classification-tree.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import keras\n",
    "import keras.backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "(X_train,Y_train),(X_test,Y_test) = keras.datasets.cifar100.load_data()\n",
    "X_train,X_test = X_train/255,X_test/255"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(100, 7, 127)"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tree_outputs = max(np.max(Y_train), np.max(Y_test)) + 1\n",
    "tree_depth = int(np.ceil(np.log2(tree_outputs)))\n",
    "tree_nodes = 2**tree_depth - 1\n",
    "tree_outputs, tree_depth, tree_nodes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def tree_path_and_label(x):\n",
    "    label,path = [],[0]\n",
    "    for i in range(tree_depth):\n",
    "        k = 2**(tree_depth-i-1)\n",
    "        if x >= k:\n",
    "            label.append(1)\n",
    "            path.append(2*path[-1]+2)\n",
    "            x = x - k\n",
    "        else:\n",
    "            label.append(0)\n",
    "            path.append(2*path[-1]+1)\n",
    "    path.pop()\n",
    "    return np.array(path), np.array(label)\n",
    "Y_train_path = np.array([tree_path_and_label(y) for y in Y_train])\n",
    "Y_train_path, Y_train_label = Y_train_path[:,0,:], Y_train_path[:,1,:]\n",
    "Y_test_path = np.array([tree_path_and_label(y) for y in Y_test])\n",
    "Y_test_path, Y_test_label = Y_test_path[:,0,:], Y_test_path[:,1,:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "input_3 (InputLayer)            (None, 32, 32, 3)    0                                            \n",
      "__________________________________________________________________________________________________\n",
      "model_1 (Model)                 (None, 256)          255807      input_3[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "input_4 (InputLayer)            (None, 7)            0                                            \n",
      "__________________________________________________________________________________________________\n",
      "repeat_vector_1 (RepeatVector)  (None, 7, 256)       0           model_1[1][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "model_2 (Model)                 (None, 7, 256)       32768       input_4[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "lambda_1 (Lambda)               (None, 7)            0           repeat_vector_1[0][0]            \n",
      "                                                                 model_2[1][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "activation_23 (Activation)      (None, 7)            0           lambda_1[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "lambda_2 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_3 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_4 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_5 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_6 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_7 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "lambda_8 (Lambda)               (None, 1)            0           activation_23[0][0]              \n",
