MNIST_baseline.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "613a3912",
   "metadata": {},
   "outputs": [],
   "source": [
    "import itertools\n",
    "import matplotlib\n",
    "\n",
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "import matplotlib.pyplot as plt\n",
    "import tensorflow.keras.backend as K\n",
    "\n",
    "from tqdm import tqdm\n",
    "from scipy.spatial import distance"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "d68b95e8",
   "metadata": {},
   "outputs": [],
   "source": [
    "font = {'family' : 'normal',\n",
    "        'weight' : 'bold',\n",
    "        'size'   : 22}\n",
    "\n",
    "matplotlib.rc('font', **font)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "15b69400",
   "metadata": {},
   "outputs": [],
   "source": [
    "np.random.seed(0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "56e08335",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[0 1 1 ... 0 0 1]\n",
      "(12593, 784)\n"
     ]
    }
   ],
   "source": [
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()\n",
    "\n",
    "x_train = x_train / 255.\n",
    "x_test = x_test / 255.\n",
    "\n",
    "y_train = tf.keras.utils.to_categorical(y_train, 10)\n",
    "y_test = tf.keras.utils.to_categorical(y_test, 10)\n",
    "\n",
    "y = np.argmax(y_train, axis=1)\n",
    "x_train_1s = x_train[np.where(y==1)]\n",
    "x_train_8s = x_train[np.where(y==8)]\n",
    "y_train_1s = y[np.where(y==1)]\n",
    "y_train_8s = y[np.where(y==8)]\n",
    "\n",
    "x_train_1s_8s = np.concatenate((x_train_1s, x_train_8s))\n",
    "y_train_1s_8s = np.concatenate((y_train_1s, y_train_8s))\n",
    "\n",
    "y_train_1s_8s[np.where(y_train_1s_8s==8)] = 0\n",
    "\n",
    "order = np.random.permutation(len(x_train_1s_8s))\n",
    "\n",
    "x_train_1s_8s = x_train_1s_8s[order]\n",
    "y_train_1s_8s = y_train_1s_8s[order]\n",
    "\n",
    "print(y_train_1s_8s)\n",
    "\n",
    "y_train_1s_8s = tf.keras.utils.to_categorical(y_train_1s_8s, 2)\n",
    "\n",
    "x_train_1s_8s_reshape = [x_train_1s_8s[i].reshape((784)) for i in range(len(x_train_1s_8s))]\n",
    "print(np.array(x_train_1s_8s_reshape).shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "6945ac65",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_hidden_neurons = 4\n",
    "lr = 0.001\n",
    "epochs = 200\n",
    "batch_size = 8\n",
    "limit = 256\n",
    "alpha = 0.0 # ignore polysemantic term in loss function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "6eb870e4",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2022-05-15 21:07:51.458356: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA\n",
      "To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
     ]
    }
   ],
   "source": [
    "i = tf.keras.layers.Input(shape=(28*28,))\n",
    "i2 = tf.keras.layers.Input(shape=(2,))\n",
    "e = tf.keras.layers.Dense(n_hidden_neurons, activation='relu', name='dense_1')(i)\n",
    "o = tf.keras.layers.Dense(2, activation='softmax', name='softmax')(e)\n",
    "model = tf.keras.Model(inputs=[i, i2], outputs=o)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "40c4f629",
   "metadata": {},
   "outputs": [],
   "source": [
    "def calculate_term(e, c0, c1):\n",
    "  a = K.expand_dims(e[c0] / (K.max(K.abs(e[c0])) + 1e-9))\n",
    "  b = K.expand_dims(e[c1] / (K.max(K.abs(e[c1])) + 1e-9))\n",
    "  if c0==c1:\n",
    "    return 0\n",
    "  else:\n",
    "    return K.dot(K.transpose(a), b)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "4aa03c6e",
   "metadata": {},
   "outputs": [],
   "source": [
    "#https://stackoverflow.com/questions/62454500/how-to-use-tensorflow-custom-loss-for-a-keras-model\n",
    "def my_loss(y_true, y_pred, e, c0, c1, alpha):\n",
    "    p_loss = 0\n",
    "    for i in range(len(c0)):\n",
    "      p_loss += calculate_term(e, c0[i], c1[i])\n",
    "\n",
    "    loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False)(y_true=y_true, \n",
    "                                                                      y_pred=y_pred)\n",
    "    loss += alpha * (p_loss / len(c0))\n",
    "    return loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "cd19bc71",
   "metadata": {},
   "outputs": [],
   "source": [
    "c0, c1 = [], []\n",
    "for c in list(itertools.combinations(range(batch_size), 2)):\n",
    "    c0.append(c[0])\n",
    "    c1.append(c[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "4fd423ac",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "28\n"
     ]
    }
   ],
   "source": [
    "print(len(c0))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "5dc6d9b3",
   "metadata": {},
   "outputs": [],
   "source": [
    "model.add_loss(my_loss(i2, o, e, np.array(c0, dtype='int32'), np.array(c1, dtype='int32'), alpha))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "ede7015a",
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=lr), \n",
    "              loss=None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "3d725091",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2022-05-15 21:07:53.796248: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/200\n",
      "32/32 [==============================] - 5s 3ms/step - loss: 0.5712\n",
      "Epoch 2/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.3271\n",
      "Epoch 3/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.1989\n",
      "Epoch 4/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.1450\n",
      "Epoch 5/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.1170\n",
      "Epoch 6/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0980\n",
      "Epoch 7/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0839\n",
      "Epoch 8/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0733\n",
      "Epoch 9/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0649\n",
      "Epoch 10/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0611\n",
      "Epoch 11/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0527\n",
      "Epoch 12/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0485\n",
      "Epoch 13/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0435\n",
