Code/self-supervised.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:06.900565Z",
     "iopub.status.busy": "2022-11-23T14:21:06.900057Z",
     "iopub.status.idle": "2022-11-23T14:21:06.910185Z",
     "shell.execute_reply": "2022-11-23T14:21:06.909088Z",
     "shell.execute_reply.started": "2022-11-23T14:21:06.900521Z"
    }
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import accuracy_score"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Data Loader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:07.946331Z",
     "iopub.status.busy": "2022-11-23T14:21:07.945966Z",
     "iopub.status.idle": "2022-11-23T14:21:07.955563Z",
     "shell.execute_reply": "2022-11-23T14:21:07.954368Z",
     "shell.execute_reply.started": "2022-11-23T14:21:07.946279Z"
    }
   },
   "outputs": [],
   "source": [
    "## Data Generator\n",
    "class DataGenerator(tf.keras.utils.Sequence):\n",
    "    'Generates data for Keras'\n",
    "    def __init__(self, xData, yData, batch_size=16, shuffle=True):\n",
    "        'Initialization'\n",
    "        self.batch_size = batch_size\n",
    "        self.X = xData\n",
    "        self.Y = yData\n",
    "        self.shuffle = shuffle\n",
    "        self.on_epoch_end()\n",
    "\n",
    "    def __len__(self):\n",
    "        'Denotes the number of batches per epoch'\n",
    "        return int(np.floor(len(self.X) / self.batch_size))\n",
    "\n",
    "    def __getitem__(self, index):\n",
    "        'Generate one batch of data'\n",
    "        # Generate indexes of the batch\n",
    "        indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]\n",
    "        return self.X[indexes]/255,self.Y[indexes]\n",
    "\n",
    "    def on_epoch_end(self):\n",
    "        'Updates indexes after each epoch'\n",
    "        self.indexes = np.arange(len(self.X))\n",
    "        if self.shuffle == True:\n",
    "            np.random.shuffle(self.indexes)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:09.160161Z",
     "iopub.status.busy": "2022-11-23T14:21:09.159802Z",
     "iopub.status.idle": "2022-11-23T14:21:09.176967Z",
     "shell.execute_reply": "2022-11-23T14:21:09.176032Z",
     "shell.execute_reply.started": "2022-11-23T14:21:09.160129Z"
    }
   },
   "outputs": [],
   "source": [
    "def ResnetBlock(in_channels, out_channels, down_sample = False):\n",
    "    inputs = tf.keras.Input((None, None, in_channels)); # inputs.shape = (batch, height, width, in_channels)\n",
    "    if down_sample:\n",
    "        shortcut = tf.keras.layers.Conv2D(out_channels, kernel_size = (1,1), strides = (2,2), padding = 'same', kernel_initializer = tf.keras.initializers.HeNormal())(inputs);\n",
    "        shortcut = tf.keras.layers.BatchNormalization()(shortcut);\n",
    "    else:\n",
    "        shortcut = inputs;\n",
    "    results = tf.keras.layers.Conv2D(out_channels, kernel_size = (3,3), strides = (2,2) if down_sample else (1,1), padding = 'same', kernel_initializer = tf.keras.initializers.HeNormal())(inputs);\n",
    "    results = tf.keras.layers.BatchNormalization()(results);\n",
    "    results = tf.keras.layers.ReLU()(results);\n",
    "    results = tf.keras.layers.Conv2D(out_channels, kernel_size = (3,3), strides = (1,1), padding = 'same', kernel_initializer = tf.keras.initializers.HeNormal())(results);\n",
    "    results = tf.keras.layers.BatchNormalization()(results);\n",
    "    results = tf.keras.layers.Add()([results, shortcut]);\n",
    "    results = tf.keras.layers.ReLU()(results);\n",
    "    return tf.keras.Model(inputs = inputs, outputs = results);\n",
    "\n",
    "def ResNet18(**kwargs):\n",
    "    inputs = tf.keras.Input((None, None, 3));\n",
    "    results = tf.keras.layers.Conv2D(64, kernel_size = (7,7), strides = (2,2), padding = 'same', kernel_initializer = tf.keras.initializers.HeNormal())(inputs);\n",
    "    results = tf.keras.layers.BatchNormalization()(results);\n",
    "    results = tf.keras.layers.ReLU()(results);\n",
    "    results = tf.keras.layers.MaxPool2D(pool_size = (3,3), strides = (2,2), padding = 'same')(results);\n",
    "    results = ResnetBlock(64, 64)(results);\n",
    "    results = ResnetBlock(64, 64)(results);\n",
    "    results = ResnetBlock(64, 128, down_sample = True)(results);\n",
    "    results = ResnetBlock(128, 128)(results);\n",
    "    results = ResnetBlock(128, 256, down_sample = True)(results);\n",
