chapter11_part04_sequence-to-sequence-learning.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wnO2lSxrC6Yd"
      },
      "source": [
        "이 노트북은 [케라스 창시자에게 배우는 딥러닝 2판](https://tensorflow.blog/kerasdl2/)의 예제 코드를 담고 있습니다.\n",
        "\n",
        "<table align=\"left\">\n",
        "    <tr>\n",
        "        <td>\n",
        "            <a href=\"https://colab.research.google.com/github/rickiepark/deep-learning-with-python-2nd/blob/main/chapter11_part04_sequence-to-sequence-learning.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
        "        </td>\n",
        "    </tr>\n",
        "</table>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hYUxgYWrC6Yg"
      },
      "source": [
        "## 텍스트 분류를 넘어: 시퀀스-투-시퀀스 학습"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GFpTmSrxC6Yg"
      },
      "source": [
        "### 기계 번역 예제"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "I1ZDSF7AC6Yh",
        "outputId": "2e5794c6-e24a-41cb-fe29-0b2102b517eb"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "--2023-11-04 03:46:10--  http://storage.googleapis.com/download.tensorflow.org/data/spa-eng.zip\n",
            "Resolving storage.googleapis.com (storage.googleapis.com)... 108.177.126.207, 172.217.218.207, 74.125.143.207, ...\n",
            "Connecting to storage.googleapis.com (storage.googleapis.com)|108.177.126.207|:80... connected.\n",
            "HTTP request sent, awaiting response... 200 OK\n",
            "Length: 2638744 (2.5M) [application/zip]\n",
            "Saving to: ‘spa-eng.zip’\n",
            "\n",
            "spa-eng.zip         100%[===================>]   2.52M  3.89MB/s    in 0.6s    \n",
            "\n",
            "2023-11-04 03:46:11 (3.89 MB/s) - ‘spa-eng.zip’ saved [2638744/2638744]\n",
            "\n"
          ]
        }
      ],
      "source": [
        "!wget http://storage.googleapis.com/download.tensorflow.org/data/spa-eng.zip\n",
        "!unzip -q spa-eng.zip"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "HwGu6EZxC6Yh"
      },
      "outputs": [],
      "source": [
        "text_file = \"spa-eng/spa.txt\"\n",
        "with open(text_file) as f:\n",
        "    lines = f.read().split(\"\\n\")[:-1]\n",
        "text_pairs = []\n",
        "for line in lines:\n",
        "    english, spanish = line.split(\"\\t\")\n",
        "    spanish = \"[start] \" + spanish + \" [end]\"\n",
        "    text_pairs.append((english, spanish))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 3,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "j3ARnP9KC6Yi",
        "outputId": "7c9971f4-4626-4c35-f6d5-3169e1acfa96"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "('Tom gave his dog a bath.', '[start] Tom bañó a su perro. [end]')\n"
          ]
        }
      ],
      "source": [
        "import random\n",
        "print(random.choice(text_pairs))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {
        "id": "gcGm7oMfC6Yi"
      },
      "outputs": [],
      "source": [
        "import random\n",
        "random.shuffle(text_pairs)\n",
        "num_val_samples = int(0.15 * len(text_pairs))\n",
        "num_train_samples = len(text_pairs) - 2 * num_val_samples\n",
        "train_pairs = text_pairs[:num_train_samples]\n",
        "val_pairs = text_pairs[num_train_samples:num_train_samples + num_val_samples]\n",
        "test_pairs = text_pairs[num_train_samples + num_val_samples:]"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1c9yTMG4C6Yj"
      },
      "source": [
        "**영어와 스페인어 텍스트 쌍을 벡터화하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 5,
      "metadata": {
        "id": "vQd-CYfvC6Yj"
      },
      "outputs": [],
      "source": [
        "import tensorflow as tf\n",
        "from tensorflow.keras import layers\n",
        "import string\n",
        "import re\n",
        "\n",
        "strip_chars = string.punctuation + \"¿\"\n",
        "strip_chars = strip_chars.replace(\"[\", \"\")\n",
        "strip_chars = strip_chars.replace(\"]\", \"\")\n",
        "\n",
        "def custom_standardization(input_string):\n",
        "    lowercase = tf.strings.lower(input_string)\n",
