6.CHATBOT/6.4.seq2seq.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 패키지 불러오기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import os\n",
    "\n",
    "from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from preprocess import *"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 시각화 함수"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "metadata": {},
   "outputs": [],
   "source": [
    "def plot_graphs(history, string):\n",
    "    plt.plot(history.history[string])\n",
    "    plt.plot(history.history['val_'+string], '')\n",
    "    plt.xlabel(\"Epochs\")\n",
    "    plt.ylabel(string)\n",
    "    plt.legend([string, 'val_'+string])\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 학습 데이터 경로 정의"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 129,
   "metadata": {},
   "outputs": [],
   "source": [
    "DATA_IN_PATH = './data_in/'\n",
    "DATA_OUT_PATH = './data_out/'\n",
    "TRAIN_INPUTS = 'train_inputs.npy'\n",
    "TRAIN_OUTPUTS = 'train_outputs.npy'\n",
    "TRAIN_TARGETS = 'train_targets.npy'\n",
    "DATA_CONFIGS = 'data_configs.json'"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 랜덤 시드 고정"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "metadata": {},
   "outputs": [],
   "source": [
    "SEED_NUM = 1234\n",
    "tf.random.set_seed(SEED_NUM)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 파일 로드"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 131,
   "metadata": {},
   "outputs": [],
   "source": [
    "index_inputs = np.load(open(DATA_IN_PATH + TRAIN_INPUTS, 'rb'))\n",
    "index_outputs = np.load(open(DATA_IN_PATH + TRAIN_OUTPUTS , 'rb'))\n",
    "index_targets = np.load(open(DATA_IN_PATH + TRAIN_TARGETS , 'rb'))\n",
    "prepro_configs = json.load(open(DATA_IN_PATH + DATA_CONFIGS, 'r'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 132,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "20 20 20\n"
     ]
    }
   ],
   "source": [
    "# Show length\n",
    "print(len(index_inputs),  len(index_outputs), len(index_targets))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 모델 만들기에 필요한 값 선언"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 133,
   "metadata": {},
   "outputs": [],
   "source": [
    "MODEL_NAME = 'seq2seq_kor'\n",
    "BATCH_SIZE = 2\n",
    "MAX_SEQUENCE = 25\n",
    "EPOCH = 30\n",
    "UNITS = 1024\n",
    "EMBEDDING_DIM = 256\n",
    "VALIDATION_SPLIT = 0.1 \n",
    "\n",
    "char2idx = prepro_configs['char2idx']\n",
    "idx2char = prepro_configs['idx2char']\n",
    "std_index = prepro_configs['std_symbol']\n",
    "end_index = prepro_configs['end_symbol']\n",
    "vocab_size = prepro_configs['vocab_size']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 모델 "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 인코더"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 134,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Encoder(tf.keras.layers.Layer):\n",
    "    def __init__(self, vocab_size, embedding_dim, enc_units, batch_sz):\n",
    "        super(Encoder, self).__init__()\n",
    "        self.batch_sz = batch_sz\n",
    "        self.enc_units = enc_units\n",
    "        self.vocab_size = vocab_size \n",
    "        self.embedding_dim = embedding_dim          \n",
    "        \n",
    "        self.embedding = tf.keras.layers.Embedding(self.vocab_size, self.embedding_dim)\n",
    "        self.gru = tf.keras.layers.GRU(self.enc_units,\n",
    "                                       return_sequences=True,\n",
    "                                       return_state=True,\n",
    "                                       recurrent_initializer='glorot_uniform')\n",
    "\n",
    "    def call(self, x, hidden):\n",
    "        x = self.embedding(x)\n",
    "        output, state = self.gru(x, initial_state = hidden)\n",
    "        return output, state\n",
    "\n",
    "    def initialize_hidden_state(self, inp):\n",
    "        return tf.zeros((tf.shape(inp)[0], self.enc_units))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 어텐션"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 135,
   "metadata": {},
   "outputs": [],
   "source": [
    "class BahdanauAttention(tf.keras.layers.Layer):\n",
    "    def __init__(self, units):\n",
    "        super(BahdanauAttention, self).__init__()\n",
    "        self.W1 = tf.keras.layers.Dense(units)\n",