    "    results = ResnetBlock(256, 256)(results);\n",
    "    results = ResnetBlock(256, 512, down_sample = True)(results);\n",
    "    results = ResnetBlock(512, 512)(results);\n",
    "    results = tf.keras.layers.GlobalAveragePooling2D()(results); # results.shape = (batch, 512)\n",
    "    return tf.keras.Model(inputs = inputs, outputs = results, **kwargs);\n",
    "\n",
    "def Model(baseModel, nClasses):\n",
    "    inputs = tf.keras.Input((None, None, 3));\n",
    "    results = baseModel(inputs);\n",
    "    results = tf.keras.layers.Dense(128,activation = tf.keras.activations.relu)(results);\n",
    "#     results = tf.keras.layers.Dense(32,activation = tf.keras.activations.relu)(results);\n",
    "    results = tf.keras.layers.Dense(nClasses, activation = tf.keras.activations.softmax)(results);\n",
    "    return tf.keras.Model(inputs = inputs, outputs = results);"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Load the Dataset & Dataloader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:10.631501Z",
     "iopub.status.busy": "2022-11-23T14:21:10.631123Z",
     "iopub.status.idle": "2022-11-23T14:21:11.637009Z",
     "shell.execute_reply": "2022-11-23T14:21:11.636045Z",
     "shell.execute_reply.started": "2022-11-23T14:21:10.631469Z"
    }
   },
   "outputs": [],
   "source": [
    "(xTrain, yTrain), (xTest, yTest) = tf.keras.datasets.cifar10.load_data()\n",
    "xSetA, xSetB, ySetA, ySetB = train_test_split(xTrain, yTrain, stratify= yTrain, test_size=0.2)\n",
    "trainRotDataloader = DataGenerator(xSetA, ySetA, batch_size = 128, shuffle = True)\n",
    "trainClfDataloader = DataGenerator(xSetB, ySetB, batch_size = 512, shuffle = True)\n",
    "testClfDataloader  = DataGenerator(xTest, yTest, batch_size = 500, shuffle = False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-19T12:13:57.301105Z",
     "iopub.status.busy": "2022-11-19T12:13:57.300693Z",
     "iopub.status.idle": "2022-11-19T12:13:57.308789Z",
     "shell.execute_reply": "2022-11-19T12:13:57.307619Z",
     "shell.execute_reply.started": "2022-11-19T12:13:57.301070Z"
    }
   },
   "source": [
    "#### Load Rotational Models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:20.714947Z",
     "iopub.status.busy": "2022-11-23T14:21:20.714549Z",
     "iopub.status.idle": "2022-11-23T14:21:21.634565Z",
     "shell.execute_reply": "2022-11-23T14:21:21.633570Z",
     "shell.execute_reply.started": "2022-11-23T14:21:20.714913Z"
    }
   },
   "outputs": [],
   "source": [
    "baseResNetModel = ResNet18()\n",
    "RotationModel = Model(baseResNetModel, nClasses = 4)\n",
    "optimizer = tf.keras.optimizers.SGD()\n",
    "loss_fn = tf.keras.losses.CategoricalCrossentropy()\n",
    "Acc = tf.keras.metrics.CategoricalAccuracy()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-19T14:30:24.345739Z",
     "iopub.status.busy": "2022-11-19T14:30:24.345357Z",
     "iopub.status.idle": "2022-11-19T14:30:24.350988Z",
     "shell.execute_reply": "2022-11-19T14:30:24.349576Z",
     "shell.execute_reply.started": "2022-11-19T14:30:24.345708Z"
    }
   },
   "source": [
    "#### Rotation of Images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:22.708375Z",
     "iopub.status.busy": "2022-11-23T14:21:22.707963Z",
     "iopub.status.idle": "2022-11-23T14:21:22.892100Z",
     "shell.execute_reply": "2022-11-23T14:21:22.890888Z",
     "shell.execute_reply.started": "2022-11-23T14:21:22.708336Z"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1008x288 with 4 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "## Get One batch\n",
    "for _,data in enumerate(trainRotDataloader):\n",
    "    break\n",
    "    \n",
    "index = 89\n",
    "plt.figure(figsize=(14,4))\n",
    "plt.subplot(141)\n",
    "plt.imshow(data[0][index])\n",
    "plt.title(\"0\"); plt.xticks([]); plt.yticks([]);\n",
    "plt.subplot(142)\n",
    "plt.imshow(np.flip(np.transpose(data[0],(0,2,1,3)),axis = 2)[index])\n",
    "plt.title(\"90\"); plt.xticks([]); plt.yticks([]);\n",
    "plt.subplot(143)\n",
    "plt.imshow(np.flip(data[0],axis= 1)[index])\n",
    "plt.title(\"180\");plt.xticks([]); plt.yticks([]);\n",
    "plt.subplot(144)\n",
    "plt.imshow(np.transpose(data[0],(0,2,1,3))[index])\n",
    "plt.title(\"270\");plt.xticks([]); plt.yticks([]);\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:21:28.636639Z",
     "iopub.status.busy": "2022-11-23T14:21:28.635735Z",