        "    return tf.strings.regex_replace(\n",
        "        lowercase, f\"[{re.escape(strip_chars)}]\", \"\")\n",
        "\n",
        "vocab_size = 15000\n",
        "sequence_length = 20\n",
        "\n",
        "source_vectorization = layers.TextVectorization(\n",
        "    max_tokens=vocab_size,\n",
        "    output_mode=\"int\",\n",
        "    output_sequence_length=sequence_length,\n",
        ")\n",
        "target_vectorization = layers.TextVectorization(\n",
        "    max_tokens=vocab_size,\n",
        "    output_mode=\"int\",\n",
        "    output_sequence_length=sequence_length + 1,\n",
        "    standardize=custom_standardization,\n",
        ")\n",
        "train_english_texts = [pair[0] for pair in train_pairs]\n",
        "train_spanish_texts = [pair[1] for pair in train_pairs]\n",
        "source_vectorization.adapt(train_english_texts)\n",
        "target_vectorization.adapt(train_spanish_texts)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4tuHosZOC6Yk"
      },
      "source": [
        "**번역 작업을 위한 데이터셋 준비하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 6,
      "metadata": {
        "id": "M7e9MC6PC6Yk"
      },
      "outputs": [],
      "source": [
        "batch_size = 64\n",
        "\n",
        "def format_dataset(eng, spa):\n",
        "    eng = source_vectorization(eng)\n",
        "    spa = target_vectorization(spa)\n",
        "    return ({\n",
        "        \"english\": eng,\n",
        "        \"spanish\": spa[:, :-1],\n",
        "    }, spa[:, 1:])\n",
        "\n",
        "def make_dataset(pairs):\n",
        "    eng_texts, spa_texts = zip(*pairs)\n",
        "    eng_texts = list(eng_texts)\n",
        "    spa_texts = list(spa_texts)\n",
        "    dataset = tf.data.Dataset.from_tensor_slices((eng_texts, spa_texts))\n",
        "    dataset = dataset.batch(batch_size)\n",
        "    dataset = dataset.map(format_dataset, num_parallel_calls=4)\n",
        "    return dataset.shuffle(2048).prefetch(16).cache()\n",
        "\n",
        "train_ds = make_dataset(train_pairs)\n",
        "val_ds = make_dataset(val_pairs)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 7,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "VzIe7PCoC6Yl",
        "outputId": "45d8013f-5ba2-47cb-93a4-234598462d41"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "inputs['english'].shape: (64, 20)\n",
            "inputs['spanish'].shape: (64, 20)\n",
            "targets.shape: (64, 20)\n"
          ]
        }
      ],
      "source": [
        "for inputs, targets in train_ds.take(1):\n",
        "    print(f\"inputs['english'].shape: {inputs['english'].shape}\")\n",
        "    print(f\"inputs['spanish'].shape: {inputs['spanish'].shape}\")\n",
        "    print(f\"targets.shape: {targets.shape}\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "D1tqIpNxC6Yl"
      },
      "source": [
        "### RNN을 사용한 시퀀스-투-시퀀스 모델"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Idn1gcR-C6Yl"
      },
      "source": [
        "**기반 인코더**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 8,
      "metadata": {
        "id": "RScgqDXWC6Ym"
      },
      "outputs": [],
      "source": [
        "from tensorflow import keras\n",
        "from tensorflow.keras import layers\n",
        "\n",
        "embed_dim = 256\n",
        "latent_dim = 1024\n",
        "\n",
        "source = keras.Input(shape=(None,), dtype=\"int64\", name=\"english\")\n",
        "x = layers.Embedding(vocab_size, embed_dim, mask_zero=True)(source)\n",
        "encoded_source = layers.Bidirectional(\n",
        "    layers.GRU(latent_dim), merge_mode=\"sum\")(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "xq5pLFNhC6Ym"
      },
      "source": [
        "**GRU 기반 디코더와 엔드-투-엔드 모델**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 9,
      "metadata": {
        "id": "-ZprypXUC6Ym"
      },
      "outputs": [],
      "source": [
        "past_target = keras.Input(shape=(None,), dtype=\"int64\", name=\"spanish\")\n",
        "x = layers.Embedding(vocab_size, embed_dim, mask_zero=True)(past_target)\n",
        "decoder_gru = layers.GRU(latent_dim, return_sequences=True)\n",