      "Epoch 14/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0394\n",
      "Epoch 15/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0363\n",
      "Epoch 16/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0348\n",
      "Epoch 17/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0301\n",
      "Epoch 18/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0280\n",
      "Epoch 19/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0263\n",
      "Epoch 20/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0238\n",
      "Epoch 21/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0224\n",
      "Epoch 22/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0203\n",
      "Epoch 23/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0189\n",
      "Epoch 24/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0176\n",
      "Epoch 25/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0159\n",
      "Epoch 26/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0149\n",
      "Epoch 27/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0146\n",
      "Epoch 28/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0130\n",
      "Epoch 29/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0122\n",
      "Epoch 30/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0111\n",
      "Epoch 31/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0105\n",
      "Epoch 32/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0097\n",
      "Epoch 33/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0092\n",
      "Epoch 34/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0086\n",
      "Epoch 35/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0081\n",
      "Epoch 36/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0077\n",
      "Epoch 37/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0074\n",
      "Epoch 38/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0069\n",
      "Epoch 39/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0063\n",
      "Epoch 40/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0060\n",
      "Epoch 41/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0057\n",
      "Epoch 42/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0054\n",
      "Epoch 43/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0052\n",
      "Epoch 44/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0049\n",
      "Epoch 45/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0047\n",
      "Epoch 46/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0045\n",
      "Epoch 47/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0041\n",
      "Epoch 48/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0040\n",
      "Epoch 49/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0038\n",
      "Epoch 50/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0038\n",
      "Epoch 51/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0035\n",
      "Epoch 52/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0033\n",
      "Epoch 53/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0032\n",
      "Epoch 54/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0031\n",
      "Epoch 55/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0029\n",
      "Epoch 56/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0028\n",
      "Epoch 57/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0027\n",
      "Epoch 58/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0026\n",
      "Epoch 59/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0025\n",
      "Epoch 60/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0024\n",
      "Epoch 61/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0023\n",
      "Epoch 62/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0022\n",
      "Epoch 63/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0021\n",
      "Epoch 64/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0020\n",
      "Epoch 65/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0020\n",
      "Epoch 66/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0019\n",
      "Epoch 67/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0018\n",
      "Epoch 68/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0018\n",
      "Epoch 69/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0017\n",
      "Epoch 70/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0017\n",
      "Epoch 71/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0016\n",
      "Epoch 72/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 0.0015\n",
      "Epoch 73/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0015\n",
      "Epoch 74/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0014\n",
      "Epoch 75/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0014\n",
      "Epoch 76/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0014\n",
      "Epoch 77/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0013\n",
      "Epoch 78/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0012\n",
      "Epoch 79/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0012\n",
      "Epoch 80/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0012\n",
      "Epoch 81/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0011\n",
      "Epoch 82/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0011\n",
      "Epoch 83/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0011\n",
      "Epoch 84/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0011\n",
      "Epoch 85/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 0.0010\n",
      "Epoch 86/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 9.8042e-04\n",
      "Epoch 87/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 9.6389e-04\n",
      "Epoch 88/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 9.2428e-04\n",
      "Epoch 89/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.9580e-04\n",
      "Epoch 90/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.7480e-04\n",
      "Epoch 91/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 8.4774e-04\n",
