQI06H/r9zbWlHPEjV9/eGo69CTa87xLgwGPc97j0RTjp1+5ipSM7+Q7Xy+uD+1139IO/E/pUNqbdNFJfBrysqoFz+Q5S1c0iMgR4T0S+UtX1gTup6lRgKriR1G0XbhsYc4krQax7140jyPnE9TyJ7x7pyFomdQikDXejzw8kQdRUue6W670eQVuXAuruqTHkJHd/jaEnN39u/4m/dI3os291J966DgJr5sCrU9zv4JJnWjYZY2uJinJ/A588BCf/unWPe8E/3URxL1zi2rpGnuuuvjs6ETj3QXchseZN14HAhCycCWIzEFgXkuktC+YyYJ9x4qq62fu5QUTmAYcB6/fftZMaehIkprkuedtWQvpBrh66Ixtxhpu8rXLX/veTaEzRendVu/49N3V01W43yVrmEa5nzdBT3En9QBpRo3xuXMI/J7hqpGvnuiqqjx9wyeLi6a3f+NwS/Q+HS6a3/nGT+7hpY164BDJGwgWPdZ56el+0+87yl4c2ytrsEc4EMR8YLiKDcYnhMmC/MeEiMhLoCXwWsKwnUKaqlSKSDhwH/DGMsbY/vhg3l/z8J9zr780Mz2yNbWnEma5b6YZ5rqjfmG0rXTJZ/97eQWc9BsHYS1wJYfAJLZtipDHdMlyj9dPnwEPjoHKna9A8874Dn5OqIxhxBvzwP656rjkJvCOIjrPk0AJhO+Ooao2I3ADMAXzANFX9WkTuwc0eWNd19TJghu47a+DBwD9FxI9rSL9PVVeEK9Z2a8zFLkEMnugGm3V0A492965Y81bjCaJwrWuIral0ieDo6107QeqQAx/70ZRBx7iuqvP+ABdMbR+jv9vSkBMjHYFpR2w21/ZM1d0jYtR5rqtmZ/DSlW7w1S2rgldhlGxx0yXUVLjpElKHtHmIgBuD0lmqWIxpRGOzudp/QHsm4qZZ6CzJAVw10+5vXRfK+sp3eFMkbHdTJkQqOYAlB2OwBGHa2rBTAdk7IrhOdbmbGqFwrZsqobW6cBpjWswShGlbSemu99HagARRW+OmRMj9zA3MGnJipKIzxgSwBGHa3ogz3GCvXfnepHO3uCkRzrrfTZFgjGkXLEGYtjfiDPdz7dvw/u/dFA8TbnVTIxhj2o0O3rHedEi9R0P3/vDevW5qjMN+4KZEMMa0K1aCMG1PxJUidue7mx6d+2D4xzcYY5rNShAmMo69yc2fdML/dvwR4sZ0UvafaSIjdTCc8ptIR2GMaYRVMRljjAnKEoQxxpigLEEYY4wJyhKEMcaYoCxBGGOMCcoShDHGmKAsQRhjjAnKEoQxxpigOs0d5USkAMg5gEOkA4WtFE5rs9haxmJrGYutZTpqbINUNSPYik6TIA6UiCxo6LZ7kWaxtYzF1jIWW8t0xtisiskYY0xQliCMMcYEZQlir6mRDqARFlvLWGwtY7G1TKeLzdogjDHGBGUlCGOMMUFZgjDGGBNUl08QInKmiKwWkXUicluk4wkkIhtF5CsRWSIiC9pBPNNEZJuILA9Ylioi74jIWu9nz3YS110istn77paIyNltHZcXxwAReV9EVojI1yJys7e8PXxvDcUW8e9OROJF5EsRWerFdre3fLCIfOH9v/5LRGLbUWxPi8g3Ad/buLaOLSBGn4gsFpH/eq9b9r2papd9AD5gPTAEiAWWAqMiHVdAfBuB9EjHERDPCcDhwPKAZX8EbvOe3wbc307iugu4tR18Z32Bw73nycAaYFQ7+d4aii3i3x0gQDfveQzwBXA0MBO4zFv+GPCTdhTb08BFkf6b8+K6BXgB+K/3ukXfW1cvQRwJrFPVDapaBcwAJkU4pnZLVT8EiustngRM955PB85v06BoMK52QVW3quoi7/kuYCXQn/bxvTUUW8Sps9t7GeM9FDgZeNlbHqnvraHY2gURyQTOAZ7wXgst/N66eoLoD2wKeJ1HO/kH8SjwtogsFJEpkQ6mAb1Vdav3PB/oHclg6rlBRJZ5VVBtXoVTn4hkAYfhrjjb1fdWLzZoB9+dV02yBNgGvIMr7e9Q1Rpvk4j9v9aPTVXrvrd7ve/tARGJi0RswIPALwC/9zqNFn5vXT1BtHfHq+rhwFnA9SJyQqQDaoy68mt7uZJ6FBgKjAO2An+JZDAi0g14BfipqpYErov09xYktnbx3alqraqOAzJxpf2RkYgjmPqxicho4HZcjEcAqcAv2zouETkX2KaqC1vjeF09QWwGBgS8zvSWtQuqutn7uQ34N+6fpL35VkT6Ang/t0U4HgBU9Vvvn9gPPE4EvzsRicGdgJ9X1Ve9xe3iewsWW3v67rx4dgDvA8cAPUQk2lsV8f/XgNjO9KrsVFUrgaeIzPd2HHCeiGzEVZmfDDxEC7+3rp4g5gPDvRb+WOAyYFaEYwJARJJEJLnuOXA6sLzxvSJiFnCF9/wK4D8RjGWPupOv5wIi9N159b9PAitV9a8BqyL+vTUUW3v47kQkQ0R6eM8TgNNwbSTvAxd5m0XqewsW26qAhC+4Ov42/95U9XZVzVTVLNz57D1V/T4t/d4i3doe6QdwNq73xnrg15GOJyCuIbheVUuBr9tDbMCLuCqHalw95jW4+s25wFrgXSC1ncT1LPAVsAx3Mu4boe/seFz10TJgifc4u518bw3FFvHvDhgLLPZiWA78xls+BPgSWAe8BMS1o9je87635cBzeD2dIvUATmRvL6YWfW821YYxxpigunoVkzHGmAZYgjDGGBOUJQhjjDFBWYIwxhgTlCUIY4wxQVmCMKYZRKQ2YLbOJdKKMwCLSFbgjLTGRFp005sYYwKUq5tiwZhOz0oQxrQCcffu+KO4+3d8KSLDvOVZIvKeN4HbXBEZ6C3vLSL/9u4psFREjvUO5RORx737DLztjdQ1JiIsQRjTPAn1qpguDVi3U1XHAA/jZtQE+DswXVXHAs8Df/OW/w34QFUPxd3L4mtv+XDgEVU9BNgBfDfMn8eYBtlIamOaQUR2q2q3IMs3Aier6gZvArx8VU0TkULcVBXV3vKtqpouIgVAprqJ3eqOkYWbOnq49/qXQIyq/i78n8yY/VkJwpjWow08b47KgOe1WDuhiSBLEMa0nksDfn7mPf8UN6smwPeBj7znc4GfwJ6bz6S0VZDGhMquToxpngTvTmJ13lLVuq6uPUVkGa4UMNlbdiPwlIj8L1AAXOUtvxmYKiLX4EoKP8HNSGtMu2FtEMa0Aq8NIltVCyMdizGtxaqYjDHGBGUlCGOMMUFZCcIYY0xQliCMMcYEZQnCGGNMUJYgjDHGBGUJwhhjTFD/D4I2jqGRr1UEAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Maximum validation accuracy:  0.7989803346420672\n",
      "Training accuracy of best model:  0.9952993406019456\n"
     ]