        "x = decoder_gru(x, initial_state=encoded_source)\n",
        "x = layers.Dropout(0.5)(x)\n",
        "target_next_step = layers.Dense(vocab_size, activation=\"softmax\")(x)\n",
        "seq2seq_rnn = keras.Model([source, past_target], target_next_step)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "D-kMw9qCC6Yn"
      },
      "source": [
        "**RNN 기반 시퀀스-투-시퀀스 모델 훈련하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "n6Yp_XqJC6Yn",
        "outputId": "08c07e02-01d6-48d6-9cc7-5713ea465f37"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/15\n",
            "1302/1302 [==============================] - 76s 46ms/step - loss: 4.6810 - accuracy: 0.3184 - val_loss: 3.8919 - val_accuracy: 0.3858\n",
            "Epoch 2/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 3.7197 - accuracy: 0.4153 - val_loss: 3.2568 - val_accuracy: 0.4645\n",
            "Epoch 3/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 3.2104 - accuracy: 0.4731 - val_loss: 2.8625 - val_accuracy: 0.5171\n",
            "Epoch 4/15\n",
            "1302/1302 [==============================] - 50s 39ms/step - loss: 2.8533 - accuracy: 0.5137 - val_loss: 2.6112 - val_accuracy: 0.5523\n",
            "Epoch 5/15\n",
            "1302/1302 [==============================] - 50s 39ms/step - loss: 2.5785 - accuracy: 0.5475 - val_loss: 2.4338 - val_accuracy: 0.5786\n",
            "Epoch 6/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 2.3565 - accuracy: 0.5748 - val_loss: 2.3013 - val_accuracy: 0.5980\n",
            "Epoch 7/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 2.1711 - accuracy: 0.5989 - val_loss: 2.2046 - val_accuracy: 0.6136\n",
            "Epoch 8/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 2.0149 - accuracy: 0.6195 - val_loss: 2.1325 - val_accuracy: 0.6257\n",
            "Epoch 9/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.8790 - accuracy: 0.6377 - val_loss: 2.0753 - val_accuracy: 0.6345\n",
            "Epoch 10/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.7594 - accuracy: 0.6548 - val_loss: 2.0304 - val_accuracy: 0.6422\n",
            "Epoch 11/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.6549 - accuracy: 0.6693 - val_loss: 1.9980 - val_accuracy: 0.6466\n",
            "Epoch 12/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.5624 - accuracy: 0.6825 - val_loss: 1.9694 - val_accuracy: 0.6514\n",
            "Epoch 13/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.4818 - accuracy: 0.6942 - val_loss: 1.9493 - val_accuracy: 0.6554\n",
            "Epoch 14/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.4104 - accuracy: 0.7043 - val_loss: 1.9303 - val_accuracy: 0.6592\n",
            "Epoch 15/15\n",
            "1302/1302 [==============================] - 51s 39ms/step - loss: 1.3504 - accuracy: 0.7131 - val_loss: 1.9213 - val_accuracy: 0.6596\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<keras.src.callbacks.History at 0x7f6055b5ad40>"
            ]
          },
          "metadata": {},
          "execution_count": 10
        }
      ],
      "source": [
        "seq2seq_rnn.compile(\n",
        "    optimizer=\"rmsprop\",\n",
        "    loss=\"sparse_categorical_crossentropy\",\n",
        "    metrics=[\"accuracy\"])\n",
        "seq2seq_rnn.fit(train_ds, epochs=15, validation_data=val_ds)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NUnTqtHQC6Yn"
      },
      "source": [
        "**RNN 인코더와 디코더로 새로운 문장 번역하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "u1he-U_1C6Yo",
        "outputId": "c9c43d38-4c4e-4a7a-f94d-4bab4615d3f9"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "-\n",
            "Nice timing.\n",
            "1/1 [==============================] - 3s 3s/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] se [UNK] [end]\n",
            "-\n",
            "The old man sat on the bench with his eyes closed.\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] el anciano se sentó en la cama con los ojos cerrados [end]\n",
            "-\n",
            "She must be angry.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 19ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] ella debe estar enojada [end]\n",