      "Epoch 92/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.2373e-04\n",
      "Epoch 93/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.9771e-04\n",
      "Epoch 94/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.7513e-04\n",
      "Epoch 95/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.5215e-04\n",
      "Epoch 96/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.4372e-04\n",
      "Epoch 97/200\n",
      "32/32 [==============================] - 0s 4ms/step - loss: 7.1150e-04\n",
      "Epoch 98/200\n",
      "32/32 [==============================] - 0s 4ms/step - loss: 6.9355e-04\n",
      "Epoch 99/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 6.7214e-04\n",
      "Epoch 100/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 6.5592e-04\n",
      "Epoch 101/200\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "32/32 [==============================] - 0s 3ms/step - loss: 6.4455e-04\n",
      "Epoch 102/200\n",
      "32/32 [==============================] - 0s 6ms/step - loss: 6.2265e-04\n",
      "Epoch 103/200\n",
      "32/32 [==============================] - 0s 4ms/step - loss: 6.0275e-04\n",
      "Epoch 104/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 5.8633e-04\n",
      "Epoch 105/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 5.7798e-04\n",
      "Epoch 106/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 5.5703e-04\n",
      "Epoch 107/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 5.4806e-04\n",
      "Epoch 108/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 5.3666e-04\n",
      "Epoch 109/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 5.1800e-04\n",
      "Epoch 110/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 5.0455e-04\n",
      "Epoch 111/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.9449e-04\n",
      "Epoch 112/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.7820e-04\n",
      "Epoch 113/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.7025e-04\n",
      "Epoch 114/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.5723e-04\n",
      "Epoch 115/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.4549e-04\n",
      "Epoch 116/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.3194e-04\n",
      "Epoch 117/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 4.2799e-04\n",
      "Epoch 118/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.1912e-04\n",
      "Epoch 119/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 4.0707e-04\n",
      "Epoch 120/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.9382e-04\n",
      "Epoch 121/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.9180e-04\n",
      "Epoch 122/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.7218e-04\n",
      "Epoch 123/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.6408e-04\n",
      "Epoch 124/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.5661e-04\n",
      "Epoch 125/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.5193e-04\n",
      "Epoch 126/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.3886e-04\n",
      "Epoch 127/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.3384e-04\n",
      "Epoch 128/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.2431e-04\n",
      "Epoch 129/200\n",
      "32/32 [==============================] - 0s 2ms/step - loss: 3.1799e-04\n",
      "Epoch 130/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.1012e-04\n",
      "Epoch 131/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 3.0968e-04\n",
      "Epoch 132/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.9513e-04\n",
      "Epoch 133/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.8950e-04\n",
      "Epoch 134/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.8226e-04\n",
      "Epoch 135/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.7558e-04\n",
      "Epoch 136/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.7053e-04\n",
      "Epoch 137/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.6508e-04\n",
      "Epoch 138/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.5731e-04\n",
      "Epoch 139/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.5326e-04\n",
      "Epoch 140/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.4766e-04\n",
      "Epoch 141/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.4637e-04\n",
      "Epoch 142/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.3596e-04\n",
      "Epoch 143/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.3024e-04\n",
      "Epoch 144/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.2510e-04\n",
      "Epoch 145/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.2090e-04\n",
      "Epoch 146/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.1545e-04\n",
      "Epoch 147/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.1038e-04\n",
      "Epoch 148/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.0685e-04\n",
      "Epoch 149/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 2.0292e-04\n",
      "Epoch 150/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.9955e-04\n",
      "Epoch 151/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.9386e-04\n",
      "Epoch 152/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.9075e-04\n",
      "Epoch 153/200\n",
      "32/32 [==============================] - 0s 4ms/step - loss: 1.8535e-04\n",
      "Epoch 154/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.8279e-04\n",
      "Epoch 155/200\n",
      "32/32 [==============================] - 0s 5ms/step - loss: 1.8093e-04\n",
      "Epoch 156/200\n",
      "32/32 [==============================] - 0s 5ms/step - loss: 1.7430e-04\n",
      "Epoch 157/200\n",
      "32/32 [==============================] - 0s 4ms/step - loss: 1.7083e-04\n",
      "Epoch 158/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.6644e-04\n",
      "Epoch 159/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.6262e-04\n",
      "Epoch 160/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.5973e-04\n",