      "==================================================================================================\n",
      "Total params: 288,575\n",
      "Trainable params: 285,561\n",
      "Non-trainable params: 3,014\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "X = X_input = keras.layers.Input(X_train.shape[1:])\n",
    "X = keras.layers.BatchNormalization()(X)\n",
    "for i in range(5):\n",
    "    for _ in range(4):\n",
    "        X = keras.layers.SeparableConv2D(2**(i+3), (3,3), padding='same')(X)\n",
    "        X = keras.layers.BatchNormalization()(X)\n",
    "        X = keras.layers.Activation('relu')(X)\n",
    "    X = keras.layers.MaxPooling2D()(X)\n",
    "X = keras.layers.Flatten()(X)\n",
    "X = keras.layers.Dense(256)(X)\n",
    "X = keras.layers.BatchNormalization()(X)\n",
    "X = keras.layers.Activation('relu')(X)\n",
    "X = keras.layers.Dense(256)(X)\n",
    "X = keras.layers.BatchNormalization()(X)\n",
    "X = keras.layers.Activation('relu')(X)\n",
    "X = keras.layers.Dense(256)(X)\n",
    "M = keras.Model(X_input, X)\n",
    "X_path = X_path_input = keras.layers.Input(Y_train_path.shape[1:])\n",
    "X_path = keras.layers.Embedding(tree_nodes+1, 256)(X_path)\n",
    "T = keras.Model(X_path_input, X_path)\n",
    "X = X_input = keras.layers.Input(X_train.shape[1:])\n",
    "X = M(X)\n",
    "X = keras.layers.RepeatVector(Y_train_path.shape[1])(X)\n",
    "X_path = X_path_input = keras.layers.Input(Y_train_path.shape[1:])\n",
    "X_path = T(X_path)\n",
    "X = keras.layers.Lambda(\n",
    "    lambda x: K.sum(x[0]*x[1], axis=-1),\n",
    "    output_shape=Y_train_label.shape[1:])([X,X_path])\n",
    "X = keras.layers.Activation('sigmoid')(X)\n",
    "X_outputs = []\n",
    "for i in range(Y_train_label.shape[-1]):\n",
    "    X_outputs.append(keras.layers.Lambda(\n",
    "        lambda x,a: x[:,a:a+1],\n",
    "        arguments={'a':i},\n",
    "        output_shape=(1,))(X))\n",
    "G = keras.Model([X_input,X_path_input], X_outputs)\n",
    "G.compile('nadam', 'binary_crossentropy', ['acc'])\n",
    "G.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 50000 samples, validate on 10000 samples\n",
      "Epoch 1/20\n",
      "50000/50000 [==============================] - 64s 1ms/step - loss: 4.2066 - lambda_2_loss: 0.6496 - lambda_3_loss: 0.5646 - lambda_4_loss: 0.6298 - lambda_5_loss: 0.6539 - lambda_6_loss: 0.6077 - lambda_7_loss: 0.5776 - lambda_8_loss: 0.5234 - lambda_2_acc: 0.6412 - lambda_3_acc: 0.6675 - lambda_4_acc: 0.6189 - lambda_5_acc: 0.5820 - lambda_6_acc: 0.6347 - lambda_7_acc: 0.6807 - lambda_8_acc: 0.7216 - val_loss: 4.5630 - val_lambda_2_loss: 0.6760 - val_lambda_3_loss: 0.5772 - val_lambda_4_loss: 0.7106 - val_lambda_5_loss: 0.7116 - val_lambda_6_loss: 0.6735 - val_lambda_7_loss: 0.6279 - val_lambda_8_loss: 0.5862 - val_lambda_2_acc: 0.6204 - val_lambda_3_acc: 0.6782 - val_lambda_4_acc: 0.6180 - val_lambda_5_acc: 0.5910 - val_lambda_6_acc: 0.6639 - val_lambda_7_acc: 0.7085 - val_lambda_8_acc: 0.7416\n",