            "-\n",
            "I'll be a few minutes late.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] estaré tarde con unos minutos [end]\n",
            "-\n",
            "Which one is mine?\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 24ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] cuál es mío [end]\n",
            "-\n",
            "What you said was very interesting.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "[start] lo que dijiste realmente era muy interesante [end]\n",
            "-\n",
            "I go to bed very early.\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 23ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] me voy a la cama temprano [end]\n",
            "-\n",
            "Tom sensed that Mary was in pain.\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 25ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] tom se [UNK] a mary que estaba en peligro [end]\n",
            "-\n",
            "This knife cuts well.\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] este cuchillo se puso bien [end]\n",
            "-\n",
            "I didn't stop.\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "[start] yo no [UNK] [end]\n",
            "-\n",
            "You both did fine.\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "[start] lo has hecho bien bien [end]\n",
            "-\n",
            "I'm really not good at that kind of thing.\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] no soy tan buena como esa buena palabra [end]\n",
            "-\n",
            "I got acquainted with him in France.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 19ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] me lo encontré con él en francia [end]\n",
            "-\n",
            "Yes, of course.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] ya no lo sabe [end]\n",
            "-\n",
            "He asked about the others.\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "[start] Él pidió de los otros [end]\n",
            "-\n",
            "Tom doesn't have a date for the Saturday night.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 23ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 24ms/step\n",
            "[start] tom no tiene una cita para la noche de noche [end]\n",
            "-\n",
            "Do you think that eating breakfast every day is important?\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 24ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "[start] crees que esa es el día de los días [end]\n",
            "-\n",
            "We don't let people like you in here.\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 23ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] no te [UNK] a la gente que nos está aquí [end]\n",
            "-\n",
            "You know you need it.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 23ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "[start] sabes que lo necesitas [end]\n",
            "-\n",
            "The wall had a crack in it.\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 21ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 20ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "1/1 [==============================] - 0s 22ms/step\n",
            "[start] la [UNK] se [UNK] en un [UNK] [end]\n"
          ]
        }
      ],
      "source": [
        "import numpy as np\n",
        "spa_vocab = target_vectorization.get_vocabulary()\n",
        "spa_index_lookup = dict(zip(range(len(spa_vocab)), spa_vocab))\n",
        "max_decoded_sentence_length = 20\n",
        "\n",
        "def decode_sequence(input_sentence):\n",
        "    tokenized_input_sentence = source_vectorization([input_sentence])\n",
        "    decoded_sentence = \"[start]\"\n",
        "    for i in range(max_decoded_sentence_length):\n",
        "        tokenized_target_sentence = target_vectorization([decoded_sentence])\n",
        "        next_token_predictions = seq2seq_rnn.predict(\n",
        "            [tokenized_input_sentence, tokenized_target_sentence])\n",