    "        self.W2 = tf.keras.layers.Dense(units)\n",
    "        self.V = tf.keras.layers.Dense(1)\n",
    "\n",
    "    def call(self, query, values):\n",
    "        hidden_with_time_axis = tf.expand_dims(query, 1)\n",
    "\n",
    "        score = self.V(tf.nn.tanh(\n",
    "            self.W1(values) + self.W2(hidden_with_time_axis)))\n",
    "\n",
    "        attention_weights = tf.nn.softmax(score, axis=1)\n",
    "\n",
    "        context_vector = attention_weights * values\n",
    "        context_vector = tf.reduce_sum(context_vector, axis=1)\n",
    "\n",
    "        return context_vector, attention_weights"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 디코더"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 136,
   "metadata": {},
   "outputs": [],
   "source": [
    "class Decoder(tf.keras.layers.Layer):\n",
    "    def __init__(self, vocab_size, embedding_dim, dec_units, batch_sz):\n",
    "        super(Decoder, self).__init__()\n",
    "        \n",
    "        self.batch_sz = batch_sz\n",
    "        self.dec_units = dec_units\n",
    "        self.vocab_size = vocab_size \n",
    "        self.embedding_dim = embedding_dim  \n",
    "        \n",
    "        self.embedding = tf.keras.layers.Embedding(self.vocab_size, self.embedding_dim)\n",
    "        self.gru = tf.keras.layers.GRU(self.dec_units,\n",
    "                                       return_sequences=True,\n",
    "                                       return_state=True,\n",
    "                                       recurrent_initializer='glorot_uniform')\n",
    "        self.fc = tf.keras.layers.Dense(self.vocab_size)\n",
    "\n",
    "        self.attention = BahdanauAttention(self.dec_units)\n",
    "        \n",
    "    def call(self, x, hidden, enc_output):\n",
    "        context_vector, attention_weights = self.attention(hidden, enc_output)\n",
    "\n",
    "        x = self.embedding(x)\n",
    "\n",
    "        x = tf.concat([tf.expand_dims(context_vector, 1), x], axis=-1)\n",
    "\n",
    "        output, state = self.gru(x)\n",
    "        output = tf.reshape(output, (-1, output.shape[2]))\n",
    "            \n",
    "        x = self.fc(output)\n",
    "        \n",
    "        return x, state, attention_weights"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 137,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = tf.keras.optimizers.Adam()\n",
    "\n",
    "loss_object = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction='none')\n",
    "\n",
    "train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy')\n",
    "\n",
    "def loss(real, pred):\n",
    "    mask = tf.math.logical_not(tf.math.equal(real, 0))\n",
    "    loss_ = loss_object(real, pred)\n",
    "    mask = tf.cast(mask, dtype=loss_.dtype)\n",
    "    loss_ *= mask\n",
    "    return tf.reduce_mean(loss_)\n",
    "\n",
    "def accuracy(real, pred):\n",
    "    mask = tf.math.logical_not(tf.math.equal(real, 0))\n",
    "    mask = tf.expand_dims(tf.cast(mask, dtype=pred.dtype), axis=-1)\n",
    "    pred *= mask    \n",
    "    acc = train_accuracy(real, pred)\n",
    "\n",
    "    return tf.reduce_mean(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 시퀀스 투 스퀀스 모델"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 138,
   "metadata": {},
   "outputs": [],
   "source": [
    "class seq2seq(tf.keras.Model):\n",
    "    def __init__(self, vocab_size, embedding_dim, enc_units, dec_units, batch_sz, end_token_idx=2):    \n",
    "        super(seq2seq, self).__init__()\n",
    "        self.end_token_idx = end_token_idx\n",
    "        self.encoder = Encoder(vocab_size, embedding_dim, enc_units, batch_sz) \n",
    "        self.decoder = Decoder(vocab_size, embedding_dim, dec_units, batch_sz) \n",
    "\n",
    "    def call(self, x):\n",
    "        inp, tar = x\n",
    "        \n",
    "        enc_hidden = self.encoder.initialize_hidden_state(inp)\n",
    "        enc_output, enc_hidden = self.encoder(inp, enc_hidden)\n",
    "\n",
    "        dec_hidden = enc_hidden\n",
    "\n",
    "        predict_tokens = list()\n",
    "        for t in range(0, tar.shape[1]):\n",
    "            dec_input = tf.dtypes.cast(tf.expand_dims(tar[:, t], 1), tf.float32) \n",