      "Epoch 2/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.8543 - lambda_2_loss: 0.6404 - lambda_3_loss: 0.5445 - lambda_4_loss: 0.6039 - lambda_5_loss: 0.6216 - lambda_6_loss: 0.5423 - lambda_7_loss: 0.4864 - lambda_8_loss: 0.4151 - lambda_2_acc: 0.6482 - lambda_3_acc: 0.6915 - lambda_4_acc: 0.6503 - lambda_5_acc: 0.6283 - lambda_6_acc: 0.7042 - lambda_7_acc: 0.7574 - lambda_8_acc: 0.8004 - val_loss: 4.4516 - val_lambda_2_loss: 0.6705 - val_lambda_3_loss: 0.6487 - val_lambda_4_loss: 0.6597 - val_lambda_5_loss: 0.6799 - val_lambda_6_loss: 0.6496 - val_lambda_7_loss: 0.6302 - val_lambda_8_loss: 0.5130 - val_lambda_2_acc: 0.6461 - val_lambda_3_acc: 0.6604 - val_lambda_4_acc: 0.6399 - val_lambda_5_acc: 0.6135 - val_lambda_6_acc: 0.6700 - val_lambda_7_acc: 0.7124 - val_lambda_8_acc: 0.7733\n",
      "Epoch 3/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.6542 - lambda_2_loss: 0.6352 - lambda_3_loss: 0.5297 - lambda_4_loss: 0.5865 - lambda_5_loss: 0.5988 - lambda_6_loss: 0.5036 - lambda_7_loss: 0.4379 - lambda_8_loss: 0.3624 - lambda_2_acc: 0.6517 - lambda_3_acc: 0.7041 - lambda_4_acc: 0.6694 - lambda_5_acc: 0.6508 - lambda_6_acc: 0.7331 - lambda_7_acc: 0.7892 - lambda_8_acc: 0.8307 - val_loss: 4.1929 - val_lambda_2_loss: 0.6380 - val_lambda_3_loss: 0.6138 - val_lambda_4_loss: 0.7201 - val_lambda_5_loss: 0.6716 - val_lambda_6_loss: 0.6163 - val_lambda_7_loss: 0.5147 - val_lambda_8_loss: 0.4183 - val_lambda_2_acc: 0.6549 - val_lambda_3_acc: 0.6889 - val_lambda_4_acc: 0.6419 - val_lambda_5_acc: 0.6251 - val_lambda_6_acc: 0.7106 - val_lambda_7_acc: 0.7703 - val_lambda_8_acc: 0.8142\n",
      "Epoch 4/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.5227 - lambda_2_loss: 0.6306 - lambda_3_loss: 0.5187 - lambda_4_loss: 0.5725 - lambda_5_loss: 0.5812 - lambda_6_loss: 0.4785 - lambda_7_loss: 0.4098 - lambda_8_loss: 0.3314 - lambda_2_acc: 0.6550 - lambda_3_acc: 0.7125 - lambda_4_acc: 0.6811 - lambda_5_acc: 0.6701 - lambda_6_acc: 0.7546 - lambda_7_acc: 0.8079 - lambda_8_acc: 0.8497 - val_loss: 4.3762 - val_lambda_2_loss: 0.6546 - val_lambda_3_loss: 0.6107 - val_lambda_4_loss: 0.6712 - val_lambda_5_loss: 0.6888 - val_lambda_6_loss: 0.6371 - val_lambda_7_loss: 0.6087 - val_lambda_8_loss: 0.5051 - val_lambda_2_acc: 0.6532 - val_lambda_3_acc: 0.6772 - val_lambda_4_acc: 0.6402 - val_lambda_5_acc: 0.6371 - val_lambda_6_acc: 0.6911 - val_lambda_7_acc: 0.7409 - val_lambda_8_acc: 0.7941\n",
