model_2.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train model\n",
    "Model is inspired by Baidu 2.\n",
    "See models/model_2.png for architecture diagram. Save model on every epoch."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Prepare train and valid data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/matplotlib/font_manager.py:278: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.\n",
      "  'Matplotlib is building the font cache using fc-list. '\n"
     ]
    }
   ],
   "source": [
    "from src.data_generator import make_audio_gen\n",
    "\n",
    "\n",
    "TRAIN_CORPUS = \"data_st/train_corpus.json\"\n",
    "VALID_CORPUS = \"data_st/valid_corpus.json\"\n",
    "\n",
    "MFCC_DIM = 13\n",
    "SPECTOGRAM = False\n",
    "EPOCHS = 2\n",
    "MODEL_NAME = \"model_2\"\n",
    "MINI_BATCH_SIZE = 250\n",
    "SORT_BY_DURATION=False\n",
    "MAX_DURATION = 10.0\n",
    "\n",
    "audio_gen = make_audio_gen(TRAIN_CORPUS, VALID_CORPUS, spectrogram=False, mfcc_dim=MFCC_DIM,\n",
    "                           minibatch_size=MINI_BATCH_SIZE, sort_by_duration=SORT_BY_DURATION,\n",
    "                           max_duration=MAX_DURATION)\n",
    "# add the training data to the generator\n",
    "audio_gen.load_train_data()\n",
    "audio_gen.load_validation_data()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Run training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From /home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Colocations handled automatically by placer.\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "the_input (InputLayer)       (None, None, 13)          0         \n",
      "_________________________________________________________________\n",
      "layer_1_conv (Conv1D)        (None, None, 200)         28800     \n",
      "_________________________________________________________________\n",
      "conv_batch_norm (BatchNormal (None, None, 200)         800       \n",
      "_________________________________________________________________\n",
      "rnn_1 (GRU)                  (None, None, 250)         338250    \n",
      "_________________________________________________________________\n",
      "bt_rnn_1 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_2 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_2 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_3 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_3 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_4 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_4 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "final_layer_of_rnn (GRU)     (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_final (BatchNormaliza (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "time_distributed_1 (TimeDist (None, None, 223)         55973     \n",
      "_________________________________________________________________\n",
      "softmax (Activation)         (None, None, 223)         0         \n",
      "=================================================================\n",
      "Total params: 1,931,823\n",
      "Trainable params: 1,928,923\n",
      "Non-trainable params: 2,900\n",
      "_________________________________________________________________\n",
      "None\n",
      "WARNING:tensorflow:From /home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py:4249: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Use tf.cast instead.\n",
      "WARNING:tensorflow:From /home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py:4229: to_int64 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Use tf.cast instead.\n",
      "WARNING:tensorflow:From /home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow/python/ops/math_grad.py:102: div (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Deprecated in favor of operator or tf.math.divide.\n",
      "Epoch 1/40\n",
      "37/37 [==============================] - 698s 19s/step - loss: 294.8893 - val_loss: 275.2110\n",
      "Epoch 2/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 183.2353 - val_loss: 221.7322\n",
      "Epoch 3/40\n",