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Plot training & validation accuracy values\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "print(np.argmax(np.array(history.history['val_acc'])))\n",
    "plt.title('Model accuracy')\n",
    "plt.ylabel('Accuracy')\n",
    "plt.xlabel('Epoch')\n",
    "plt.legend(['Train', 'Test'], loc='upper left')\n",
    "plt.show()\n",
    "print('Maximum validation accuracy: ',np.max(np.array(history.history['val_acc'])))\n",
    "print('Training accuracy of best model: ',np.array(history.history['acc'])[np.argmax(np.array(history.history['val_acc']))])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 180,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[250  98]\n",
      " [115 871]]\n",
      "the mean-f1 score: 0.80\n",
      "accuracy is: 0.84\n"
     ]
    }
   ],
   "source": [
    "model = load_model(model_name)\n",
    "y_pred = np.argmax(model.predict(X_test), axis=1)\n",
    "y_true = np.argmax(y_test, axis=1)\n",
    "cf_matrix = confusion_matrix(y_true, y_pred)\n",
    "print(cf_matrix)\n",
    "class_wise_f1 = np.round(f1_score(y_true, y_pred, average=None)*100)*0.01\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "\n",
    "\n",
    "print('the mean-f1 score: {:.2f}'.format(np.mean(class_wise_f1)))\n",
    "print('accuracy is: {:.2f}'.format(accuracy))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 181,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total_count of test files: 50\n",
      "correctly predicted test files: 44\n",
      "Percentage accuracy: 88.0\n",
      "total_count of train files: 188\n",
      "correctly predicted train files: 186\n",
      "Percentage accuracy: 98.93617021276596\n",
      "total_count of val files: 50\n",
      "correctly predicted val files: 43\n",
      "Percentage accuracy: 86.0\n"
     ]
    }
   ],
   "source": [
    "# Test data accuracy:\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_test, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "# Train accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_train, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of train files:', total_count)\n",
    "print('correctly predicted train files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)\n",
    "#val accuracy\n",
    "total_count, correct_count = get_prediction_accuracy(data_dir_val, sub_dirs,pos_lab,neg_lab,labelfile,model)\n",
    "print('total_count of val files:', total_count)\n",
    "print('correctly predicted val files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 182,
   "metadata": {},
   "outputs": [],
   "source": [
    "del train_x, train_y, train_files, val_x, val_y, val_files, test_x, test_y,test_files\n",
    "del train_x2, train_y2, all_counts\n",
    "del train_x3, train_y3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 199,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_dir_test='test_micro_ov/'\n",
    "labelfile = 'LabelTable.csv'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 200,
   "metadata": {},
   "outputs": [],
   "source": [
    "modelcut = load_model('DCBL_micro_Cut_jitter.hdf5')\n",
    "modelwash = load_model('DCBL_micro_Wash_jitter.hdf5')\n",
    "modeltake = load_model('DCBL_micro_Take_jitter.hdf5')\n",
    "modelput = load_model('DCBL_micro_Put_jitter.hdf5')\n",
    "modelpeel = load_model('DCBL_micro_Peel_jitter.hdf5')\n",
    "modeladd = load_model('DCBL_micro_ADD_jitter.hdf5')\n",
    "modelmix = load_model('DCBL_micro_Mix_jitter.hdf5')\n",
    "modelpour = load_model('DCBL_micro_Pour_jitter.hdf5')\n",
    "modelopen = load_model('DCBL_micro_Open_jitter.hdf5')\n",
