EXPLORATION/EXPLORATION 11. 작사가 인공지능 만들기.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# EXPLORATION 11. 작사가 인공지능 만들기"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Import"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import glob\n",
    "import re\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "\n",
    "import tensorflow as tf\n",
    "from sklearn.model_selection import train_test_split"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## utils"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "def preprocess_sentence(sentence):\n",
    "    sentence = sentence.lower().strip()       # 소문자로 바꾸고 양쪽 공백을 삭제\n",
    "  \n",
    "    # 아래 3단계를 거쳐 sentence는 스페이스 1개를 delimeter로 하는 소문자 단어 시퀀스로 바뀝니다.\n",
    "    sentence = re.sub(r\"([?.!,¿])\", r\" \\1 \", sentence)        # 패턴의 특수문자를 만나면 특수문자 양쪽에 공백을 추가\n",
    "    sentence = re.sub(r'[\" \"]+', \" \", sentence)                  # 공백 패턴을 만나면 스페이스 1개로 치환\n",
    "    sentence = re.sub(r\"[^a-zA-Z?.!,¿]+\", \" \", sentence)  # a-zA-Z?.!,¿ 패턴을 제외한 모든 문자(공백문자까지도)를 스페이스 1개로 치환\n",
    "\n",
    "    sentence = sentence.strip()\n",
    "\n",
    "    sentence = '<start> ' + sentence + ' <end>'      # 이전 스텝에서 본 것처럼 문장 앞뒤로 <start>와 <end>를 단어처럼 붙여 줍니다\n",
    "    \n",
    "    return sentence"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "def tokenize(corpus):\n",
    "    # 텐서플로우에서 제공하는 Tokenizer 패키지를 생성\n",
    "    tokenizer = tf.keras.preprocessing.text.Tokenizer(\n",
    "        num_words=7000,  # 전체 단어의 개수 \n",
    "        filters=' ',    # 별도로 전처리 로직을 추가할 수 있습니다. 이번에는 사용하지 않겠습니다.\n",
    "        oov_token=\"<unk>\"  # out-of-vocabulary, 사전에 없었던 단어는 어떤 토큰으로 대체할지\n",
    "    )\n",
    "    tokenizer.fit_on_texts(corpus)   # 우리가 구축한 corpus로부터 Tokenizer가 사전을 자동구축하게 됩니다.\n",
    "\n",
    "    # 이후 tokenizer를 활용하여 모델에 입력할 데이터셋을 구축하게 됩니다.\n",
    "    tensor = tokenizer.texts_to_sequences(corpus)   # tokenizer는 구축한 사전으로부터 corpus를 해석해 Tensor로 변환합니다.\n",
    "\n",
    "    # 입력 데이터의 시퀀스 길이를 일정하게 맞추기 위한 padding  메소드를 제공합니다.\n",
    "    # maxlen의 디폴트값은 None입니다. 이 경우 corpus의 가장 긴 문장을 기준으로 시퀀스 길이가 맞춰집니다.\n",
    "    tensor = tf.keras.preprocessing.sequence.pad_sequences(tensor, padding='post', maxlen=15)  \n",
    "\n",
    "    print(tensor,tokenizer)\n",
    "    return tensor, tokenizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def plot_curve(epochs, hist, list_of_metrics):\n",
    "    \n",
    "    fig, ax = plt.subplots(1,2,figsize = (12, 8))\n",
    "    \n",
    "    for i in range(len(ax)):\n",
    "        ax[i].set_xlabel('Epochs')\n",
    "        ax[i].set_ylabel('Value')\n",
    "        \n",
    "        for n in range(len(list_of_metrics)):\n",
    "            if i == 0:\n",
    "                y = hist[list_of_metrics[n]]\n",
    "                if n == 0:\n",
    "                    ax[i].plot(epochs, y, label=\"train\")\n",
    "                else:\n",
    "                    ax[i].plot(epochs, y, label=\"val\")\n",
    "                ax[i].set_title('Loss')\n",
    "                ax[i].legend(loc='upper right')\n",
    "                if n == 1:\n",
    "                    break\n",
    "            else:\n",
    "                if n >= 2:\n",
    "                    y = hist[list_of_metrics[n]]\n",
    "                    if n == 2:\n",
    "                        ax[i].plot(epochs, y, label=\"train\")\n",
    "                    else:\n",
    "                        ax[i].plot(epochs, y, label=\"val\")\n",
    "                    ax[i].set_title('Accuracy')\n",
    "                    ax[i].legend(loc='lower right')\n",
    "                    \n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def generate_text(model, tokenizer, init_sentence=\"<start>\", max_len=20):\n",
    "    # 테스트를 위해서 입력받은 init_sentence도 일단 텐서로 변환합니다.\n",
    "    test_input = tokenizer.texts_to_sequences([init_sentence])\n",