      "37/37 [==============================] - 284s 8s/step - loss: 173.1160 - val_loss: 206.8790\n",
      "Epoch 4/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 163.8706 - val_loss: 204.2038\n",
      "Epoch 5/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 147.8705 - val_loss: 194.3479\n",
      "Epoch 6/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 126.8548 - val_loss: 183.7972\n",
      "Epoch 7/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 109.6716 - val_loss: 169.7674\n",
      "Epoch 8/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 96.6256 - val_loss: 143.0346\n",
      "Epoch 9/40\n",
      "37/37 [==============================] - 277s 7s/step - loss: 86.5622 - val_loss: 128.1374\n",
      "Epoch 10/40\n",
      "37/37 [==============================] - 283s 8s/step - loss: 79.6420 - val_loss: 114.9585\n",
      "Epoch 11/40\n",
      "37/37 [==============================] - 283s 8s/step - loss: 73.9136 - val_loss: 123.6323\n",
      "Epoch 12/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 69.9558 - val_loss: 105.7059\n",
      "Epoch 13/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 66.1719 - val_loss: 101.0773\n",
      "Epoch 14/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 63.0848 - val_loss: 98.2723\n",
      "Epoch 15/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 60.2478 - val_loss: 99.2324\n",
      "Epoch 16/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 57.8849 - val_loss: 95.9872\n",
      "Epoch 17/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 55.6378 - val_loss: 97.1039\n",
      "Epoch 18/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 53.7870 - val_loss: 93.6262\n",
      "Epoch 19/40\n",
      "37/37 [==============================] - 282s 8s/step - loss: 51.7757 - val_loss: 89.8333\n",
      "Epoch 20/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 49.8312 - val_loss: 96.8465\n",
      "Epoch 21/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 48.2688 - val_loss: 95.9883\n",
      "Epoch 22/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 46.5187 - val_loss: 93.4521\n",
      "Epoch 23/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 44.9839 - val_loss: 91.4823\n",
      "Epoch 24/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 43.4967 - val_loss: 88.6644\n",
      "Epoch 25/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 42.0769 - val_loss: 95.9970\n",
      "Epoch 26/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 40.7368 - val_loss: 93.1917\n",
      "Epoch 27/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 39.3175 - val_loss: 93.2005\n",
      "Epoch 28/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 37.9768 - val_loss: 99.2874\n",
      "Epoch 29/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 36.6046 - val_loss: 97.9306\n",
      "Epoch 30/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 35.3490 - val_loss: 94.7189\n",
      "Epoch 31/40\n",
      "37/37 [==============================] - 282s 8s/step - loss: 34.0487 - val_loss: 96.6904\n",
      "Epoch 32/40\n",
      "37/37 [==============================] - 278s 8s/step - loss: 33.0520 - val_loss: 97.5740\n",
      "Epoch 33/40\n",
      "37/37 [==============================] - 283s 8s/step - loss: 31.9078 - val_loss: 100.0249\n",
      "Epoch 34/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 30.6181 - val_loss: 99.6203\n",
      "Epoch 35/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 29.6347 - val_loss: 105.1588\n",
      "Epoch 36/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 28.4633 - val_loss: 106.3366\n",
      "Epoch 37/40\n",
      "37/37 [==============================] - 279s 8s/step - loss: 27.5134 - val_loss: 103.4251\n",
      "Epoch 38/40\n",
      "37/37 [==============================] - 283s 8s/step - loss: 26.2988 - val_loss: 103.9742\n",
      "Epoch 39/40\n",
      "37/37 [==============================] - 281s 8s/step - loss: 25.2729 - val_loss: 103.1728\n",
      "Epoch 40/40\n",
      "37/37 [==============================] - 280s 8s/step - loss: 24.2762 - val_loss: 111.1886\n"