        "        sampled_token_index = np.argmax(next_token_predictions[0, i, :])\n",
        "        sampled_token = spa_index_lookup[sampled_token_index]\n",
        "        decoded_sentence += \" \" + sampled_token\n",
        "        if sampled_token == \"[end]\":\n",
        "            break\n",
        "    return decoded_sentence\n",
        "\n",
        "test_eng_texts = [pair[0] for pair in test_pairs]\n",
        "for _ in range(20):\n",
        "    input_sentence = random.choice(test_eng_texts)\n",
        "    print(\"-\")\n",
        "    print(input_sentence)\n",
        "    print(decode_sequence(input_sentence))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ThpTHoiqC6Yo"
      },
      "source": [
        "### 트랜스포머를 사용한 시퀀스-투-시퀀스 모델"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "pdRs7iLLC6Yo"
      },
      "source": [
        "#### 트랜스포머 디코더"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3bgMELmgC6Yo"
      },
      "source": [
        "**`TransformerDecoder` 클래스**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {
        "id": "NMlKqNtqC6Yo"
      },
      "outputs": [],
      "source": [
        "class TransformerDecoder(layers.Layer):\n",
        "    def __init__(self, embed_dim, dense_dim, num_heads, **kwargs):\n",
        "        super().__init__(**kwargs)\n",
        "        self.embed_dim = embed_dim\n",
        "        self.dense_dim = dense_dim\n",
        "        self.num_heads = num_heads\n",
        "        self.attention_1 = layers.MultiHeadAttention(\n",
        "            num_heads=num_heads, key_dim=embed_dim)\n",
        "        self.attention_2 = layers.MultiHeadAttention(\n",
        "            num_heads=num_heads, key_dim=embed_dim)\n",
        "        self.dense_proj = keras.Sequential(\n",
        "            [layers.Dense(dense_dim, activation=\"relu\"),\n",
        "             layers.Dense(embed_dim),]\n",
        "        )\n",
        "        self.layernorm_1 = layers.LayerNormalization()\n",
        "        self.layernorm_2 = layers.LayerNormalization()\n",
        "        self.layernorm_3 = layers.LayerNormalization()\n",
        "        self.supports_masking = True\n",
        "\n",
        "    def get_config(self):\n",
        "        config = super().get_config()\n",
        "        config.update({\n",
        "            \"embed_dim\": self.embed_dim,\n",
        "            \"num_heads\": self.num_heads,\n",
        "            \"dense_dim\": self.dense_dim,\n",
        "        })\n",
        "        return config\n",
        "\n",
        "    def call(self, inputs, encoder_outputs, mask=None):\n",
        "        attention_output_1 = self.attention_1(\n",
        "            query=inputs,\n",
        "            value=inputs,\n",
        "            key=inputs,\n",
        "            use_causal_mask=True)\n",
        "        attention_output_1 = self.layernorm_1(inputs + attention_output_1)\n",
        "        attention_output_2 = self.attention_2(\n",
        "            query=attention_output_1,\n",
        "            value=encoder_outputs,\n",
        "            key=encoder_outputs\n",
        "        )\n",
        "        attention_output_2 = self.layernorm_2(\n",
        "            attention_output_1 + attention_output_2)\n",
        "        proj_output = self.dense_proj(attention_output_2)\n",
        "        return self.layernorm_3(attention_output_2 + proj_output)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "yF6EcciHC6Yp"
      },
      "source": [
        "#### 기계 번역을 위한 트랜스포머"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_sOybVcwC6Yp"
      },
      "source": [
        "**`PositionalEmbedding` 층**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {
        "id": "L5e12udRC6Yp"
      },
      "outputs": [],
      "source": [
        "class PositionalEmbedding(layers.Layer):\n",
        "    def __init__(self, sequence_length, input_dim, output_dim, **kwargs):\n",
        "        super().__init__(**kwargs)\n",
        "        self.token_embeddings = layers.Embedding(\n",
        "            input_dim=input_dim, output_dim=output_dim)\n",