    "    test_tensor = tf.convert_to_tensor(test_input, dtype=tf.int64)\n",
    "    end_token = tokenizer.word_index[\"<end>\"]\n",
    "\n",
    "    # 텍스트를 실제로 생성할때는 루프를 돌면서 단어 하나씩 생성해야 합니다. \n",
    "    while True:\n",
    "        predict = model(test_tensor)  # 입력받은 문장의 텐서를 입력합니다. \n",
    "        predict_word = tf.argmax(tf.nn.softmax(predict, axis=-1), axis=-1)[:, -1]   # 우리 모델이 예측한 마지막 단어가 바로 새롭게 생성한 단어가 됩니다. \n",
    "\n",
    "        # 우리 모델이 새롭게 예측한 단어를 입력 문장의 뒤에 붙여 줍니다. \n",
    "        test_tensor = tf.concat([test_tensor, \n",
    "        tf.expand_dims(predict_word, axis=0)], axis=-1)\n",
    "\n",
    "        # 우리 모델이 <END>를 예측했거나, max_len에 도달하지 않았다면  while 루프를 또 돌면서 다음 단어를 예측해야 합니다.\n",
    "        if predict_word.numpy()[0] == end_token: break\n",
    "        if test_tensor.shape[1] >= max_len: break\n",
    "\n",
    "    generated = \"\"\n",
    "    # 생성된 tensor 안에 있는 word index를 tokenizer.index_word 사전을 통해 실제 단어로 하나씩 변환합니다. \n",
    "    for word_index in test_tensor[0].numpy():\n",
    "        generated += tokenizer.index_word[word_index] + \" \"\n",
    "\n",
    "    return generated   # 이것이 최종적으로 모델이 생성한 자연어 문장입니다."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "데이터 크기: 187088\n",
      "Examples:\n",
      " ['Hey, Vietnam, Vietnam, Vietnam, Vietnam', 'Vietnam, Vietnam, Vietnam Yesterday I got a letter from my friend', 'Fighting in Vietnam']\n"
     ]
    }
   ],
   "source": [
    "txt_file_path = os.getenv('HOME')+'/aiffel//lyricist/data/lyrics/*'\n",
    "\n",
    "txt_list = glob.glob(txt_file_path)\n",
    "\n",
    "raw_corpus = []\n",
    "\n",
    "# 여러개의 txt 파일을 모두 읽어서 raw_corpus 에 담습니다.\n",
    "for txt_file in txt_list:\n",
    "    with open(txt_file, \"r\") as f:\n",
    "        raw = f.read().splitlines()\n",
    "        raw_corpus.extend(raw)\n",
    "\n",
    "print(\"데이터 크기:\", len(raw_corpus))\n",
    "print(\"Examples:\\n\", raw_corpus[:3])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Data Preprocessing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<start> hey , vietnam , vietnam , vietnam , vietnam <end>',\n",
       " '<start> vietnam , vietnam , vietnam yesterday i got a letter from my friend <end>',\n",
       " '<start> fighting in vietnam <end>',\n",
       " '<start> and this is what he had to say <end>',\n",
       " '<start> tell all my friends that i ll be coming home soon <end>',\n",
       " '<start> my time it ll be up some time in june <end>',\n",
       " '<start> don t forget , he said to tell my sweet mary <end>',\n",
       " '<start> her golden lips as sweet as cherries and it came from <end>',\n",
       " '<start> vietnam , vietnam , vietnam , vietnam <end>',\n",
       " '<start> vietnam , vietnam , vietnam it was just the next day his mother got a telegram <end>']"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "corpus = []\n",
    "\n",
    "for sentence in raw_corpus:\n",
    "    if len(sentence) == 0: continue\n",
    "    if sentence[-1] == \":\": continue\n",
    "        \n",
    "    corpus.append(preprocess_sentence(sentence))\n",
    "        \n",
    "corpus[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[   2  135    4 ...    0    0    0]\n",
      " [   2 1350    4 ...   13  308    3]\n",
      " [   2 1043   14 ...    0    0    0]\n",
      " ...\n",
      " [   2    5   61 ...    0    0    0]\n",
      " [   2  113  682 ...    0    0    0]\n",
      " [   2    8   50 ...    0    0    0]] <keras_preprocessing.text.Tokenizer object at 0x7f1714f8fa10>\n"
     ]
    }
   ],
   "source": [
    "tensor, tokenizer = tokenize(corpus)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[   2  135    4 1350    4 1350    4 1350    4 1350    3    0    0    0]\n",
      "[ 135    4 1350    4 1350    4 1350    4 1350    3    0    0    0    0]\n"
     ]
    }
   ],
   "source": [
    "src_input = tensor[:, :-1]  # tensor에서 마지막 토큰을 잘라내서 소스 문장을 생성합니다. 마지막 토큰은 <END>가 아니라 <pad>일 가능성이 높습니다.\n",