     ]
    }
   ],
   "source": [
    "from src.train import train\n",
    "from src.models import model_2\n",
    "from src.char_map import char_map, index_map\n",
    "\n",
    "EPOCHS = 40\n",
    "MODEL_NAME = \"model_2\"\n",
    "\n",
    "model = model_2(input_dim=13,\n",
    "                filters=200,\n",
    "                kernel_size=11, \n",
    "                conv_stride=2,\n",
    "                conv_border_mode='valid',\n",
    "                units=250,\n",
    "                activation='relu',\n",
    "                dropout_rate=1,\n",
    "                number_of_layers=5,\n",
    "                output_dim=len(char_map)+1)\n",
    "\n",
    "train(audio_gen, input_to_softmax=model, model_name=MODEL_NAME, epochs=EPOCHS, minibatch_size=MINI_BATCH_SIZE)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Load saved model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "the_input (InputLayer)       (None, None, 13)          0         \n",
      "_________________________________________________________________\n",
      "layer_1_conv (Conv1D)        (None, None, 200)         28800     \n",
      "_________________________________________________________________\n",
      "conv_batch_norm (BatchNormal (None, None, 200)         800       \n",
      "_________________________________________________________________\n",
      "rnn_1 (GRU)                  (None, None, 250)         338250    \n",
      "_________________________________________________________________\n",
      "bt_rnn_1 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_2 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_2 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_3 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_3 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "rnn_4 (GRU)                  (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_4 (BatchNormalization (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "final_layer_of_rnn (GRU)     (None, None, 250)         375750    \n",
      "_________________________________________________________________\n",
      "bt_rnn_final (BatchNormaliza (None, None, 250)         1000      \n",
      "_________________________________________________________________\n",
      "time_distributed_2 (TimeDist (None, None, 223)         55973     \n",
      "_________________________________________________________________\n",
      "softmax (Activation)         (None, None, 223)         0         \n",
      "=================================================================\n",
      "Total params: 1,931,823\n",
      "Trainable params: 1,928,923\n",
      "Non-trainable params: 2,900\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "from src.char_map import char_map, index_map\n",
    "from src.models import model_2\n",
    "\n",
    "model = model_2(input_dim=13,\n",
    "                filters=200,\n",
    "                kernel_size=11, \n",
    "                conv_stride=2,\n",
    "                conv_border_mode='valid',\n",
    "                units=250,\n",
    "                activation='relu',\n",
    "                dropout_rate=1,\n",
    "                number_of_layers=5,\n",
    "                output_dim=len(char_map)+1)\n",
    "model_name = \"model_2\"\n",
    "model.load_weights('models/' + model_name + '.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from src.plot import plot_hist\n",
    "\n",
    "model_name = \"model_2\"\n",
    "plot_hist(model_name)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Word Error Rate - train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From /home/ubuntu/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py:4303: sparse_to_dense (from tensorflow.python.ops.sparse_ops) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Create a `tf.sparse.SparseTensor` and use `tf.sparse.to_dense` instead.\n",
      "wer: 13\n",
      "wer: 4\n",
      "wer: 23\n",
      "processed 5560 in 0 minutes\n",
      "wer: 4\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 6\n",
      "wer: 11\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 5\n",
