normalize.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Normalization Script\n",
    "\n",
    "Author: Kristina Preuer\n",
    "\n",
    "This script shows how the data was split and how the features were normalized. The data is then saved in a pickle file. Which will be loaded during the cross validation procedure."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "import pickle \n",
    "import gzip"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "##### Define the parameters for data generation: folds for testing and validation and normalization strategy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# in this example tanh normalization is used\n",
    "# fold 0 is used for testing and fold 1 for validation (hyperparamter selection)\n",
    "norm = 'tanh'\n",
    "test_fold = 0\n",
    "val_fold = 1"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Define nomalization function\n",
    "It normalizes the input data X. If X is used for training the mean and the standard deviation is calculated during normalization. If X is used for validation or testing, the previously calculated mean and standard deviation of the training data should be used. If \"tanh_norm\" is used as normalization strategy, then the mean and standard deviation are calculated twice. The features with a standard deviation of 0 are filtered out. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def normalize(X, means1=None, std1=None, means2=None, std2=None, feat_filt=None, norm='tanh_norm'):\n",
    "    if std1 is None:\n",
    "        std1 = np.nanstd(X, axis=0)\n",
    "    if feat_filt is None:\n",
    "        feat_filt = std1!=0\n",
    "    X = X[:,feat_filt]\n",
    "    X = np.ascontiguousarray(X)\n",
    "    if means1 is None:\n",
    "        means1 = np.mean(X, axis=0)\n",
    "    X = (X-means1)/std1[feat_filt]\n",
    "    if norm == 'norm':\n",
    "        return(X, means1, std1, feat_filt)\n",
    "    elif norm == 'tanh':\n",
    "        return(np.tanh(X), means1, std1, feat_filt)\n",
    "    elif norm == 'tanh_norm':\n",
    "        X = np.tanh(X)\n",
    "        if means2 is None:\n",
    "            means2 = np.mean(X, axis=0)\n",
    "        if std2 is None:\n",
    "            std2 = np.std(X, axis=0)\n",
    "        X = (X-means2)/std2\n",
    "        X[:,std2==0]=0\n",
    "        return(X, means1, std1, means2, std2, feat_filt)        "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Load features and labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "#contains the data in both feature ordering ways (drug A - drug B - cell line and drug B - drug A - cell line)\n",
    "#in the first half of the data the features are ordered (drug A - drug B - cell line)\n",
    "#in the second half of the data the features are ordered (drug B - drug A - cell line)\n",
    "file = gzip.open('X.p.gz', 'rb')\n",
    "X = pickle.load(file)\n",
    "file.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "#contains synergy values and fold split (numbers 0-4)\n",
    "labels = pd.read_csv('labels.csv', index_col=0) \n",
    "#labels are duplicated for the two different ways of ordering in the data\n",
    "labels = pd.concat([labels, labels]) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Define indices for splitting"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "#indices of training data for hyperparameter selection: fold 2, 3, 4\n",
    "idx_tr = np.where(np.logical_and(labels['fold']!=test_fold, labels['fold']!=val_fold))\n",
    "#indices of validation data for hyperparameter selection: fold 1\n",
    "idx_val = np.where(labels['fold']==val_fold)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "#indices of training data for model testing: fold 1, 2, 3, 4\n",
    "idx_train = np.where(labels['fold']!=test_fold)\n",
    "#indices of test data for model testing: fold 0\n",
    "idx_test = np.where(labels['fold']==test_fold)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Split data "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_tr = X[idx_tr]\n",
    "X_val = X[idx_val]\n",
    "X_train = X[idx_train]\n",
    "X_test = X[idx_test]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_tr = labels.iloc[idx_tr]['synergy'].values\n",
    "y_val = labels.iloc[idx_val]['synergy'].values\n",
    "y_train = labels.iloc[idx_train]['synergy'].values\n",
    "y_test = labels.iloc[idx_test]['synergy'].values"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Normalize training and validation data for hyperparameter selection"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "if norm == \"tanh_norm\":\n",
    "    X_tr, mean, std, mean2, std2, feat_filt = normalize(X_tr, norm=norm)\n",
    "    X_val, mean, std, mean2, std2, feat_filt = normalize(X_val, mean, std, mean2, std2, \n",
    "                                                          feat_filt=feat_filt, norm=norm)\n",
    "else:\n",
    "    X_tr, mean, std, feat_filt = normalize(X_tr, norm=norm)\n",
    "    X_val, mean, std, feat_filt = normalize(X_val, mean, std, feat_filt=feat_filt, norm=norm)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Normalize training and test data for methods comparison"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "if norm == \"tanh_norm\":\n",
    "    X_train, mean, std, mean2, std2, feat_filt = normalize(X_train, norm=norm)\n",
    "    X_test, mean, std, mean2, std2, feat_filt = normalize(X_test, mean, std, mean2, std2, \n",
    "                                                          feat_filt=feat_filt, norm=norm)\n",
    "else:\n",
    "    X_train, mean, std, feat_filt = normalize(X_train, norm=norm)\n",
    "    X_test, mean, std, feat_filt = normalize(X_test, mean, std, feat_filt=feat_filt, norm=norm)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Save data as pickle file"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "pickle.dump((X_tr, X_val, X_train, X_test, y_tr, y_val, y_train, y_test), \n",
    "            open('data_test_fold%d_%s.p'%(test_fold, norm), 'wb'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "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": "3.6.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}