     "iopub.status.idle": "2022-11-23T14:47:20.301733Z",
     "shell.execute_reply": "2022-11-23T14:47:20.300781Z",
     "shell.execute_reply.started": "2022-11-23T14:21:28.636589Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "epoch 1/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 1.1410 - Accuracy: 0.4155\n",
      "epoch 2/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 1.0290 - Accuracy: 0.5188\n",
      "epoch 3/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.9810 - Accuracy: 0.5528\n",
      "epoch 4/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.9470 - Accuracy: 0.5707\n",
      "epoch 5/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.9190 - Accuracy: 0.5828\n",
      "epoch 6/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.9270 - Accuracy: 0.5929\n",
      "epoch 7/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.8780 - Accuracy: 0.6010\n",
      "epoch 8/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.8600 - Accuracy: 0.6104\n",
      "epoch 9/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.8580 - Accuracy: 0.6172\n",
      "epoch 10/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.8360 - Accuracy: 0.6237\n",
      "epoch 11/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.8210 - Accuracy: 0.6302\n",
      "epoch 12/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7990 - Accuracy: 0.6369\n",
      "epoch 13/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7920 - Accuracy: 0.6426\n",
      "epoch 14/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7770 - Accuracy: 0.6476\n",
      "epoch 15/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7520 - Accuracy: 0.6538\n",
      "epoch 16/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7650 - Accuracy: 0.6596\n",
      "epoch 17/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7380 - Accuracy: 0.6637\n",
      "epoch 18/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7390 - Accuracy: 0.6691\n",
      "epoch 19/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7320 - Accuracy: 0.6745\n",
      "epoch 20/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7170 - Accuracy: 0.6780\n",
      "epoch 21/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7050 - Accuracy: 0.6830\n",
      "epoch 22/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.7300 - Accuracy: 0.6869\n",
      "epoch 23/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6900 - Accuracy: 0.6907\n",
      "epoch 24/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6750 - Accuracy: 0.6954\n",
      "epoch 25/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6640 - Accuracy: 0.6977\n",
      "epoch 26/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6550 - Accuracy: 0.7034\n",
      "epoch 27/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6720 - Accuracy: 0.7068\n",
      "epoch 28/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6310 - Accuracy: 0.7107\n",
      "epoch 29/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6330 - Accuracy: 0.7135\n",
      "epoch 30/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6570 - Accuracy: 0.7197\n",
      "epoch 31/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6150 - Accuracy: 0.7211\n",
      "epoch 32/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6010 - Accuracy: 0.7262\n",
      "epoch 33/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6040 - Accuracy: 0.7287\n",
      "epoch 34/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.6010 - Accuracy: 0.7330\n",
      "epoch 35/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5700 - Accuracy: 0.7361\n",
      "epoch 36/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5720 - Accuracy: 0.7410\n",
      "epoch 37/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5770 - Accuracy: 0.7435\n",
      "epoch 38/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5870 - Accuracy: 0.7472\n",
      "epoch 39/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5380 - Accuracy: 0.7510\n",
      "epoch 40/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5610 - Accuracy: 0.7555\n",
      "epoch 41/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5460 - Accuracy: 0.7577\n",
      "epoch 42/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5310 - Accuracy: 0.7608\n",
      "epoch 43/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5350 - Accuracy: 0.7643\n",