      "Epoch 5/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.4226 - lambda_2_loss: 0.6262 - lambda_3_loss: 0.5101 - lambda_4_loss: 0.5629 - lambda_5_loss: 0.5652 - lambda_6_loss: 0.4584 - lambda_7_loss: 0.3894 - lambda_8_loss: 0.3103 - lambda_2_acc: 0.6584 - lambda_3_acc: 0.7193 - lambda_4_acc: 0.6906 - lambda_5_acc: 0.6878 - lambda_6_acc: 0.7691 - lambda_7_acc: 0.8183 - lambda_8_acc: 0.8609 - val_loss: 5.5482 - val_lambda_2_loss: 0.7123 - val_lambda_3_loss: 0.6967 - val_lambda_4_loss: 0.8418 - val_lambda_5_loss: 0.8601 - val_lambda_6_loss: 0.9683 - val_lambda_7_loss: 0.8468 - val_lambda_8_loss: 0.6223 - val_lambda_2_acc: 0.6468 - val_lambda_3_acc: 0.6886 - val_lambda_4_acc: 0.6156 - val_lambda_5_acc: 0.6094 - val_lambda_6_acc: 0.6508 - val_lambda_7_acc: 0.7183 - val_lambda_8_acc: 0.7733\n",
      "Epoch 6/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.3336 - lambda_2_loss: 0.6202 - lambda_3_loss: 0.5016 - lambda_4_loss: 0.5526 - lambda_5_loss: 0.5522 - lambda_6_loss: 0.4415 - lambda_7_loss: 0.3711 - lambda_8_loss: 0.2944 - lambda_2_acc: 0.6640 - lambda_3_acc: 0.7270 - lambda_4_acc: 0.6995 - lambda_5_acc: 0.6997 - lambda_6_acc: 0.7798 - lambda_7_acc: 0.8276 - lambda_8_acc: 0.8683 - val_loss: 3.9256 - val_lambda_2_loss: 0.6472 - val_lambda_3_loss: 0.5607 - val_lambda_4_loss: 0.6074 - val_lambda_5_loss: 0.6134 - val_lambda_6_loss: 0.5561 - val_lambda_7_loss: 0.5230 - val_lambda_8_loss: 0.4178 - val_lambda_2_acc: 0.6519 - val_lambda_3_acc: 0.7102 - val_lambda_4_acc: 0.6730 - val_lambda_5_acc: 0.6637 - val_lambda_6_acc: 0.7471 - val_lambda_7_acc: 0.7787 - val_lambda_8_acc: 0.8252\n",
      "Epoch 7/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.2524 - lambda_2_loss: 0.6148 - lambda_3_loss: 0.4933 - lambda_4_loss: 0.5429 - lambda_5_loss: 0.5398 - lambda_6_loss: 0.4257 - lambda_7_loss: 0.3551 - lambda_8_loss: 0.2807 - lambda_2_acc: 0.6673 - lambda_3_acc: 0.7353 - lambda_4_acc: 0.7079 - lambda_5_acc: 0.7080 - lambda_6_acc: 0.7893 - lambda_7_acc: 0.8352 - lambda_8_acc: 0.8759 - val_loss: 3.9325 - val_lambda_2_loss: 0.6466 - val_lambda_3_loss: 0.5403 - val_lambda_4_loss: 0.6238 - val_lambda_5_loss: 0.6487 - val_lambda_6_loss: 0.5525 - val_lambda_7_loss: 0.5234 - val_lambda_8_loss: 0.3972 - val_lambda_2_acc: 0.6610 - val_lambda_3_acc: 0.7154 - val_lambda_4_acc: 0.6611 - val_lambda_5_acc: 0.6673 - val_lambda_6_acc: 0.7376 - val_lambda_7_acc: 0.7763 - val_lambda_8_acc: 0.8266\n",
      "Epoch 8/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.1807 - lambda_2_loss: 0.6082 - lambda_3_loss: 0.4857 - lambda_4_loss: 0.5351 - lambda_5_loss: 0.5264 - lambda_6_loss: 0.4147 - lambda_7_loss: 0.3436 - lambda_8_loss: 0.2671 - lambda_2_acc: 0.6731 - lambda_3_acc: 0.7395 - lambda_4_acc: 0.7145 - lambda_5_acc: 0.7209 - lambda_6_acc: 0.7979 - lambda_7_acc: 0.8423 - lambda_8_acc: 0.8830 - val_loss: 3.5108 - val_lambda_2_loss: 0.6221 - val_lambda_3_loss: 0.5147 - val_lambda_4_loss: 0.5789 - val_lambda_5_loss: 0.5728 - val_lambda_6_loss: 0.4645 - val_lambda_7_loss: 0.4033 - val_lambda_8_loss: 0.3545 - val_lambda_2_acc: 0.6661 - val_lambda_3_acc: 0.7274 - val_lambda_4_acc: 0.6863 - val_lambda_5_acc: 0.6899 - val_lambda_6_acc: 0.7741 - val_lambda_7_acc: 0.8196 - val_lambda_8_acc: 0.8507\n",
