triplet_loss.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import keras\n",
    "import numpy as np\n",
    "import tensorflow as tf\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Conv2D, MaxPooling2D, BatchNormalization, ZeroPadding2D, Dropout, Activation, Flatten, Dense\n",
    "from keras.layers.normalization import BatchNormalization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.datasets import mnist\n",
    "from collections import defaultdict\n",
    "from matplotlib import pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.metrics import pairwise_distances\n",
    "from sklearn.manifold import TSNE\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "(x_train, y_train), (x_test, y_test) = mnist.load_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def mnist_generator():\n",
    "    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
    "    train_dict = defaultdict(lambda: [])\n",
    "    for i in range(len(x_train)):\n",
    "        train_dict[y_train[i]].append(x_train[i])\n",
    "    for keys in train_dict:\n",
    "        train_dict[keys] = np.array(train_dict[keys])\n",
    "    labels = np.arange(9)\n",
    "    \n",
    "    while(True):\n",
    "        \n",
    "        sample = np.random.choice(labels, 6, replace=False)\n",
    "        p_choices =np.random.choice(len(train_dict[sample[0]])-1, 10, replace=False)\n",
    "        n_choices =[np.random.choice(len(train_dict[sample[i]])-1, 1, replace=False)[0] for i in range(1, 6)]\n",
    "        \n",
    "        positive_image = train_dict[sample[0]][p_choices].reshape(10, 28, 28, 1)\n",
    "        negative_image = np.array([train_dict[sample[i]][n_choices[i-1]] for i in range(1, 6)]).reshape(5, 28, 28, 1)\n",
    "        \n",
    "        \n",
    "        y = [sample[0] for i in range(10)]\n",
    "        for i in range(1, 6):\n",
    "            y.append(sample[i])\n",
    "#         print(np.vstack((positive_image / 255., negative_image / 255.)).shape, y.shape)\n",
    "        yield np.vstack((positive_image / 255., negative_image / 255.)), np.array(y)\n",
    "        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 15 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 15 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Visualize the generator mini batch, expect to see 10 sample which same class and 5 sample from different class\n",
    "k = 0\n",
    "x_ = None\n",
    "columns = 3\n",
    "rows = 5\n",
    "for train, labels in mnist_generator():\n",
    "    k += 1\n",
    "    if k >= 3:\n",
    "        break\n",
    "    # Show random images\n",
    "    x_ = train\n",
    "\n",
    "    fig=plt.figure()\n",
    "\n",
    "    for i in range(1, columns*rows +1):\n",
    "        fig.add_subplot(rows, columns, i)\n",
    "        plt.imshow(x_[i-1].reshape(28, 28))\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def triplets_loss(y_true, y_pred):\n",
    "    \n",
    "#     embeddings = K.cast(embeddings, 'float32')\n",
    "#     with sess.as_default():\n",
    "#         print(embeddings.eval())\n",
    "    \n",
    "    embeddings = y_pred\n",
    "    anchor_positive = embeddings[:10]\n",
    "    negative = embeddings[10:]\n",
    "#     print(anchor_positive)\n",
    "\n",
    "    # Compute pairwise distance between all of anchor-positive\n",
    "    dot_product = K.dot(anchor_positive, K.transpose(anchor_positive))\n",
    "    square = K.square(anchor_positive)\n",
    "    a_p_distance = K.reshape(K.sum(square, axis=1), (-1,1)) - 2.*dot_product  + K.sum(K.transpose(square), axis=0) + 1e-6\n",
    "    a_p_distance = K.maximum(a_p_distance, 0.0) ## Numerical stability\n",
    "#     with K.get_session().as_default():\n",
    "#         print(a_p_distance.eval())\n",
    "#     print(\"Pairwise shape: \", a_p_distance)\n",
    "#     print(\"Negative shape: \", negative)\n",
    "\n",
    "    # Compute distance between anchor and negative\n",
    "    dot_product_2 = K.dot(anchor_positive, K.transpose(negative))\n",
    "    negative_square = K.square(negative)\n",
    "    a_n_distance = K.reshape(K.sum(square, axis=1), (-1,1)) - 2.*dot_product_2  + K.sum(K.transpose(negative_square), axis=0)  + 1e-6\n",
    "    a_n_distance = K.maximum(a_n_distance, 0.0) ## Numerical stability\n",
    "    \n",
    "    hard_negative = K.reshape(K.min(a_n_distance, axis=1), (-1, 1))\n",
    "    \n",
    "    distance = (a_p_distance - hard_negative + 0.2)\n",
    "    loss = K.mean(K.maximum(distance, 0.0))/(2.)\n",
    "\n",
    "#     with K.get_session().as_default():\n",
    "#             print(loss.eval())\n",
    "            \n",
    "    return loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1.516795159936006\n"
     ]
    }
   ],
   "source": [
    "# Sanity check the implementation\n",
    "test_batch = np.random.rand(15, 128)\n",
    "test_batch\n",
    "with K.get_session().as_default():\n",
    "        print(triplets_loss(None, tf.convert_to_tensor(test_batch)).eval())\n",
    "            \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "# declare a simple CNN network\n",