    "modelother = load_model('DCBL_micro_other_jitter.hdf5')\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 211,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_test_prediction_accuracy(data_dir, sub_dirs,labelfile,modelcut,modelwash,modeltake,modelput,modelpeel,modeladd,modelmix,modelpour,modelopen,modelother):\n",
    "\n",
    "    files = os.listdir(data_dir+'/left_hip')\n",
    "    number_of_samples = 500\n",
    "\n",
    "    labels_micro =dict()\n",
    "    \n",
    "    labels_micro['Cut'] = 0\n",
    "    labels_micro['Wash'] = 1\n",
    "    labels_micro['Take'] = 2\n",
    "    labels_micro['Put'] = 3\n",
    "    labels_micro['Peel'] = 4\n",
    "    labels_micro['Add'] = 5\n",
    "    labels_micro['Mix'] = 6\n",
    "    labels_micro['Pour'] = 7\n",
    "    labels_micro['Open'] = 8\n",
    "    labels_micro['other'] = 9 \n",
    "    \n",
    "    \n",
    "    #read the labels\n",
    "    labels_loc = labelfile\n",
    "    file_label = open(labels_loc, newline='')\n",
    "    label_reader = csv.reader(file_label)\n",
    "    file_label_mapping = dict()\n",
    "    \n",
    "    for row in label_reader:\n",
    "        label = row[2].split(',')\n",
    "        \n",
    "        label_file_micro = [0]*10\n",
    "        for lab in label:\n",
    "            if lab!='' and lab !='micro':\n",
    "                label_file_micro[labels_micro[lab]] = 1\n",
    "                \n",
    "        \n",
    "        file_label_mapping[row[0]+'.csv'] = label_file_micro\n",
    "\n",
    "\n",
    "    total_count = 0\n",
    "    correct_count = 0\n",
    "\n",
    "    for f in files:\n",
    "\n",
    "        file_data = []\n",
    "        file_label = []\n",
    "\n",
    "        st_index = 0\n",
    "        end_index = 30000\n",
    "        step = 1000 #overlapping window, step: 1000. \n",
    "        window_index = 10000 #10 second window\n",
    "\n",
    "        #print('reading file:',f)\n",
    "        f_name = f\n",
    "\n",
    "        if f_name == '.DS_Store':\n",
    "            continue\n",
    "\n",
    "        total_count = total_count+1\n",
    "\n",
    "        curr_label_file = file_label_mapping[f_name]\n",
    "\n",
    "        while st_index+step < end_index:\n",
    "\n",
    "            data, data_count = parse_IMU_files(data_dir, sub_dirs, st_index, st_index+window_index,  f, number_of_samples)\n",
    "            st_index = st_index+step\n",
    "\n",
    "            if data_count<data_min_count:\n",
    "                continue\n",
    "\n",
    "            train_data_sample  = np.zeros((9, number_of_samples))\n",
    "            train_data_label   = curr_label_file\n",
    "            for i in range(len(data)):\n",
    "                for j in range(len(data[i])):\n",
    "                    train_data_sample[i*3,j]=data[i][j][0]\n",
    "                    train_data_sample[i*3+1,j]=data[i][j][1]\n",
    "                    train_data_sample[i*3+2,j]=data[i][j][2]\n",
    "\n",
    "            file_data.append(train_data_sample)\n",
    "            #file_label.append(train_data_label)\n",
    "\n",
    "        file_data = np.array(file_data)\n",
    "        file_label = curr_label_file\n",
    "\n",
    "        file_data = file_data.reshape((-1, 9,500, 1))\n",
    "\n",
    "        #y_pred = np.argmax(model.predict(file_data), axis=1) \n",
    "        \n",
    "        y_pred_cut = modelcut.predict(file_data)\n",
    "     \n",