    "tgt_input = tensor[:, 1:]    # tensor에서 <START>를 잘라내서 타겟 문장을 생성합니다.\n",
    "\n",
    "print(src_input[0])\n",
    "print(tgt_input[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "enc_train, enc_val, dec_train, dec_val = train_test_split(src_input, tgt_input, test_size=0.2, random_state=2020)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "BUFFER_SIZE = len(src_input)\n",
    "BATCH_SIZE = 256\n",
    "steps_per_epoch = len(src_input) // BATCH_SIZE\n",
    "\n",
    "VOCAB_SIZE = tokenizer.num_words + 1    # tokenizer가 구축한 단어사전 내 7000개와, 여기 포함되지 않은 0:<pad>를 포함하여 7001개"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<BatchDataset shapes: ((256, 14), (256, 14)), types: (tf.int32, tf.int32)>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "dataset = tf.data.Dataset.from_tensor_slices((enc_train, dec_train)).shuffle(BUFFER_SIZE)\n",
    "dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)\n",
    "dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<BatchDataset shapes: ((256, 14), (256, 14)), types: (tf.int32, tf.int32)>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "val_dataset = tf.data.Dataset.from_tensor_slices((enc_val, dec_val)).shuffle(BUFFER_SIZE)\n",
    "val_dataset = val_dataset.batch(BATCH_SIZE, drop_remainder=True)\n",
    "val_dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "class TextGenerator(tf.keras.Model):\n",
    "    def __init__(self, vocab_size, embedding_size, hidden_size):\n",
    "        super(TextGenerator, self).__init__()\n",
    "        \n",
    "        self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_size)\n",
    "        self.rnn_1 = tf.keras.layers.LSTM(hidden_size, return_sequences=True)\n",
    "        self.drop1 = tf.keras.layers.Dropout(0.3)\n",
    "        self.rnn_2 = tf.keras.layers.LSTM(hidden_size, return_sequences=True)\n",
    "        self.linear = tf.keras.layers.Dense(vocab_size)\n",
    "        \n",
    "    def call(self, x):\n",
    "        out = self.embedding(x)\n",
    "        out = self.rnn_1(out)\n",
    "        out = self.drop1(out)\n",
    "        out = self.rnn_2(out)\n",
    "        out = self.linear(out)\n",
    "        \n",
    "        return out\n",
    "    \n",
    "embedding_size = 2048\n",
    "hidden_size = 2048\n",
    "model = TextGenerator(tokenizer.num_words + 1, embedding_size , hidden_size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<tf.Tensor: shape=(256, 14, 7001), dtype=float32, numpy=\n",
       "array([[[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [-1.6151645e-04,  8.5445371e-04,  6.3464144e-04, ...,\n",
       "          1.1881832e-03, -2.7399493e-04,  1.5679382e-03],\n",
       "        [-3.9252752e-04,  6.3220767e-04,  8.5498398e-04, ...,\n",
       "          1.4483191e-03, -9.3175890e-04,  1.8981083e-03],\n",
       "        ...,\n",
       "        [ 3.9022549e-03,  1.5955748e-03,  1.8970459e-04, ...,\n",
       "          3.5897671e-04, -9.2148337e-05,  1.9062869e-03],\n",
       "        [ 4.0667811e-03,  1.3115890e-03, -2.7780482e-04, ...,\n",
       "          1.0592234e-04,  2.5221612e-04,  1.7512824e-03],\n",
       "        [ 3.4290787e-03,  7.4810215e-04, -7.5264013e-04, ...,\n",
       "         -5.5265392e-04,  1.1828416e-03,  2.2386902e-03]],\n",
       "\n",
       "       [[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [-6.8696530e-04, -7.1341189e-04,  1.0045312e-03, ...,\n",
       "          1.7465261e-03, -8.3888200e-04,  1.5486403e-03],\n",
       "        [-8.2833448e-04, -1.7398659e-03,  8.3019154e-04, ...,\n",
       "          1.7954031e-03, -1.2055320e-03,  2.1898146e-03],\n",
       "        ...,\n",
       "        [-1.2948883e-03, -3.9993445e-03, -1.2534321e-03, ...,\n",
       "         -2.7231500e-03,  5.6345146e-03,  4.2364146e-03],\n",
       "        [-1.7263282e-03, -4.4809030e-03, -1.3912532e-03, ...,\n",
       "         -3.9886227e-03,  6.5975836e-03,  5.7826969e-03],\n",