    "            predictions, dec_hidden, _ = self.decoder(dec_input, dec_hidden, enc_output)\n",
    "            predict_tokens.append(tf.dtypes.cast(predictions, tf.float32))   \n",
    "        return tf.stack(predict_tokens, axis=1)\n",
    "    \n",
    "    def inference(self, x):\n",
    "        inp  = x\n",
    "\n",
    "        enc_hidden = self.encoder.initialize_hidden_state(inp)\n",
    "        enc_output, enc_hidden = self.encoder(inp, enc_hidden)\n",
    "\n",
    "        dec_hidden = enc_hidden\n",
    "        \n",
    "        dec_input = tf.expand_dims([char2idx[std_index]], 1)\n",
    "        \n",
    "        predict_tokens = list()\n",
    "        for t in range(0, MAX_SEQUENCE):\n",
    "            predictions, dec_hidden, _ = self.decoder(dec_input, dec_hidden, enc_output)\n",
    "            predict_token = tf.argmax(predictions[0])\n",
    "            \n",
    "            if predict_token == self.end_token_idx:\n",
    "                break\n",
    "            \n",
    "            predict_tokens.append(predict_token)\n",
    "            dec_input = tf.dtypes.cast(tf.expand_dims([predict_token], 0), tf.float32)   \n",
    "            \n",
    "        return tf.stack(predict_tokens, axis=0).numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 139,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = seq2seq(vocab_size, EMBEDDING_DIM, UNITS, UNITS, BATCH_SIZE, char2idx[end_index])\n",
    "model.compile(loss=loss, optimizer=tf.keras.optimizers.Adam(1e-3), metrics=[accuracy])\n",
    "#model.run_eagerly = True"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 학습 진행"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 140,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.8365 - accuracy: 0.8360\n",
      "Epoch 00001: val_accuracy improved from -inf to 0.85400, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 5s 599ms/step - loss: 0.8365 - accuracy: 0.8360 - val_loss: 0.6929 - val_accuracy: 0.8540\n",
      "Epoch 2/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.7576 - accuracy: 0.8527\n",
      "Epoch 00002: val_accuracy improved from 0.85400 to 0.85600, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 91ms/step - loss: 0.7576 - accuracy: 0.8527 - val_loss: 0.6273 - val_accuracy: 0.8560\n",
      "Epoch 3/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.6806 - accuracy: 0.8572\n",
      "Epoch 00003: val_accuracy improved from 0.85600 to 0.85667, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 70ms/step - loss: 0.6806 - accuracy: 0.8572 - val_loss: 0.5928 - val_accuracy: 0.8567\n",
      "Epoch 4/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.6296 - accuracy: 0.8572\n",
      "Epoch 00004: val_accuracy improved from 0.85667 to 0.85700, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 88ms/step - loss: 0.6296 - accuracy: 0.8572 - val_loss: 0.5681 - val_accuracy: 0.8570\n",
      "Epoch 5/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.5831 - accuracy: 0.8582\n",
      "Epoch 00005: val_accuracy improved from 0.85700 to 0.85800, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 72ms/step - loss: 0.5831 - accuracy: 0.8582 - val_loss: 0.5146 - val_accuracy: 0.8580\n",
      "Epoch 6/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.5308 - accuracy: 0.8577\n",
      "Epoch 00006: val_accuracy improved from 0.85800 to 0.85900, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 89ms/step - loss: 0.5308 - accuracy: 0.8577 - val_loss: 0.4513 - val_accuracy: 0.8590\n",
      "Epoch 7/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.4739 - accuracy: 0.8598\n",
      "Epoch 00007: val_accuracy improved from 0.85900 to 0.86114, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 72ms/step - loss: 0.4739 - accuracy: 0.8598 - val_loss: 0.4118 - val_accuracy: 0.8611\n",
      "Epoch 8/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.4216 - accuracy: 0.8632\n",
      "Epoch 00008: val_accuracy improved from 0.86114 to 0.86450, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 90ms/step - loss: 0.4216 - accuracy: 0.8632 - val_loss: 0.3691 - val_accuracy: 0.8645\n",
      "Epoch 9/30\n",
      "8/9 [=========================>....] - ETA: 0s - loss: 0.3744 - accuracy: 0.8667\n",
      "Epoch 00009: val_accuracy improved from 0.86450 to 0.86933, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 87ms/step - loss: 0.3652 - accuracy: 0.8669 - val_loss: 0.3232 - val_accuracy: 0.8693\n",