      "Epoch 9/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.1184 - lambda_2_loss: 0.6037 - lambda_3_loss: 0.4813 - lambda_4_loss: 0.5268 - lambda_5_loss: 0.5166 - lambda_6_loss: 0.4031 - lambda_7_loss: 0.3281 - lambda_8_loss: 0.2588 - lambda_2_acc: 0.6774 - lambda_3_acc: 0.7420 - lambda_4_acc: 0.7201 - lambda_5_acc: 0.7294 - lambda_6_acc: 0.8027 - lambda_7_acc: 0.8504 - lambda_8_acc: 0.8869 - val_loss: 3.6035 - val_lambda_2_loss: 0.6334 - val_lambda_3_loss: 0.5323 - val_lambda_4_loss: 0.5724 - val_lambda_5_loss: 0.5838 - val_lambda_6_loss: 0.4890 - val_lambda_7_loss: 0.4471 - val_lambda_8_loss: 0.3455 - val_lambda_2_acc: 0.6461 - val_lambda_3_acc: 0.7222 - val_lambda_4_acc: 0.6967 - val_lambda_5_acc: 0.6969 - val_lambda_6_acc: 0.7685 - val_lambda_7_acc: 0.8095 - val_lambda_8_acc: 0.8519\n",
      "Epoch 10/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 3.0587 - lambda_2_loss: 0.5996 - lambda_3_loss: 0.4731 - lambda_4_loss: 0.5199 - lambda_5_loss: 0.5060 - lambda_6_loss: 0.3932 - lambda_7_loss: 0.3188 - lambda_8_loss: 0.2481 - lambda_2_acc: 0.6788 - lambda_3_acc: 0.7516 - lambda_4_acc: 0.7278 - lambda_5_acc: 0.7341 - lambda_6_acc: 0.8100 - lambda_7_acc: 0.8549 - lambda_8_acc: 0.8909 - val_loss: 3.3546 - val_lambda_2_loss: 0.6193 - val_lambda_3_loss: 0.5181 - val_lambda_4_loss: 0.5458 - val_lambda_5_loss: 0.5417 - val_lambda_6_loss: 0.4421 - val_lambda_7_loss: 0.3758 - val_lambda_8_loss: 0.3118 - val_lambda_2_acc: 0.6650 - val_lambda_3_acc: 0.7280 - val_lambda_4_acc: 0.7135 - val_lambda_5_acc: 0.7112 - val_lambda_6_acc: 0.7831 - val_lambda_7_acc: 0.8293 - val_lambda_8_acc: 0.8632\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 11/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.9984 - lambda_2_loss: 0.5944 - lambda_3_loss: 0.4675 - lambda_4_loss: 0.5114 - lambda_5_loss: 0.4982 - lambda_6_loss: 0.3806 - lambda_7_loss: 0.3066 - lambda_8_loss: 0.2397 - lambda_2_acc: 0.6830 - lambda_3_acc: 0.7564 - lambda_4_acc: 0.7336 - lambda_5_acc: 0.7420 - lambda_6_acc: 0.8187 - lambda_7_acc: 0.8623 - lambda_8_acc: 0.8946 - val_loss: 3.5727 - val_lambda_2_loss: 0.6343 - val_lambda_3_loss: 0.5172 - val_lambda_4_loss: 0.5825 - val_lambda_5_loss: 0.5789 - val_lambda_6_loss: 0.4992 - val_lambda_7_loss: 0.4218 - val_lambda_8_loss: 0.3388 - val_lambda_2_acc: 0.6599 - val_lambda_3_acc: 0.7368 - val_lambda_4_acc: 0.6933 - val_lambda_5_acc: 0.7014 - val_lambda_6_acc: 0.7647 - val_lambda_7_acc: 0.8130 - val_lambda_8_acc: 0.8577\n",