    "        y_pred_wash = modelwash.predict(file_data)\n",
    "\n",
    "        y_pred_take = modeltake.predict(file_data)\n",
    "\n",
    "        y_pred_put = modelput.predict(file_data)\n",
    "\n",
    "        y_pred_peel = modelpeel.predict(file_data)\n",
    "  \n",
    "        y_pred_add = modeladd.predict(file_data)\n",
    " \n",
    "        y_pred_mix = modelmix.predict(file_data)\n",
    "    \n",
    "        y_pred_pour = modelpour.predict(file_data)\n",
    "  \n",
    "        y_pred_open = modelopen.predict(file_data)\n",
    "    \n",
    "        y_pred_other = modelother.predict(file_data)\n",
    "\n",
    "\n",
    "        \n",
    "        y_pred = np.zeros((file_data.shape[0],10))\n",
    "        \n",
    "        for j in range(0,y_pred.shape[0]):\n",
    "            if(y_pred_cut[j,0]>=y_pred_cut[j,1]):\n",
    "                y_pred[j,0] = 1\n",
    "            if(y_pred_wash[j,0]>=y_pred_wash[j,1]):\n",
    "                y_pred[j,1] = 1\n",
    "            if(y_pred_take[j,0]>=y_pred_take[j,1]):\n",
    "                y_pred[j,2] = 1\n",
    "            if(y_pred_put[j,0]>=y_pred_put[j,1]):\n",
    "                y_pred[j,3] = 1\n",
    "            if(y_pred_peel[j,0]>=y_pred_peel[j,1]):\n",
    "                y_pred[j,4] = 1\n",
    "            if(y_pred_add[j,0]>=y_pred_add[j,1]):\n",
    "                y_pred[j,5] = 1\n",
    "            if(y_pred_mix[j,0]>=y_pred_mix[j,1]):\n",
    "                y_pred[j,6] = 1\n",
    "            if(y_pred_pour[j,0]>=y_pred_pour[j,1]):\n",
    "                y_pred[j,7] = 1\n",
    "            if(y_pred_open[j,0]>=y_pred_open[j,1]):\n",
    "                y_pred[j,8] = 1\n",
    "            if(y_pred_other[j,0]>=y_pred_other[j,1]):\n",
    "                y_pred[j,9] = 1\n",
    "        \n",
    "        #print(y_pred.shape)\n",
    "        \n",
    "        #print(y_pred)\n",
    "        #y_pred = np.where(y_pred > 0, 1, 0)\n",
    "        \n",
    "        loc_arr = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0])\n",
    "        \n",
    "        \n",
    "        for i in range(y_pred.shape[0]):\n",
    "            loc_arr = loc_arr + y_pred[i]\n",
    "        \n",
    "        \n",
    "        loc_arr = loc_arr/y_pred.shape[0]\n",
    "        \n",
    "        max_val = loc_arr.max()\n",
    "        \n",
    "        loc_arr = np.where(loc_arr > max_val/3.0, 1, 0)\n",
    "        \n",
    "        #print(loc_arr)\n",
    "        \n",
    "        file_label = np.array(file_label)\n",
    "        \n",
    "        #print(file_label)\n",
    "\n",
    "        #correct prediction\n",
    "        \n",
    "        if np.array_equal(file_label, loc_arr):\n",
    "            correct_count = correct_count+1\n",
    "        \n",
    "        #print('*'*10)\n",
    "        #break\n",
    "    \n",
    "    return total_count, correct_count"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "# Test data accuracy:\n",
    "total_count, correct_count = get_test_prediction_accuracy(data_dir_test, sub_dirs,labelfile,modelcut,modelwash,modeltake,modelput,modelpeel,modeladd,modelmix,modelpour,modelopen,modelother)\n",
    "print('total_count of test files:', total_count)\n",
    "print('correctly predicted test files:', correct_count)\n",
    "print('Percentage accuracy:', (correct_count/total_count)*100.0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "colab": {
   "collapsed_sections": [],
   "name": "Macro-Activity",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}