      "Epoch 10/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.3247 - accuracy: 0.8736\n",
      "Epoch 00010: val_accuracy improved from 0.86933 to 0.87780, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 90ms/step - loss: 0.3247 - accuracy: 0.8736 - val_loss: 0.2706 - val_accuracy: 0.8778\n",
      "Epoch 11/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.2797 - accuracy: 0.8816\n",
      "Epoch 00011: val_accuracy improved from 0.87780 to 0.88491, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 89ms/step - loss: 0.2797 - accuracy: 0.8816 - val_loss: 0.2418 - val_accuracy: 0.8849\n",
      "Epoch 12/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.2250 - accuracy: 0.8882\n",
      "Epoch 00012: val_accuracy improved from 0.88491 to 0.89117, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 87ms/step - loss: 0.2250 - accuracy: 0.8882 - val_loss: 0.2023 - val_accuracy: 0.8912\n",
      "Epoch 13/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1984 - accuracy: 0.8938\n",
      "Epoch 00013: val_accuracy improved from 0.89117 to 0.89615, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 72ms/step - loss: 0.1984 - accuracy: 0.8938 - val_loss: 0.2193 - val_accuracy: 0.8962\n",
      "Epoch 14/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1793 - accuracy: 0.8985\n",
      "Epoch 00014: val_accuracy improved from 0.89615 to 0.90043, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 88ms/step - loss: 0.1793 - accuracy: 0.8985 - val_loss: 0.1197 - val_accuracy: 0.9004\n",
      "Epoch 15/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1552 - accuracy: 0.9023\n",
      "Epoch 00015: val_accuracy improved from 0.90043 to 0.90400, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 91ms/step - loss: 0.1552 - accuracy: 0.9023 - val_loss: 0.1431 - val_accuracy: 0.9040\n",
      "Epoch 16/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1389 - accuracy: 0.9059\n",
      "Epoch 00016: val_accuracy improved from 0.90400 to 0.90750, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 90ms/step - loss: 0.1389 - accuracy: 0.9059 - val_loss: 0.1279 - val_accuracy: 0.9075\n",
      "Epoch 17/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1268 - accuracy: 0.9090\n",
      "Epoch 00017: val_accuracy improved from 0.90750 to 0.91047, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 77ms/step - loss: 0.1268 - accuracy: 0.9090 - val_loss: 0.1207 - val_accuracy: 0.9105\n",
      "Epoch 18/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1297 - accuracy: 0.9119\n",
      "Epoch 00018: val_accuracy improved from 0.91047 to 0.91322, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 87ms/step - loss: 0.1297 - accuracy: 0.9119 - val_loss: 0.1242 - val_accuracy: 0.9132\n",
      "Epoch 19/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1235 - accuracy: 0.9146\n",
      "Epoch 00019: val_accuracy improved from 0.91322 to 0.91568, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 73ms/step - loss: 0.1235 - accuracy: 0.9146 - val_loss: 0.1305 - val_accuracy: 0.9157\n",
      "Epoch 20/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1107 - accuracy: 0.9169\n",
      "Epoch 00020: val_accuracy improved from 0.91568 to 0.91790, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 92ms/step - loss: 0.1107 - accuracy: 0.9169 - val_loss: 0.1224 - val_accuracy: 0.9179\n",
      "Epoch 21/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.1014 - accuracy: 0.9190\n",
      "Epoch 00021: val_accuracy improved from 0.91790 to 0.92010, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 82ms/step - loss: 0.1014 - accuracy: 0.9190 - val_loss: 0.1348 - val_accuracy: 0.9201\n",
      "Epoch 22/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0860 - accuracy: 0.9210\n",
      "Epoch 00022: val_accuracy improved from 0.92010 to 0.92209, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 85ms/step - loss: 0.0860 - accuracy: 0.9210 - val_loss: 0.1393 - val_accuracy: 0.9221\n",
      "Epoch 23/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0679 - accuracy: 0.9230\n",
      "Epoch 00023: val_accuracy improved from 0.92209 to 0.92409, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 75ms/step - loss: 0.0679 - accuracy: 0.9230 - val_loss: 0.1486 - val_accuracy: 0.9241\n",