      "wer: 4\n",
      "processed 5570 in 1 minutes\n",
      "wer: 15\n",
      "wer: 2\n",
      "wer: 11\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 10\n",
      "wer: 15\n",
      "wer: 5\n",
      "wer: 1\n",
      "wer: 4\n",
      "processed 5580 in 2 minutes\n",
      "wer: 7\n",
      "wer: 14\n",
      "wer: 19\n",
      "wer: 10\n",
      "wer: 9\n",
      "wer: 11\n",
      "wer: 2\n",
      "wer: 7\n",
      "wer: 16\n",
      "wer: 9\n",
      "processed 5590 in 3 minutes\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 14\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 5\n",
      "wer: 5\n",
      "wer: 8\n",
      "wer: 6\n",
      "processed 5600 in 4 minutes\n",
      "wer: 5\n",
      "wer: 23\n",
      "wer: 10\n",
      "wer: 8\n",
      "wer: 2\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 1\n",
      "processed 5610 in 5 minutes\n",
      "wer: 8\n",
      "wer: 5\n",
      "wer: 6\n",
      "wer: 1\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 1\n",
      "processed 5620 in 6 minutes\n",
      "wer: 9\n",
      "wer: 4\n",
      "wer: 8\n",
      "wer: 5\n",
      "wer: 5\n",
      "wer: 3\n",
      "wer: 2\n",
      "wer: 5\n",
      "wer: 8\n",
      "wer: 7\n",
      "processed 5630 in 6 minutes\n",
      "wer: 2\n",
      "wer: 3\n",
      "wer: 5\n",
      "wer: 9\n",
      "wer: 13\n",
      "wer: 10\n",
      "wer: 14\n",
      "wer: 9\n",
      "wer: 10\n",
      "wer: 12\n",
      "processed 5640 in 8 minutes\n",
      "wer: 7\n",
      "wer: 5\n",
      "wer: 2\n",
      "wer: 13\n",
      "wer: 3\n",
      "wer: 1\n",
      "wer: 1\n",
      "wer: 8\n",
      "wer: 8\n",
      "wer: 18\n",
      "processed 5650 in 9 minutes\n",
      "wer: 8\n",
      "wer: 1\n",
      "wer: 9\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 9\n",
      "wer: 4\n",
      "wer: 4\n",
      "wer: 19\n",
      "wer: 16\n",
      "processed 5660 in 10 minutes\n",
      "wer: 15\n",
      "wer: 6\n",
      "wer: 2\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 9\n",
      "wer: 3\n",
      "wer: 6\n",
      "wer: 2\n",
      "wer: 17\n",
      "processed 5670 in 11 minutes\n",
      "wer: 16\n",
      "wer: 6\n",
      "wer: 3\n",
      "wer: 3\n",
      "wer: 13\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 5\n",
      "wer: 7\n",
      "wer: 15\n",
      "processed 5680 in 12 minutes\n",
      "wer: 9\n",
      "wer: 7\n",
      "wer: 5\n",
      "wer: 8\n",
      "wer: 2\n",
      "wer: 2\n",
      "wer: 2\n",
      "wer: 15\n",
      "wer: 8\n",
      "wer: 6\n",
      "processed 5690 in 13 minutes\n",
      "wer: 11\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 11\n",
      "wer: 10\n",
      "wer: 10\n",
      "wer: 3\n",
      "wer: 3\n",
      "processed 5700 in 14 minutes\n",
      "wer: 4\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 1\n",
      "wer: 5\n",
      "wer: 2\n",
      "wer: 8\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 2\n",
      "processed 5710 in 15 minutes\n",
      "wer: 7\n",
      "wer: 10\n",
      "wer: 11\n",
      "wer: 3\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 9\n",
      "wer: 7\n",
      "wer: 10\n",
      "wer: 18\n",
      "processed 5720 in 16 minutes\n",
      "wer: 18\n",
      "wer: 3\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 2\n",
      "wer: 8\n",
      "wer: 3\n",
      "wer: 10\n",
      "processed 5730 in 17 minutes\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 15\n",
      "wer: 0\n",
      "wer: 6\n",
      "wer: 2\n",
      "wer: 8\n",
      "wer: 6\n",
      "wer: 8\n",
      "wer: 5\n",
      "processed 5740 in 18 minutes\n",
      "wer: 4\n",
      "wer: 12\n",
      "wer: 11\n",
      "wer: 2\n",
      "wer: 7\n",
      "wer: 1\n",
      "wer: 15\n",
      "wer: 17\n",
      "wer: 3\n",
      "wer: 9\n",
      "processed 5750 in 19 minutes\n",
      "wer: 10\n",
      "wer: 4\n",
      "wer: 9\n",
      "wer: 10\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 12\n",
      "wer: 11\n",
      "processed 5760 in 20 minutes\n",
      "wer: 20\n",