      "epoch 44/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5340 - Accuracy: 0.7695\n",
      "epoch 45/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.4940 - Accuracy: 0.7742\n",
      "epoch 46/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.4990 - Accuracy: 0.7748\n",
      "epoch 47/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.4560 - Accuracy: 0.7799\n",
      "epoch 48/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5560 - Accuracy: 0.7827\n",
      "epoch 49/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.4590 - Accuracy: 0.7871\n",
      "epoch 50/50\n",
      "311/312 [============================>.] - ETA: 0s - Loss: 0.5310 - Accuracy: 0.7898"
     ]
    }
   ],
   "source": [
    "epochs = 50\n",
    "for epoch in range(epochs):\n",
    "    print(\"\\nepoch {}/{}\".format(epoch+1,epochs))\n",
    "    Acc.reset_states()\n",
    "    pb = tf.keras.utils.Progbar(len(trainRotDataloader),stateful_metrics=[\"Loss\",\"Accuracy\"])\n",
    "    for batch,data in enumerate(trainRotDataloader):\n",
    "        imgs, _ = data\n",
    "        size = imgs.shape[0]\n",
    "        \n",
    "        ## Generate Rotated Images and stack them\n",
    "        transformedImgs = np.vstack([imgs,np.flip(np.transpose(imgs,(0,2,1,3)),axis = 2),\n",
    "                              np.flip(imgs,axis= 1),np.transpose(imgs,(0,2,1,3))])\n",
    "        \n",
    "        ## Generate one-hot labels for rotation\n",
    "        labels = np.zeros((size*4,4))\n",
    "        for i in range(4):\n",
    "            labels[(i*size):((i+1)*size),i] = 1\n",
    "            \n",
    "        ##Forward pass the model\n",
    "        with tf.GradientTape() as tape:\n",
    "            pred = RotationModel(transformedImgs)\n",
    "            loss = loss_fn(labels,pred)\n",
    "            Acc.update_state(labels,pred)\n",
    "           \n",
    "        grads = tape.gradient(loss, RotationModel.trainable_weights)\n",
    "        optimizer.apply_gradients(zip(grads,RotationModel.trainable_weights))\n",
    "        pb.update(batch,values=[(\"Loss\",round(loss.numpy(),3)),(\"Accuracy\",Acc.result().numpy())])        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-19T14:24:49.297326Z",
     "iopub.status.busy": "2022-11-19T14:24:49.296909Z",
     "iopub.status.idle": "2022-11-19T14:24:49.644775Z",
     "shell.execute_reply": "2022-11-19T14:24:49.643791Z",
     "shell.execute_reply.started": "2022-11-19T14:24:49.297294Z"
    }
   },
   "outputs": [],
   "source": [
    "baseResNetModel.save(\"baseResNetModel.h5\")\n",
    "RotationModel.save(\"RotationModel.h5\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Train Classifier Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:47:20.304063Z",
     "iopub.status.busy": "2022-11-23T14:47:20.303713Z",
     "iopub.status.idle": "2022-11-23T14:47:20.472677Z",
     "shell.execute_reply": "2022-11-23T14:47:20.471649Z",
     "shell.execute_reply.started": "2022-11-23T14:47:20.304026Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model_22\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_23 (InputLayer)        [(None, None, None, 3)]   0         \n",
      "_________________________________________________________________\n",
      "model_20 (Functional)        (None, 512)               11190912  \n",
      "_________________________________________________________________\n",
      "dense_8 (Dense)              (None, 128)               65664     \n",
      "_________________________________________________________________\n",
      "dense_9 (Dense)              (None, 10)                1290      \n",
      "=================================================================\n",
      "Total params: 11,257,866\n",
      "Trainable params: 66,954\n",
      "Non-trainable params: 11,190,912\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "baseResNetModel.trainable = False\n",
    "ClassifierModel = Model(baseResNetModel, nClasses = 10)\n",
    "ClassifierModel.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T14:47:20.474409Z",
     "iopub.status.busy": "2022-11-23T14:47:20.474062Z",
     "iopub.status.idle": "2022-11-23T14:50:25.176842Z",
     "shell.execute_reply": "2022-11-23T14:50:25.175910Z",
     "shell.execute_reply.started": "2022-11-23T14:47:20.474374Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "epoch 1/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.2760 - Accuracy: 0.1378\n",