      "Epoch 24/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0564 - accuracy: 0.9254\n",
      "Epoch 00024: val_accuracy improved from 0.92409 to 0.92642, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 87ms/step - loss: 0.0564 - accuracy: 0.9254 - val_loss: 0.1438 - val_accuracy: 0.9264\n",
      "Epoch 25/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0442 - accuracy: 0.9275\n",
      "Epoch 00025: val_accuracy improved from 0.92642 to 0.92848, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 72ms/step - loss: 0.0442 - accuracy: 0.9275 - val_loss: 0.1351 - val_accuracy: 0.9285\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 26/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0379 - accuracy: 0.9297\n",
      "Epoch 00026: val_accuracy improved from 0.92848 to 0.93069, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 87ms/step - loss: 0.0379 - accuracy: 0.9297 - val_loss: 0.1366 - val_accuracy: 0.9307\n",
      "Epoch 27/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0187 - accuracy: 0.9319\n",
      "Epoch 00027: val_accuracy improved from 0.93069 to 0.93289, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 88ms/step - loss: 0.0187 - accuracy: 0.9319 - val_loss: 0.1678 - val_accuracy: 0.9329\n",
      "Epoch 28/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0352 - accuracy: 0.9339\n",
      "Epoch 00028: val_accuracy improved from 0.93289 to 0.93493, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 92ms/step - loss: 0.0352 - accuracy: 0.9339 - val_loss: 0.1361 - val_accuracy: 0.9349\n",
      "Epoch 29/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0210 - accuracy: 0.9359\n",
      "Epoch 00029: val_accuracy improved from 0.93493 to 0.93676, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 77ms/step - loss: 0.0210 - accuracy: 0.9359 - val_loss: 0.1303 - val_accuracy: 0.9368\n",
      "Epoch 30/30\n",
      "9/9 [==============================] - ETA: 0s - loss: 0.0240 - accuracy: 0.9377\n",
      "Epoch 00030: val_accuracy improved from 0.93676 to 0.93847, saving model to ./data_out/seq2seq_kor/weights.h5\n",
      "9/9 [==============================] - 1s 88ms/step - loss: 0.0240 - accuracy: 0.9377 - val_loss: 0.1384 - val_accuracy: 0.9385\n"
     ]
    }
   ],
   "source": [
    "PATH = DATA_OUT_PATH + MODEL_NAME\n",
    "if not(os.path.isdir(PATH)):\n",
    "        os.makedirs(os.path.join(PATH))\n",
    "        \n",
    "checkpoint_path = DATA_OUT_PATH + MODEL_NAME + '/weights.h5'\n",
    "    \n",
    "cp_callback = ModelCheckpoint(\n",
    "    checkpoint_path, monitor='val_accuracy', verbose=1, save_best_only=True, save_weights_only=True)\n",
    "\n",
    "earlystop_callback = EarlyStopping(monitor='val_accuracy', min_delta=0.0001, patience=10)\n",
    "\n",
    "history = model.fit([index_inputs, index_outputs], index_targets,\n",
    "                    batch_size=BATCH_SIZE, epochs=EPOCH,\n",
    "                    validation_split=VALIDATION_SPLIT, callbacks=[earlystop_callback, cp_callback])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 결과 플롯"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 141,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_graphs(history, 'accuracy')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 142,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_graphs(history, 'loss')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 결과 확인"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 143,
   "metadata": {},
   "outputs": [],
   "source": [
    "SAVE_FILE_NM = \"weights.h5\"\n",
    "model.load_weights(os.path.join(DATA_OUT_PATH, MODEL_NAME, SAVE_FILE_NM))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 144,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[83 79 98 97 21 56]\n",
      "평소에 필요했던 게 좋을 것 생각해보세요\n"
     ]
    }
   ],
   "source": [
    "query = \"남자친구 승진 선물로 뭐가 좋을까?\"\n",
    "\n",
    "test_index_inputs, _ = enc_processing([query], char2idx)    \n",
    "predict_tokens = model.inference(test_index_inputs)\n",
    "print(predict_tokens)\n",
    "\n",
    "print(' '.join([idx2char[str(t)] for t in predict_tokens]))"
   ]
  },
  {
   "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": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}