      "wer: 7\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 8\n",
      "wer: 1\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 21\n",
      "wer: 9\n",
      "processed 5770 in 21 minutes\n",
      "wer: 4\n",
      "wer: 5\n",
      "wer: 8\n",
      "wer: 12\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 3\n",
      "wer: 1\n",
      "processed 5780 in 22 minutes\n",
      "wer: 9\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 5\n",
      "wer: 14\n",
      "wer: 13\n",
      "wer: 4\n",
      "wer: 5\n",
      "wer: 1\n",
      "wer: 6\n",
      "processed 5790 in 23 minutes\n",
      "wer: 1\n",
      "wer: 5\n",
      "wer: 1\n",
      "wer: 11\n",
      "wer: 7\n",
      "wer: 16\n",
      "wer: 5\n",
      "wer: 2\n",
      "wer: 1\n",
      "wer: 4\n",
      "processed 5800 in 24 minutes\n",
      "wer: 11\n",
      "wer: 12\n",
      "wer: 6\n",
      "wer: 15\n",
      "wer: 15\n",
      "wer: 13\n",
      "wer: 15\n",
      "wer: 14\n",
      "wer: 9\n",
      "wer: 4\n",
      "processed 5810 in 25 minutes\n",
      "wer: 2\n",
      "wer: 8\n",
      "wer: 11\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 4\n",
      "wer: 17\n",
      "wer: 6\n",
      "wer: 9\n",
      "wer: 8\n",
      "processed 5820 in 26 minutes\n",
      "wer: 4\n",
      "wer: 13\n",
      "wer: 8\n",
      "wer: 9\n",
      "wer: 2\n",
      "wer: 15\n",
      "wer: 8\n",
      "wer: 6\n",
      "wer: 10\n",
      "wer: 4\n",
      "processed 5830 in 27 minutes\n",
      "wer: 12\n",
      "wer: 4\n",
      "wer: 5\n",
      "wer: 17\n",
      "wer: 13\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 13\n",
      "wer: 12\n",
      "wer: 3\n",
      "processed 5840 in 28 minutes\n",
      "wer: 5\n",
      "wer: 10\n",
      "wer: 17\n",
      "wer: 6\n",
      "wer: 10\n",
      "wer: 1\n",
      "wer: 12\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 4\n",
      "processed 5850 in 29 minutes\n",
      "wer: 8\n",
      "wer: 6\n",
      "wer: 13\n",
      "wer: 11\n",
      "wer: 8\n",
      "wer: 0\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 8\n",
      "wer: 13\n",
      "processed 5860 in 30 minutes\n",
      "wer: 3\n",
      "wer: 7\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 9\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 10\n",
      "processed 5870 in 31 minutes\n",
      "wer: 12\n",
      "wer: 8\n",
      "wer: 12\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 20\n",
      "wer: 12\n",
      "wer: 9\n",
      "wer: 5\n",
      "wer: 5\n",
      "processed 5880 in 32 minutes\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 6\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 13\n",
      "wer: 2\n",
      "wer: 7\n",
      "wer: 5\n",
      "processed 5890 in 33 minutes\n",
      "wer: 16\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 13\n",
      "wer: 7\n",
      "wer: 14\n",
      "wer: 4\n",
      "wer: 8\n",
      "wer: 13\n",
      "wer: 15\n",
      "processed 5900 in 35 minutes\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 8\n",
      "wer: 7\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 5\n",
      "wer: 7\n",
      "wer: 7\n",
      "wer: 16\n",
      "processed 5910 in 36 minutes\n",
      "wer: 10\n",
      "wer: 8\n",
      "wer: 2\n",
      "wer: 6\n",
      "wer: 15\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 3\n",
      "wer: 4\n",
      "processed 5920 in 37 minutes\n",
      "wer: 11\n",
      "wer: 6\n",
      "wer: 12\n",
      "wer: 11\n",
      "wer: 7\n",
      "wer: 4\n",
      "wer: 26\n",
      "wer: 4\n",
      "wer: 16\n",
      "wer: 7\n",
      "processed 5930 in 38 minutes\n",
      "wer: 4\n",
      "wer: 4\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 12\n",
      "wer: 9\n",
      "wer: 18\n",
      "wer: 1\n",
      "processed 5940 in 39 minutes\n",
      "wer: 8\n",
      "wer: 5\n",
      "wer: 1\n",
      "wer: 7\n",
      "wer: 6\n",
      "wer: 8\n",
      "wer: 3\n",
      "wer: 4\n",
      "wer: 11\n",
      "wer: 12\n",
      "processed 5950 in 40 minutes\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 12\n",