      "Epoch 12/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.9489 - lambda_2_loss: 0.5889 - lambda_3_loss: 0.4633 - lambda_4_loss: 0.5045 - lambda_5_loss: 0.4906 - lambda_6_loss: 0.3706 - lambda_7_loss: 0.2974 - lambda_8_loss: 0.2335 - lambda_2_acc: 0.6868 - lambda_3_acc: 0.7561 - lambda_4_acc: 0.7387 - lambda_5_acc: 0.7474 - lambda_6_acc: 0.8245 - lambda_7_acc: 0.8679 - lambda_8_acc: 0.8996 - val_loss: 3.2808 - val_lambda_2_loss: 0.6038 - val_lambda_3_loss: 0.5060 - val_lambda_4_loss: 0.5432 - val_lambda_5_loss: 0.5381 - val_lambda_6_loss: 0.4307 - val_lambda_7_loss: 0.3656 - val_lambda_8_loss: 0.2934 - val_lambda_2_acc: 0.6742 - val_lambda_3_acc: 0.7335 - val_lambda_4_acc: 0.7115 - val_lambda_5_acc: 0.7223 - val_lambda_6_acc: 0.7940 - val_lambda_7_acc: 0.8339 - val_lambda_8_acc: 0.8753\n",
      "Epoch 13/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.9000 - lambda_2_loss: 0.5860 - lambda_3_loss: 0.4578 - lambda_4_loss: 0.4958 - lambda_5_loss: 0.4830 - lambda_6_loss: 0.3639 - lambda_7_loss: 0.2892 - lambda_8_loss: 0.2243 - lambda_2_acc: 0.6884 - lambda_3_acc: 0.7619 - lambda_4_acc: 0.7447 - lambda_5_acc: 0.7541 - lambda_6_acc: 0.8261 - lambda_7_acc: 0.8711 - lambda_8_acc: 0.9042 - val_loss: 3.4836 - val_lambda_2_loss: 0.6352 - val_lambda_3_loss: 0.5063 - val_lambda_4_loss: 0.5559 - val_lambda_5_loss: 0.5509 - val_lambda_6_loss: 0.4807 - val_lambda_7_loss: 0.4170 - val_lambda_8_loss: 0.3376 - val_lambda_2_acc: 0.6750 - val_lambda_3_acc: 0.7315 - val_lambda_4_acc: 0.7146 - val_lambda_5_acc: 0.7152 - val_lambda_6_acc: 0.7778 - val_lambda_7_acc: 0.8188 - val_lambda_8_acc: 0.8610\n",
      "Epoch 14/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.8485 - lambda_2_loss: 0.5800 - lambda_3_loss: 0.4530 - lambda_4_loss: 0.4891 - lambda_5_loss: 0.4735 - lambda_6_loss: 0.3540 - lambda_7_loss: 0.2810 - lambda_8_loss: 0.2180 - lambda_2_acc: 0.6913 - lambda_3_acc: 0.7657 - lambda_4_acc: 0.7491 - lambda_5_acc: 0.7596 - lambda_6_acc: 0.8344 - lambda_7_acc: 0.8747 - lambda_8_acc: 0.9054 - val_loss: 3.3087 - val_lambda_2_loss: 0.6277 - val_lambda_3_loss: 0.5070 - val_lambda_4_loss: 0.5440 - val_lambda_5_loss: 0.5297 - val_lambda_6_loss: 0.4340 - val_lambda_7_loss: 0.3672 - val_lambda_8_loss: 0.2991 - val_lambda_2_acc: 0.6542 - val_lambda_3_acc: 0.7413 - val_lambda_4_acc: 0.7203 - val_lambda_5_acc: 0.7279 - val_lambda_6_acc: 0.7997 - val_lambda_7_acc: 0.8407 - val_lambda_8_acc: 0.8774\n",
      "Epoch 15/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.7989 - lambda_2_loss: 0.5737 - lambda_3_loss: 0.4481 - lambda_4_loss: 0.4841 - lambda_5_loss: 0.4665 - lambda_6_loss: 0.3450 - lambda_7_loss: 0.2711 - lambda_8_loss: 0.2104 - lambda_2_acc: 0.6960 - lambda_3_acc: 0.7698 - lambda_4_acc: 0.7522 - lambda_5_acc: 0.7655 - lambda_6_acc: 0.8387 - lambda_7_acc: 0.8809 - lambda_8_acc: 0.9089 - val_loss: 3.5182 - val_lambda_2_loss: 0.6295 - val_lambda_3_loss: 0.5174 - val_lambda_4_loss: 0.5673 - val_lambda_5_loss: 0.5596 - val_lambda_6_loss: 0.4803 - val_lambda_7_loss: 0.4228 - val_lambda_8_loss: 0.3413 - val_lambda_2_acc: 0.6691 - val_lambda_3_acc: 0.7282 - val_lambda_4_acc: 0.7033 - val_lambda_5_acc: 0.7076 - val_lambda_6_acc: 0.7749 - val_lambda_7_acc: 0.8188 - val_lambda_8_acc: 0.8572\n",