        "        self.position_embeddings = layers.Embedding(\n",
        "            input_dim=sequence_length, output_dim=output_dim)\n",
        "        self.sequence_length = sequence_length\n",
        "        self.input_dim = input_dim\n",
        "        self.output_dim = output_dim\n",
        "\n",
        "    def call(self, inputs):\n",
        "        length = tf.shape(inputs)[-1]\n",
        "        positions = tf.range(start=0, limit=length, delta=1)\n",
        "        embedded_tokens = self.token_embeddings(inputs)\n",
        "        embedded_positions = self.position_embeddings(positions)\n",
        "        return embedded_tokens + embedded_positions\n",
        "\n",
        "    def compute_mask(self, inputs, mask=None):\n",
        "        return tf.math.not_equal(inputs, 0)\n",
        "\n",
        "    def get_config(self):\n",
        "        config = super(PositionalEmbedding, self).get_config()\n",
        "        config.update({\n",
        "            \"output_dim\": self.output_dim,\n",
        "            \"sequence_length\": self.sequence_length,\n",
        "            \"input_dim\": self.input_dim,\n",
        "        })\n",
        "        return config"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1H7QlHwQC6Yp"
      },
      "source": [
        "**엔드-투-엔드 트랜스포머**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 14,
      "metadata": {
        "id": "vIM5VwfzN4Co"
      },
      "outputs": [],
      "source": [
        "class TransformerEncoder(layers.Layer):\n",
        "    def __init__(self, embed_dim, dense_dim, num_heads, **kwargs):\n",
        "        super().__init__(**kwargs)\n",
        "        self.embed_dim = embed_dim\n",
        "        self.dense_dim = dense_dim\n",
        "        self.num_heads = num_heads\n",
        "        self.attention = layers.MultiHeadAttention(\n",
        "            num_heads=num_heads, key_dim=embed_dim)\n",
        "        self.dense_proj = keras.Sequential(\n",
        "            [layers.Dense(dense_dim, activation=\"relu\"),\n",
        "             layers.Dense(embed_dim),]\n",
        "        )\n",
        "        self.layernorm_1 = layers.LayerNormalization()\n",
        "        self.layernorm_2 = layers.LayerNormalization()\n",
        "\n",
        "    def call(self, inputs, mask=None):\n",
        "        if mask is not None:\n",
        "            mask = mask[:, tf.newaxis, :]\n",
        "        attention_output = self.attention(\n",
        "            inputs, inputs, attention_mask=mask)\n",
        "        proj_input = self.layernorm_1(inputs + attention_output)\n",
        "        proj_output = self.dense_proj(proj_input)\n",
        "        return self.layernorm_2(proj_input + proj_output)\n",
        "\n",
        "    def get_config(self):\n",
        "        config = super().get_config()\n",
        "        config.update({\n",
        "            \"embed_dim\": self.embed_dim,\n",
        "            \"num_heads\": self.num_heads,\n",
        "            \"dense_dim\": self.dense_dim,\n",
        "        })\n",
        "        return config"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 15,
      "metadata": {
        "id": "M-s0ZCSQC6Yp"
      },
      "outputs": [],
      "source": [
        "embed_dim = 256\n",
        "dense_dim = 2048\n",
        "num_heads = 8\n",
        "\n",
        "encoder_inputs = keras.Input(shape=(None,), dtype=\"int64\", name=\"english\")\n",
        "x = PositionalEmbedding(sequence_length, vocab_size, embed_dim)(encoder_inputs)\n",
        "encoder_outputs = TransformerEncoder(embed_dim, dense_dim, num_heads)(x)\n",
        "\n",
        "decoder_inputs = keras.Input(shape=(None,), dtype=\"int64\", name=\"spanish\")\n",
        "x = PositionalEmbedding(sequence_length, vocab_size, embed_dim)(decoder_inputs)\n",
        "x = TransformerDecoder(embed_dim, dense_dim, num_heads)(x, encoder_outputs)\n",
        "x = layers.Dropout(0.5)(x)\n",
        "decoder_outputs = layers.Dense(vocab_size, activation=\"softmax\")(x)\n",