      "Epoch 161/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.5586e-04\n",
      "Epoch 162/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.5337e-04\n",
      "Epoch 163/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.4939e-04\n",
      "Epoch 164/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.4698e-04\n",
      "Epoch 165/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.4889e-04\n",
      "Epoch 166/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.4036e-04\n",
      "Epoch 167/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.3761e-04\n",
      "Epoch 168/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.3526e-04\n",
      "Epoch 169/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.3190e-04\n",
      "Epoch 170/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.3003e-04\n",
      "Epoch 171/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.2693e-04\n",
      "Epoch 172/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.2357e-04\n",
      "Epoch 173/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.2188e-04\n",
      "Epoch 174/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.1939e-04\n",
      "Epoch 175/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.1688e-04\n",
      "Epoch 176/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.1443e-04\n",
      "Epoch 177/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.1158e-04\n",
      "Epoch 178/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.0965e-04\n",
      "Epoch 179/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.0738e-04\n",
      "Epoch 180/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.0634e-04\n",
      "Epoch 181/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.0287e-04\n",
      "Epoch 182/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 1.0153e-04\n",
      "Epoch 183/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.9369e-05\n",
      "Epoch 184/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.6963e-05\n",
      "Epoch 185/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.4928e-05\n",
      "Epoch 186/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.3903e-05\n",
      "Epoch 187/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.1657e-05\n",
      "Epoch 188/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 9.0158e-05\n",
      "Epoch 189/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.8159e-05\n",
      "Epoch 190/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.5915e-05\n",
      "Epoch 191/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.4655e-05\n",
      "Epoch 192/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.2842e-05\n",
      "Epoch 193/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 8.1362e-05\n",
      "Epoch 194/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.9701e-05\n",
      "Epoch 195/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.7788e-05\n",
      "Epoch 196/200\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "32/32 [==============================] - 0s 3ms/step - loss: 7.6585e-05\n",
      "Epoch 197/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.5104e-05\n",
      "Epoch 198/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.3629e-05\n",
      "Epoch 199/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.2199e-05\n",
      "Epoch 200/200\n",
      "32/32 [==============================] - 0s 3ms/step - loss: 7.0674e-05\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x18ede1b50>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.fit([np.array(x_train_1s_8s_reshape)[:limit], y_train_1s_8s[:limit]],\n",
    "          y=None,\n",
    "          batch_size=batch_size,\n",
    "          epochs=epochs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "ffe33452",
   "metadata": {},
   "outputs": [],
   "source": [
    "embed = tf.keras.Model(inputs=i, outputs=e)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "9e351734",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|███████████████████████████████████████████████████████████████████████| 256/256 [00:00<00:00, 7672.43it/s]\n"
     ]
    }
   ],
   "source": [
    "embed.layers[1].set_weights(model.layers[1].get_weights())\n",
    "activations_by_class = {}\n",
    "a = embed.predict(np.array(x_train_1s_8s_reshape))\n",
    "for i in tqdm(range(len(x_train_1s_8s_reshape[:limit]))):\n",
    "    try:\n",
    "        activations_by_class[np.argmax(y_train_1s_8s, axis=1)[i]] += np.squeeze(a[i])\n",
    "    except KeyError:\n",
    "        activations_by_class[np.argmax(y_train_1s_8s, axis=1)[i]] = np.squeeze(a[i])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "27ef5d7c",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "findfont: Font family ['normal'] not found. Falling back to DejaVu Sans.\n",
      "findfont: Font family ['normal'] not found. Falling back to DejaVu Sans.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x1440 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig = plt.figure(figsize=(20,20))\n",
    "for i,k in enumerate(activations_by_class.keys()):\n",
    "    y_pos = np.arange(len(activations_by_class[k]))\n",
    "    ax = fig.add_subplot(1, 2, i+1)\n",
    "    ax.set_title(k)\n",
    "    ax.barh(y_pos, np.squeeze(activations_by_class[k]), align='center')\n",
    "    ax.plot([0,0],[-1,len(np.squeeze(activations_by_class[k]))], 'k-')\n",
    "    ax.set_ylim(-1,len(np.squeeze(activations_by_class[k])))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "4a5b5998",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x1440 with 4 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "W = embed.layers[1].get_weights()\n",
    "fig = plt.figure(figsize=(20,20))\n",
    "dim = int(np.ceil(np.sqrt(n_hidden_neurons)))\n",
    "for j in range(n_hidden_neurons):\n",
    "  x_j = W[0][:,j] / np.sqrt(np.sum(np.dot(W[0][:,j], W[0][:,j].T)))\n",
    "  ax = fig.add_subplot(dim,dim,j+1)\n",
    "  ax.imshow(x_j.reshape((28,28)))\n",
    "  plt.axis('off')\n",
    "plt.show()"
   ]
  }
 ],
 "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": "3.7.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}