      "Epoch 16/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.7545 - lambda_2_loss: 0.5710 - lambda_3_loss: 0.4432 - lambda_4_loss: 0.4758 - lambda_5_loss: 0.4598 - lambda_6_loss: 0.3382 - lambda_7_loss: 0.2634 - lambda_8_loss: 0.2032 - lambda_2_acc: 0.6999 - lambda_3_acc: 0.7713 - lambda_4_acc: 0.7572 - lambda_5_acc: 0.7688 - lambda_6_acc: 0.8428 - lambda_7_acc: 0.8847 - lambda_8_acc: 0.9129 - val_loss: 3.6118 - val_lambda_2_loss: 0.6447 - val_lambda_3_loss: 0.5295 - val_lambda_4_loss: 0.5759 - val_lambda_5_loss: 0.6007 - val_lambda_6_loss: 0.4789 - val_lambda_7_loss: 0.4300 - val_lambda_8_loss: 0.3520 - val_lambda_2_acc: 0.6682 - val_lambda_3_acc: 0.7355 - val_lambda_4_acc: 0.7181 - val_lambda_5_acc: 0.7089 - val_lambda_6_acc: 0.7865 - val_lambda_7_acc: 0.8203 - val_lambda_8_acc: 0.8588\n",
      "Epoch 17/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.7188 - lambda_2_loss: 0.5670 - lambda_3_loss: 0.4394 - lambda_4_loss: 0.4723 - lambda_5_loss: 0.4546 - lambda_6_loss: 0.3317 - lambda_7_loss: 0.2572 - lambda_8_loss: 0.1965 - lambda_2_acc: 0.7041 - lambda_3_acc: 0.7738 - lambda_4_acc: 0.7601 - lambda_5_acc: 0.7718 - lambda_6_acc: 0.8461 - lambda_7_acc: 0.8886 - lambda_8_acc: 0.9159 - val_loss: 3.3536 - val_lambda_2_loss: 0.5982 - val_lambda_3_loss: 0.5052 - val_lambda_4_loss: 0.5441 - val_lambda_5_loss: 0.5470 - val_lambda_6_loss: 0.4491 - val_lambda_7_loss: 0.4022 - val_lambda_8_loss: 0.3078 - val_lambda_2_acc: 0.6817 - val_lambda_3_acc: 0.7343 - val_lambda_4_acc: 0.7284 - val_lambda_5_acc: 0.7191 - val_lambda_6_acc: 0.7942 - val_lambda_7_acc: 0.8247 - val_lambda_8_acc: 0.8664\n",
      "Epoch 18/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.6755 - lambda_2_loss: 0.5630 - lambda_3_loss: 0.4362 - lambda_4_loss: 0.4638 - lambda_5_loss: 0.4448 - lambda_6_loss: 0.3260 - lambda_7_loss: 0.2484 - lambda_8_loss: 0.1933 - lambda_2_acc: 0.7032 - lambda_3_acc: 0.7770 - lambda_4_acc: 0.7657 - lambda_5_acc: 0.7778 - lambda_6_acc: 0.8502 - lambda_7_acc: 0.8928 - lambda_8_acc: 0.9170 - val_loss: 3.2978 - val_lambda_2_loss: 0.6184 - val_lambda_3_loss: 0.5049 - val_lambda_4_loss: 0.5296 - val_lambda_5_loss: 0.5240 - val_lambda_6_loss: 0.4422 - val_lambda_7_loss: 0.3779 - val_lambda_8_loss: 0.3009 - val_lambda_2_acc: 0.6677 - val_lambda_3_acc: 0.7474 - val_lambda_4_acc: 0.7307 - val_lambda_5_acc: 0.7355 - val_lambda_6_acc: 0.8026 - val_lambda_7_acc: 0.8432 - val_lambda_8_acc: 0.8788\n",