      "wer: 8\n",
      "wer: 8\n",
      "wer: 8\n",
      "wer: 4\n",
      "wer: 13\n",
      "wer: 6\n",
      "processed 5960 in 41 minutes\n",
      "wer: 6\n",
      "wer: 10\n",
      "wer: 7\n",
      "wer: 2\n",
      "wer: 0\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 2\n",
      "wer: 6\n",
      "wer: 4\n",
      "processed 5970 in 42 minutes\n",
      "wer: 2\n",
      "wer: 9\n",
      "wer: 6\n",
      "wer: 3\n",
      "wer: 22\n",
      "wer: 6\n",
      "wer: 10\n",
      "wer: 31\n",
      "wer: 3\n",
      "wer: 6\n",
      "processed 5980 in 43 minutes\n",
      "wer: 11\n",
      "wer: 2\n",
      "wer: 7\n",
      "wer: 7\n",
      "wer: 13\n",
      "wer: 18\n",
      "wer: 16\n",
      "wer: 12\n",
      "wer: 2\n",
      "wer: 11\n",
      "processed 5990 in 44 minutes\n",
      "wer: 14\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 5\n",
      "wer: 7\n",
      "wer: 1\n",
      "wer: 2\n",
      "wer: 5\n",
      "wer: 9\n",
      "wer: 7\n",
      "processed 6000 in 45 minutes\n",
      "wer: 10\n",
      "wer: 19\n",
      "wer: 11\n",
      "wer: 7\n",
      "wer: 15\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 4\n",
      "wer: 9\n",
      "processed 6010 in 46 minutes\n",
      "wer: 3\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 1\n",
      "wer: 13\n",
      "wer: 2\n",
      "wer: 9\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 1\n",
      "processed 6020 in 47 minutes\n",
      "wer: 10\n",
      "wer: 7\n",
      "wer: 7\n",
      "wer: 19\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 8\n",
      "wer: 9\n",
      "wer: 1\n",
      "wer: 9\n",
      "processed 6030 in 48 minutes\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 14\n",
      "wer: 8\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 8\n",
      "wer: 9\n",
      "wer: 1\n",
      "processed 6040 in 49 minutes\n",
      "wer: 4\n",
      "wer: 10\n",
      "wer: 3\n",
      "wer: 5\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 4\n",
      "wer: 9\n",
      "processed 6050 in 50 minutes\n",
      "wer: 3\n",
      "wer: 3\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 1\n",
      "wer: 5\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 9\n",
      "processed 6060 in 51 minutes\n",
      "wer: 23\n",
      "wer: 15\n",
      "wer: 3\n",
      "wer: 15\n",
      "wer: 9\n",
      "wer: 4\n",
      "wer: 6\n",
      "wer: 4\n",
      "wer: 12\n",
      "wer: 1\n",
      "processed 6070 in 52 minutes\n",
      "wer: 10\n",
      "wer: 3\n",
      "wer: 3\n",
      "wer: 4\n",
      "wer: 10\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 15\n",
      "wer: 10\n",
      "processed 6080 in 54 minutes\n",
      "wer: 6\n",
      "wer: 21\n",
      "wer: 13\n",
      "wer: 2\n",
      "wer: 7\n",
      "wer: 4\n",
      "wer: 13\n",
      "wer: 4\n",
      "wer: 1\n",
      "wer: 17\n",
      "processed 6090 in 55 minutes\n",
      "wer: 5\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 7\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 11\n",
      "wer: 5\n",
      "wer: 2\n",
      "wer: 1\n",
      "processed 6100 in 56 minutes\n",
      "wer: 6\n",
      "wer: 12\n",
      "wer: 4\n",
      "wer: 6\n",
      "wer: 14\n",
      "wer: 10\n",
      "wer: 23\n",
      "wer: 11\n",
      "wer: 15\n",
      "wer: 2\n",
      "processed 6110 in 57 minutes\n",
      "wer: 8\n",
      "wer: 9\n",
      "wer: 5\n",
      "wer: 4\n",
      "wer: 3\n",
      "wer: 9\n",
      "wer: 3\n",
      "wer: 2\n",
      "wer: 4\n",
      "wer: 5\n",
      "processed 6120 in 58 minutes\n",
      "wer: 12\n",
      "wer: 4\n",
      "wer: 6\n",
      "wer: 6\n",
      "wer: 23\n",
      "wer: 2\n",
      "wer: 5\n",
      "wer: 8\n",
      "wer: 2\n",
      "wer: 14\n",
      "processed 6130 in 59 minutes\n",
      "wer: 16\n",
      "wer: 9\n",
      "wer: 6\n",
      "wer: 3\n",
      "wer: 7\n",
      "wer: 12\n",
      "wer: 12\n",
      "wer: 5\n",
      "wer: 10\n",
      "wer: 5\n",
      "processed 6140 in 60 minutes\n",
      "wer: 7\n",