        "transformer = keras.Model([encoder_inputs, decoder_inputs], decoder_outputs)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "IstNU4guC6Yp"
      },
      "source": [
        "**시퀀스-투-시퀀스 트랜스포머 훈련하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 16,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ox_6aZfxC6Yq",
        "outputId": "cb832822-b483-486a-fa59-153f3476b39f"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/30\n",
            "1302/1302 [==============================] - 57s 39ms/step - loss: 3.9515 - accuracy: 0.4188 - val_loss: 2.7509 - val_accuracy: 0.5770\n",
            "Epoch 2/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 2.5898 - accuracy: 0.6019 - val_loss: 2.2324 - val_accuracy: 0.6471\n",
            "Epoch 3/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 2.2178 - accuracy: 0.6515 - val_loss: 2.0779 - val_accuracy: 0.6707\n",
            "Epoch 4/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 2.0307 - accuracy: 0.6788 - val_loss: 2.0129 - val_accuracy: 0.6806\n",
            "Epoch 5/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.9215 - accuracy: 0.6963 - val_loss: 1.9987 - val_accuracy: 0.6851\n",
            "Epoch 6/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.8491 - accuracy: 0.7085 - val_loss: 1.9938 - val_accuracy: 0.6901\n",
            "Epoch 7/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.8019 - accuracy: 0.7176 - val_loss: 2.0158 - val_accuracy: 0.6868\n",
            "Epoch 8/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.7651 - accuracy: 0.7252 - val_loss: 2.0363 - val_accuracy: 0.6910\n",
            "Epoch 9/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.7325 - accuracy: 0.7315 - val_loss: 2.0417 - val_accuracy: 0.6921\n",
            "Epoch 10/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.7054 - accuracy: 0.7369 - val_loss: 2.0829 - val_accuracy: 0.6892\n",
            "Epoch 11/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.6810 - accuracy: 0.7419 - val_loss: 2.0481 - val_accuracy: 0.6955\n",
            "Epoch 12/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.6540 - accuracy: 0.7475 - val_loss: 2.0796 - val_accuracy: 0.6926\n",
            "Epoch 13/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.6327 - accuracy: 0.7515 - val_loss: 2.1090 - val_accuracy: 0.6929\n",
            "Epoch 14/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.6138 - accuracy: 0.7551 - val_loss: 2.1333 - val_accuracy: 0.6959\n",
            "Epoch 15/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.5929 - accuracy: 0.7588 - val_loss: 2.1477 - val_accuracy: 0.6907\n",
            "Epoch 16/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.5730 - accuracy: 0.7628 - val_loss: 2.1654 - val_accuracy: 0.6905\n",
            "Epoch 17/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.5561 - accuracy: 0.7665 - val_loss: 2.1738 - val_accuracy: 0.6930\n",
            "Epoch 18/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.5415 - accuracy: 0.7693 - val_loss: 2.1982 - val_accuracy: 0.6954\n",
            "Epoch 19/30\n",
            "1302/1302 [==============================] - 43s 33ms/step - loss: 1.5250 - accuracy: 0.7724 - val_loss: 2.2561 - val_accuracy: 0.6918\n",
            "Epoch 20/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.5086 - accuracy: 0.7753 - val_loss: 2.2229 - val_accuracy: 0.6932\n",
            "Epoch 21/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4935 - accuracy: 0.7777 - val_loss: 2.2340 - val_accuracy: 0.6957\n",
            "Epoch 22/30\n",
            "1302/1302 [==============================] - 42s 32ms/step - loss: 1.4799 - accuracy: 0.7804 - val_loss: 2.2960 - val_accuracy: 0.6892\n",
            "Epoch 23/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4679 - accuracy: 0.7827 - val_loss: 2.2763 - val_accuracy: 0.6947\n",
            "Epoch 24/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4544 - accuracy: 0.7854 - val_loss: 2.3117 - val_accuracy: 0.6914\n",
            "Epoch 25/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4406 - accuracy: 0.7879 - val_loss: 2.3158 - val_accuracy: 0.6948\n",