      "epoch 2/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.2110 - Accuracy: 0.1728\n",
      "epoch 3/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.1600 - Accuracy: 0.1992\n",
      "epoch 4/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.1160 - Accuracy: 0.2282\n",
      "epoch 5/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.0790 - Accuracy: 0.2499\n",
      "epoch 6/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.0460 - Accuracy: 0.2696\n",
      "epoch 7/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 2.0160 - Accuracy: 0.2855\n",
      "epoch 8/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.9900 - Accuracy: 0.2986\n",
      "epoch 9/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.9660 - Accuracy: 0.3097\n",
      "epoch 10/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.9440 - Accuracy: 0.3188\n",
      "epoch 11/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.9240 - Accuracy: 0.3267\n",
      "epoch 12/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.9050 - Accuracy: 0.3336\n",
      "epoch 13/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8880 - Accuracy: 0.3413\n",
      "epoch 14/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8720 - Accuracy: 0.3482\n",
      "epoch 15/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8570 - Accuracy: 0.3530\n",
      "epoch 16/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8430 - Accuracy: 0.3569\n",
      "epoch 17/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8310 - Accuracy: 0.3602\n",
      "epoch 18/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8190 - Accuracy: 0.3645\n",
      "epoch 19/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.8080 - Accuracy: 0.3689\n",
      "epoch 20/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7970 - Accuracy: 0.3721\n",
      "epoch 21/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7870 - Accuracy: 0.3756\n",
      "epoch 22/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7780 - Accuracy: 0.3789\n",
      "epoch 23/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7700 - Accuracy: 0.3817\n",
      "epoch 24/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7620 - Accuracy: 0.3848\n",
      "epoch 25/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7540 - Accuracy: 0.3891\n",
      "epoch 26/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7470 - Accuracy: 0.3907\n",
      "epoch 27/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7400 - Accuracy: 0.3943\n",
      "epoch 28/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7330 - Accuracy: 0.3974\n",
      "epoch 29/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7270 - Accuracy: 0.3996\n",
      "epoch 30/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7210 - Accuracy: 0.4014\n",
      "epoch 31/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7160 - Accuracy: 0.4035\n",
      "epoch 32/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7100 - Accuracy: 0.4050\n",
      "epoch 33/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7050 - Accuracy: 0.4073\n",
      "epoch 34/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.7000 - Accuracy: 0.4081\n",
      "epoch 35/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6950 - Accuracy: 0.4108\n",
      "epoch 36/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6900 - Accuracy: 0.4119\n",
      "epoch 37/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6860 - Accuracy: 0.4134\n",
      "epoch 38/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6820 - Accuracy: 0.4146\n",
      "epoch 39/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6780 - Accuracy: 0.4166\n",
      "epoch 40/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6740 - Accuracy: 0.4176\n",
      "epoch 41/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6700 - Accuracy: 0.4197\n",
      "epoch 42/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6660 - Accuracy: 0.4217\n",
      "epoch 43/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6620 - Accuracy: 0.4232\n",
      "epoch 44/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6590 - Accuracy: 0.4242\n",
      "epoch 45/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6550 - Accuracy: 0.4256\n",
      "epoch 46/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6520 - Accuracy: 0.4273\n",