      "Epoch 19/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.6476 - lambda_2_loss: 0.5616 - lambda_3_loss: 0.4330 - lambda_4_loss: 0.4592 - lambda_5_loss: 0.4401 - lambda_6_loss: 0.3172 - lambda_7_loss: 0.2480 - lambda_8_loss: 0.1884 - lambda_2_acc: 0.7063 - lambda_3_acc: 0.7788 - lambda_4_acc: 0.7690 - lambda_5_acc: 0.7829 - lambda_6_acc: 0.8551 - lambda_7_acc: 0.8931 - lambda_8_acc: 0.9200 - val_loss: 3.6673 - val_lambda_2_loss: 0.6389 - val_lambda_3_loss: 0.5325 - val_lambda_4_loss: 0.5561 - val_lambda_5_loss: 0.5940 - val_lambda_6_loss: 0.5124 - val_lambda_7_loss: 0.4736 - val_lambda_8_loss: 0.3597 - val_lambda_2_acc: 0.6560 - val_lambda_3_acc: 0.7334 - val_lambda_4_acc: 0.7128 - val_lambda_5_acc: 0.7110 - val_lambda_6_acc: 0.7761 - val_lambda_7_acc: 0.8043 - val_lambda_8_acc: 0.8540\n",
      "Epoch 20/20\n",
      "50000/50000 [==============================] - 56s 1ms/step - loss: 2.6084 - lambda_2_loss: 0.5561 - lambda_3_loss: 0.4264 - lambda_4_loss: 0.4568 - lambda_5_loss: 0.4347 - lambda_6_loss: 0.3139 - lambda_7_loss: 0.2369 - lambda_8_loss: 0.1836 - lambda_2_acc: 0.7111 - lambda_3_acc: 0.7843 - lambda_4_acc: 0.7712 - lambda_5_acc: 0.7833 - lambda_6_acc: 0.8570 - lambda_7_acc: 0.8962 - lambda_8_acc: 0.9214 - val_loss: 3.3757 - val_lambda_2_loss: 0.6453 - val_lambda_3_loss: 0.5109 - val_lambda_4_loss: 0.5298 - val_lambda_5_loss: 0.5360 - val_lambda_6_loss: 0.4545 - val_lambda_7_loss: 0.3841 - val_lambda_8_loss: 0.3151 - val_lambda_2_acc: 0.6657 - val_lambda_3_acc: 0.7338 - val_lambda_4_acc: 0.7252 - val_lambda_5_acc: 0.7371 - val_lambda_6_acc: 0.7944 - val_lambda_7_acc: 0.8334 - val_lambda_8_acc: 0.8735\n"
     ]
    }
   ],
   "source": [
    "hist = G.fit(\n",
    "    [X_train, Y_train_path],\n",
    "    np.split(Y_train_label, Y_train_label.shape[-1], axis=-1),\n",
    "    validation_data=([X_test, Y_test_path],\n",
    "        np.split(Y_test_label, Y_test_label.shape[-1], axis=-1)),\n",
    "    batch_size=128, epochs=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fe75a7246a0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(hist.history['loss'], label='loss')\n",
    "plt.plot(hist.history['val_loss'], label='val_loss')\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.2078"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def sigmoid(x):\n",
    "    return np.where(x >= 0, \n",
    "        1 / (1 + np.exp(-x)), \n",
    "        np.exp(x) / (1 + np.exp(x)))\n",
    "acc = M.predict(X_test, batch_size=128)\n",
    "tree_embedding = T.get_weights()[0]\n",
    "acc = np.tile(acc[:,np.newaxis,:], (1,Y_test_path.shape[-1],1))\n",
    "acc = acc * tree_embedding[Y_test_path]\n",
    "acc = np.sum(acc, axis=-1)\n",
    "acc = sigmoid(acc)\n",
    "acc = np.where(acc >= 0.5, 1, 0)\n",
    "acc = np.all(acc == Y_test_label, axis=-1)\n",
    "np.sum(acc) / Y_test_label.shape[0]"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "ml",
   "language": "python",
   "name": "ml"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}