      "wer: 3\n",
      "wer: 10\n",
      "wer: 6\n",
      "wer: 8\n",
      "wer: 3\n",
      "wer: 3\n",
      "wer: 7\n",
      "wer: 1\n",
      "wer: 9\n",
      "processed 6150 in 61 minutes\n",
      "wer: 4\n",
      "wer: 2\n",
      "wer: 13\n",
      "wer: 4\n",
      "wer: 15\n",
      "wer: 5\n",
      "wer: 6\n",
      "wer: 7\n",
      "wer: 6\n",
      "wer: 4\n",
      "processed 6160 in 62 minutes\n",
      "wer: 8\n",
      "wer: 1\n",
      "wer: 11\n",
      "wer: 1\n",
      "wer: 10\n",
      "wer: 11\n",
      "wer: 6\n",
      "wer: 5\n",
      "wer: 3\n",
      "wer: 11\n",
      "processed 6170 in 63 minutes\n",
      "wer: 7\n",
      "wer: 11\n",
      "wer: 16\n",
      "wer: 8\n",
      "wer: 7\n",
      "wer: 11\n",
      "wer: 5\n",
      "wer: 1\n",
      "wer: 15\n",
      "wer: 8\n",
      "processed 6180 in 64 minutes\n",
      "wer: 18\n",
      "wer: 6\n",
      "wer: 9\n",
      "wer: 18\n",
      "wer: 12\n",
      "wer: 15\n",
      "wer: 2\n"
     ]
    }
   ],
   "source": [
    "from src.predict import calculate_wer\n",
    "\n",
    "calculate_wer(model, model_name, audio_gen, 'train', audio_gen.train_length)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from src.predict import load_wer\n",
    "\n",
    "wer = load_wer(model_name, 'train')\n",
    "print(\"WER mean=%f std=%f\" % (np.std(wer), np.mean(wer)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Word Error Rate - validation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from src.predict import calculate_wer\n",
    "from src.data_generator import VALID_LENGTH\n",
    "\n",
    "calculate_wer(model, model_name, 'validation', VALID_LENGTH)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from src.predict import load_wer\n",
    "\n",
    "wer = load_wer(model_name, 'validation')\n",
    "print(\"WER mean=%f std=%f\" % (np.std(wer), np.mean(wer)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Sample prediction - training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pickle\n",
    "from src.predict import predict\n",
    "\n",
    "predict(0, 'train', model)\n",
    "predict(1, 'train', model)\n",
    "predict(2, 'train', model)\n",
    "predict(3, 'train', model)\n",
    "predict(4, 'train', model)\n",
    "predict(5, 'train', model)\n",
    "predict(6, 'train', model)\n",
    "predict(7, 'train', model)\n",
    "predict(8, 'train', model)\n",
    "predict(9, 'train', model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Sample prediction - validation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pickle\n",
    "from src.predict import predict\n",
    "\n",
    "wer = pickle.load(open('models/' + model_name + '_validation_wer.pickle', \"rb\" ))\n",
    "wer_index = np.argsort(wer)\n",
    "predict(wer_index[0], 'validation', model)\n",
    "predict(wer_index[1], 'validation', model)\n",
    "predict(wer_index[2], 'validation', model)\n",
    "predict(wer_index[3], 'validation', model)\n",
    "predict(wer_index[4], 'validation', model)\n",
    "predict(wer_index[5], 'validation', model)\n",
    "predict(wer_index[-1], 'validation', model)\n",
    "predict(wer_index[-2], 'validation', model)\n",
    "predict(wer_index[-3], 'validation', model)\n",
    "predict(wer_index[-4], 'validation', model)\n",
    "predict(wer_index[-5], 'validation', model)\n",
    "predict(wer_index[-6], 'validation', model)"
   ]
  },
  {
   "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": "3.5.2"
  },
  "varInspector": {
   "cols": {
    "lenName": 16,
    "lenType": 16,
    "lenVar": 40
   },
   "kernels_config": {
    "python": {
     "delete_cmd_postfix": "",
     "delete_cmd_prefix": "del ",
     "library": "var_list.py",
     "varRefreshCmd": "print(var_dic_list())"
    },
    "r": {
     "delete_cmd_postfix": ") ",
     "delete_cmd_prefix": "rm(",
     "library": "var_list.r",
     "varRefreshCmd": "cat(var_dic_list()) "
    }
   },
   "types_to_exclude": [
    "module",
    "function",
    "builtin_function_or_method",
    "instance",
    "_Feature"
   ],
   "window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}