    "def baseline_model():\n",
    "    # create model\n",
    "    model = Sequential()\n",
    "    \n",
    "    model.add(Conv2D(32, (3, 3), input_shape=(28, 28, 1), activation='relu'))\n",
    "    model.add(BatchNormalization())\n",
    "    model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "    \n",
    "    model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\n",
    "    model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\n",
    "    model.add(BatchNormalization())\n",
    "    model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "    \n",
    "    model.add(Conv2D(128, (5, 5), padding='same', activation='relu'))\n",
    "    model.add(BatchNormalization())\n",
    "    model.add(MaxPooling2D(pool_size=(2, 2)))\n",
    "    model.add(Dropout(0.2))\n",
    "    model.add(Flatten())\n",
    "    model.add(Dense(256, activation='relu'))\n",
    "    model.add(Dense(32))\n",
    "    # Compile model\n",
    "    optim = keras.optimizers.Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=True)\n",
    "    model.compile(loss=triplets_loss, optimizer=optim)\n",
    "    model.summary()\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_1 (Conv2D)            (None, 26, 26, 32)        320       \n",
      "_________________________________________________________________\n",
      "batch_normalization_1 (Batch (None, 26, 26, 32)        128       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_1 (MaxPooling2 (None, 13, 13, 32)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_2 (Conv2D)            (None, 13, 13, 64)        18496     \n",
      "_________________________________________________________________\n",
      "conv2d_3 (Conv2D)            (None, 13, 13, 64)        36928     \n",
      "_________________________________________________________________\n",
      "batch_normalization_2 (Batch (None, 13, 13, 64)        256       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_2 (MaxPooling2 (None, 6, 6, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv2d_4 (Conv2D)            (None, 6, 6, 128)         204928    \n",
      "_________________________________________________________________\n",
      "batch_normalization_3 (Batch (None, 6, 6, 128)         512       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_3 (MaxPooling2 (None, 3, 3, 128)         0         \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 3, 3, 128)         0         \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 1152)              0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 256)               295168    \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 32)                8224      \n",
      "=================================================================\n",
      "Total params: 564,960\n",
      "Trainable params: 564,512\n",
      "Non-trainable params: 448\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = baseline_model()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "5000/5000 [==============================] - 64s 13ms/step - loss: 0.0021\n",
      "Epoch 2/10\n",
      "5000/5000 [==============================] - 63s 13ms/step - loss: 0.0021\n",
      "Epoch 3/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0017\n",
      "Epoch 4/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0017\n",
      "Epoch 5/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0016\n",
      "Epoch 6/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0014\n",
      "Epoch 7/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0014 \n",
      "Epoch 8/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0012\n",
      "Epoch 9/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 0.0010\n",
      "Epoch 10/10\n",
      "5000/5000 [==============================] - 62s 12ms/step - loss: 9.5238e-04\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x21786cb7710>"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.fit_generator(mnist_generator(), steps_per_epoch=5000, epochs=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "tsne = TSNE()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "choices = np.random.choice(np.arange(len(x_test)), len(x_test), replace=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_sample, y_sample = x_test[choices], y_test[choices]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "embeddings = model.predict(x_sample.reshape((len(x_sample), 28, 28, 1)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000, 32)"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "embeddings.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tsne_train = tsne.fit_transform(embeddings)\n",
    "    \n",
    "plt.figure(figsize=(16, 16))\n",
    "colors = 'r', 'g', 'b', 'c', 'm', 'y', 'k', 'w', 'orange', 'purple'\n",
    "for i, c in zip(range(10), colors):\n",
    "    plt.scatter(tsne_train[y_sample == i, 0], tsne_train[y_sample == i, 1], c=c, label=str(i))\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "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": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}