       "        [-2.0366788e-03, -4.9541630e-03, -1.4510246e-03, ...,\n",
       "         -5.2017937e-03,  7.3515037e-03,  7.3413234e-03]],\n",
       "\n",
       "       [[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [-1.6151645e-04,  8.5445371e-04,  6.3464144e-04, ...,\n",
       "          1.1881832e-03, -2.7399493e-04,  1.5679382e-03],\n",
       "        [ 6.6782944e-05,  2.3104174e-03,  5.4162613e-04, ...,\n",
       "          9.9611655e-04, -1.0417688e-03,  9.0497907e-04],\n",
       "        ...,\n",
       "        [-8.0262529e-05, -1.6905076e-04, -2.3874662e-04, ...,\n",
       "         -2.1839379e-03,  2.2557771e-03,  1.1943307e-03],\n",
       "        [-8.3167118e-04, -1.1789228e-03, -5.7587068e-04, ...,\n",
       "         -3.2669466e-03,  3.5289738e-03,  3.2040044e-03],\n",
       "        [-1.4508509e-03, -2.1531312e-03, -8.1860326e-04, ...,\n",
       "         -4.3523549e-03,  4.6602092e-03,  5.2094175e-03]],\n",
       "\n",
       "       ...,\n",
       "\n",
       "       [[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [ 3.7171389e-04, -1.7844359e-04,  3.4582618e-04, ...,\n",
       "          1.8330604e-03,  5.3297938e-04,  1.4624995e-03],\n",
       "        [ 4.6573003e-04, -6.9836696e-04,  1.0530804e-03, ...,\n",
       "          1.9795257e-03,  1.0253018e-04,  2.9105335e-03],\n",
       "        ...,\n",
       "        [ 1.2134666e-03, -3.3916526e-03, -1.7461932e-03, ...,\n",
       "         -2.0929223e-03,  4.2845905e-03,  3.4357670e-03],\n",
       "        [ 3.4030175e-04, -3.9708572e-03, -1.9584231e-03, ...,\n",
       "         -3.2081313e-03,  5.6069987e-03,  4.8073176e-03],\n",
       "        [-4.0537192e-04, -4.5428206e-03, -2.0644744e-03, ...,\n",
       "         -4.3288544e-03,  6.7030299e-03,  6.2830178e-03]],\n",
       "\n",
       "       [[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [-1.6151645e-04,  8.5445371e-04,  6.3464144e-04, ...,\n",
       "          1.1881832e-03, -2.7399493e-04,  1.5679382e-03],\n",
       "        [-4.1117321e-05,  9.1524969e-04,  4.5021577e-04, ...,\n",
       "          1.7758497e-03, -1.4138516e-04,  1.0144478e-03],\n",
       "        ...,\n",
       "        [-1.1542349e-03,  2.5190383e-03, -2.7552820e-03, ...,\n",
       "         -1.1187278e-03, -7.1458955e-04, -2.8003692e-03],\n",
       "        [ 1.0159365e-03,  2.3766442e-03, -2.6961949e-03, ...,\n",
       "         -1.2698271e-03, -1.3945791e-03, -4.2080944e-03],\n",
       "        [ 2.0043494e-03,  1.9516401e-03, -2.7839586e-03, ...,\n",
       "         -1.5520008e-03, -9.9816953e-04, -4.1264049e-03]],\n",
       "\n",
       "       [[ 1.3450939e-04,  1.9431842e-04,  6.5826345e-04, ...,\n",
       "          6.7252561e-04, -2.5257989e-04,  1.0152693e-03],\n",
       "        [ 2.2641444e-04, -9.2051101e-05,  7.1450032e-04, ...,\n",
       "          2.6143808e-04, -3.7541753e-04,  1.3495351e-03],\n",
       "        [-4.9780577e-04, -9.2131709e-04,  3.3080493e-04, ...,\n",
       "          4.2659434e-05, -4.4952723e-04,  1.5020686e-03],\n",
       "        ...,\n",
       "        [ 2.5158385e-03, -3.3051395e-03, -1.7769173e-03, ...,\n",
       "         -1.2499240e-03, -2.9406648e-03, -1.9939812e-03],\n",
       "        [ 2.4392256e-03, -3.5622136e-03, -2.1106920e-03, ...,\n",
       "         -1.8713351e-03, -1.6534531e-03, -1.3120663e-03],\n",
       "        [ 1.9852710e-03, -3.9218934e-03, -2.3530663e-03, ...,\n",
       "         -2.6956543e-03, -8.6609107e-05, -5.3011110e-05]]], dtype=float32)>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "for src_sample, tgt_sample in dataset.take(1): break\n",
    "model(src_sample)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"text_generator\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "embedding (Embedding)        multiple                  14338048  \n",
      "_________________________________________________________________\n",