      "epoch 47/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6490 - Accuracy: 0.4292\n",
      "epoch 48/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6450 - Accuracy: 0.4307\n",
      "epoch 49/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6420 - Accuracy: 0.4316\n",
      "epoch 50/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6390 - Accuracy: 0.4325\n",
      "epoch 51/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6360 - Accuracy: 0.4335\n",
      "epoch 52/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6330 - Accuracy: 0.4340\n",
      "epoch 53/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6300 - Accuracy: 0.4353\n",
      "epoch 54/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6280 - Accuracy: 0.4365\n",
      "epoch 55/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6250 - Accuracy: 0.4377\n",
      "epoch 56/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6220 - Accuracy: 0.4390\n",
      "epoch 57/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6190 - Accuracy: 0.4404\n",
      "epoch 58/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6170 - Accuracy: 0.4415\n",
      "epoch 59/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6140 - Accuracy: 0.4420\n",
      "epoch 60/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6120 - Accuracy: 0.4438\n",
      "epoch 61/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6090 - Accuracy: 0.4442\n",
      "epoch 62/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6070 - Accuracy: 0.4448\n",
      "epoch 63/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6040 - Accuracy: 0.4454\n",
      "epoch 64/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6020 - Accuracy: 0.4460\n",
      "epoch 65/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.6000 - Accuracy: 0.4476\n",
      "epoch 66/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5970 - Accuracy: 0.4484\n",
      "epoch 67/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5950 - Accuracy: 0.4491\n",
      "epoch 68/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5930 - Accuracy: 0.4505\n",
      "epoch 69/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5910 - Accuracy: 0.4518\n",
      "epoch 70/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5880 - Accuracy: 0.4530\n",
      "epoch 71/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5860 - Accuracy: 0.4539\n",
      "epoch 72/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5840 - Accuracy: 0.4547\n",
      "epoch 73/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5820 - Accuracy: 0.4554\n",
      "epoch 74/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5800 - Accuracy: 0.4563\n",
      "epoch 75/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5780 - Accuracy: 0.4578\n",
      "epoch 76/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5760 - Accuracy: 0.4588\n",
      "epoch 77/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5740 - Accuracy: 0.4597\n",
      "epoch 78/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5720 - Accuracy: 0.4604\n",
      "epoch 79/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5700 - Accuracy: 0.4609\n",
      "epoch 80/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5690 - Accuracy: 0.4618\n",
      "epoch 81/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5670 - Accuracy: 0.4624\n",
      "epoch 82/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5650 - Accuracy: 0.4634\n",
      "epoch 83/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5630 - Accuracy: 0.4647\n",
      "epoch 84/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5610 - Accuracy: 0.4650\n",
      "epoch 85/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5600 - Accuracy: 0.4659\n",
      "epoch 86/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5580 - Accuracy: 0.4664\n",
      "epoch 87/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5560 - Accuracy: 0.4666\n",
      "epoch 88/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5550 - Accuracy: 0.4674\n",
      "epoch 89/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5530 - Accuracy: 0.4685\n",
      "epoch 90/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5510 - Accuracy: 0.4697\n",
      "epoch 91/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5500 - Accuracy: 0.4701\n",
      "epoch 92/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5480 - Accuracy: 0.4703\n",
      "epoch 93/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5470 - Accuracy: 0.4710\n",
      "epoch 94/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5450 - Accuracy: 0.4714\n",
      "epoch 95/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5440 - Accuracy: 0.4723\n",
      "epoch 96/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5420 - Accuracy: 0.4732\n",
      "epoch 97/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5400 - Accuracy: 0.4735\n",
      "epoch 98/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5390 - Accuracy: 0.4742\n",
      "epoch 99/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5370 - Accuracy: 0.4746\n",
      "epoch 100/100\n",
      "18/19 [===========================>..] - ETA: 0s - Loss: 1.5360 - Accuracy: 0.4754"
     ]
    }
   ],
   "source": [
    "epochs = 100\n",
    "for epoch in range(epochs):\n",
    "    print(\"\\nepoch {}/{}\".format(epoch+1,epochs))\n",
    "    Acc.reset_states()\n",
    "    pb = tf.keras.utils.Progbar(len(trainClfDataloader),stateful_metrics=[\"Loss\",\"Accuracy\"])\n",
    "    for batch,data in enumerate(trainClfDataloader):\n",
    "        imgs, labels = data\n",
    "        labels = tf.keras.utils.to_categorical(labels,num_classes=10)\n",
    "        \n",
    "        ##Forward pass the model\n",
    "        with tf.GradientTape() as tape:\n",
    "            pred = ClassifierModel(imgs)\n",
    "            loss = loss_fn(labels,pred)\n",
    "            Acc.update_state(labels,pred)\n",
    "           \n",
    "        grads = tape.gradient(loss, ClassifierModel.trainable_weights)\n",
    "        optimizer.apply_gradients(zip(grads,ClassifierModel.trainable_weights))\n",
    "        pb.update(batch,values=[(\"Loss\",round(loss.numpy(),3)),(\"Accuracy\",Acc.result().numpy())])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ClassifierModel.save(\"ClassifierModel.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T15:24:27.762350Z",
     "iopub.status.busy": "2022-11-23T15:24:27.761943Z",
     "iopub.status.idle": "2022-11-23T15:24:28.627354Z",
     "shell.execute_reply": "2022-11-23T15:24:28.626237Z",
     "shell.execute_reply.started": "2022-11-23T15:24:27.762285Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Classifier Accuracy on cifar 10 test dataset 0.4391\n"
     ]
    }
   ],
   "source": [
    "### Checking classifier accuracy on cifar 10 test dataset\n",
    "Acc.reset_states()\n",
    "for batch,data in enumerate(testClfDataloader):\n",
    "    imgs, labels = data\n",
    "    labels = tf.keras.utils.to_categorical(labels)\n",
    "    \n",
    "    ##Forward pass the model\n",
    "    with tf.GradientTape() as tape:\n",
    "        pred = ClassifierModel(imgs,training=False)\n",
    "        Acc.update_state(labels,pred)\n",
    "print(\"Classifier Accuracy on cifar 10 test dataset\",Acc.result().numpy())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2022-11-23T15:24:28.630851Z",
     "iopub.status.busy": "2022-11-23T15:24:28.630170Z",
     "iopub.status.idle": "2022-11-23T15:24:31.349053Z",
     "shell.execute_reply": "2022-11-23T15:24:31.348000Z",
     "shell.execute_reply.started": "2022-11-23T15:24:28.630806Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Rototional Accuracy on cifar 10 test set 0.6663\n"
     ]
    }
   ],
   "source": [
    "### Checking rotational accuracy on classifier test dataset\n",
    "Acc.reset_states()\n",
    "for batch,data in enumerate(testClfDataloader):\n",
    "    imgs, labels = data\n",
    "    size = imgs.shape[0]\n",
    "\n",
    "    ## Generate Rotated Images and stack them\n",
    "    transformedImgs = np.vstack([imgs,np.flip(np.transpose(imgs,(0,2,1,3)),axis = 2),\n",
    "                          np.flip(imgs,axis= 1),np.transpose(imgs,(0,2,1,3))])\n",
    "\n",
    "    ## Generate one-hot labels for rotation\n",
    "    labels = np.zeros((size*4,4))\n",
    "    for i in range(4):\n",
    "        labels[(i*size):((i+1)*size),i] = 1\n",
    "        \n",
    "    ##Forward pass the model\n",
    "    with tf.GradientTape() as tape:\n",
    "        pred = RotationModel(transformedImgs,training=False)\n",
    "        Acc.update_state(labels,pred)\n",
    "print(\"Rototional Accuracy on cifar 10 test set\",Acc.result().numpy())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}