      "lstm (LSTM)                  multiple                  33562624  \n",
      "_________________________________________________________________\n",
      "dropout (Dropout)            multiple                  0         \n",
      "_________________________________________________________________\n",
      "lstm_1 (LSTM)                multiple                  33562624  \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          multiple                  0         \n",
      "_________________________________________________________________\n",
      "dense (Dense)                multiple                  14345049  \n",
      "=================================================================\n",
      "Total params: 95,808,345\n",
      "Trainable params: 95,808,345\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)\n",
    "loss = tf.keras.losses.SparseCategoricalCrossentropy(\n",
    "    from_logits=True,\n",
    "    reduction='none'\n",
    ")\n",
    "\n",
    "model.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/7\n",
      "549/549 [==============================] - 288s 525ms/step - loss: 3.2488 - accuracy: 0.4806 - val_loss: 2.8874 - val_accuracy: 0.5103\n",
      "Epoch 2/7\n",
      "549/549 [==============================] - 289s 527ms/step - loss: 2.7417 - accuracy: 0.5219 - val_loss: 2.6394 - val_accuracy: 0.5355\n",
      "Epoch 3/7\n",
      "549/549 [==============================] - 289s 526ms/step - loss: 2.4669 - accuracy: 0.5491 - val_loss: 2.4792 - val_accuracy: 0.5569\n",
      "Epoch 4/7\n",
      "549/549 [==============================] - 289s 527ms/step - loss: 2.2227 - accuracy: 0.5777 - val_loss: 2.3597 - val_accuracy: 0.5769\n",
      "Epoch 5/7\n",
      "549/549 [==============================] - 289s 527ms/step - loss: 2.0015 - accuracy: 0.6075 - val_loss: 2.2731 - val_accuracy: 0.5948\n",
      "Epoch 6/7\n",
      "549/549 [==============================] - 290s 528ms/step - loss: 1.8090 - accuracy: 0.6369 - val_loss: 2.2132 - val_accuracy: 0.6091\n",
      "Epoch 7/7\n",
      "549/549 [==============================] - 290s 528ms/step - loss: 1.6495 - accuracy: 0.6638 - val_loss: 2.1772 - val_accuracy: 0.6212\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(dataset,\n",
    "          epochs=7,\n",
    "          validation_data=val_dataset,\n",
    "          verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x576 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_curve(history.epoch, history.history, ['loss', 'val_loss', 'accuracy', 'val_accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generate lyrics"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'<start> i love you , liberian girl <end> '"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "generate_text(model, tokenizer, init_sentence=\"<start> i love\", max_len=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'<start> i am not throwing away my shot <end> '"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "generate_text(model, tokenizer, init_sentence=\"<start> i am\", max_len=20)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Report"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Model  \n",
    "LSTM layer 사이에 Dropout(0.3)을 적용하였습니다.  \n",
    "embedding_size, hidden_size를 각각 다음과 같이 설정하였습니다.  \n",
    "embedding_size = 2048, hidden_size = 2048  \n",
    "\n",
    "<br>\n",
    "\n",
    "최종적으로 val_loss를 2.1772을 얻었습니다.  \n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "aiffel",
   "language": "python",
   "name": "aiffel"
  },
  "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"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": true,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {
    "height": "758px",
    "left": "168px",
    "top": "199px",
    "width": "280.797px"
   },
   "toc_section_display": true,
   "toc_window_display": true
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}