            "Epoch 26/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4293 - accuracy: 0.7901 - val_loss: 2.3182 - val_accuracy: 0.6957\n",
            "Epoch 27/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4177 - accuracy: 0.7917 - val_loss: 2.3373 - val_accuracy: 0.6951\n",
            "Epoch 28/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4110 - accuracy: 0.7933 - val_loss: 2.3765 - val_accuracy: 0.6953\n",
            "Epoch 29/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.4010 - accuracy: 0.7955 - val_loss: 2.3944 - val_accuracy: 0.6930\n",
            "Epoch 30/30\n",
            "1302/1302 [==============================] - 42s 33ms/step - loss: 1.3931 - accuracy: 0.7969 - val_loss: 2.3732 - val_accuracy: 0.6939\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<keras.src.callbacks.History at 0x7f5f8f7e4c40>"
            ]
          },
          "metadata": {},
          "execution_count": 16
        }
      ],
      "source": [
        "transformer.compile(\n",
        "    optimizer=\"rmsprop\",\n",
        "    loss=\"sparse_categorical_crossentropy\",\n",
        "    metrics=[\"accuracy\"])\n",
        "transformer.fit(train_ds, epochs=30, validation_data=val_ds)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-rE4EM5PC6Yq"
      },
      "source": [
        "**트랜스포머 모델을 사용해 새로운 문장 번역하기**"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 17,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "gQ3O0QCLC6Yq",
        "outputId": "94025507-4b41-4936-975c-13a4d60d449d"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "-\n",
            "The pain was almost unbearable.\n",
            "[start] el dolor era [UNK] casi rica [end]\n",
            "-\n",
            "I have to go to the toilet.\n",
            "[start] tengo que ir al baño [end]\n",
            "-\n",
            "He looked around the waiting room.\n",
            "[start] Él miró a su alrededor de la sala [end]\n",
            "-\n",
            "Tom pulled Mary out of the burning car.\n",
            "[start] tom sacó a mary el coche se está en francia [end]\n",
            "-\n",
            "They had wonderful lives.\n",
            "[start] ellos lo hicieron vive [end]\n",
            "-\n",
            "Keep in mind that smoking is not good for your health.\n",
            "[start] sigue usted de fumar es una buena para la salud [end]\n",
            "-\n",
            "I'm three years older than my brother.\n",
            "[start] yo soy tres años mayor que mi hermano [end]\n",
            "-\n",
            "Come home with me.\n",
            "[start] ven a mi casa conmigo [end]\n",
            "-\n",
            "Strange things happened on her birthday.\n",
            "[start] tan raro pasó cosas en su cumpleaños [end]\n",
            "-\n",
            "Two against one is not a fair fight.\n",
            "[start] dos es uno de no es solo una lloviendo [end]\n",
            "-\n",
            "Will you have some more pie?\n",
            "[start] te [UNK] algo más de torta [end]\n",
            "-\n",
            "Put some clothes on.\n",
            "[start] [UNK] algo de ropa [end]\n",
            "-\n",
            "I think I still have time for another cup of coffee.\n",
            "[start] creo que todavía tengo tiempo para otra taza de café [end]\n",
            "-\n",
            "I tried to get Tom to help Mary.\n",
            "[start] traté de recibir a tom de ayuda a mary [end]\n",
            "-\n",
            "These socks don't stretch when you wash them.\n",
            "[start] estos calcetines no te [UNK] cuándo [end]\n",
            "-\n",
            "This isn't my first time to be here.\n",
            "[start] esto no es mi primera vez en que esté aquí [end]\n",
            "-\n",
            "Do you want my help?\n",
            "[start] quieres mi ayuda [end]\n",
            "-\n",
            "I don't think I could spend six hours alone with you.\n",
            "[start] no creo que podría pasar seis horas solo contigo [end]\n",
            "-\n",
            "Can we really learn to speak a foreign language like a native?\n",
            "[start] nos podemos aprender a hablar un idioma extranjero como un idioma extranjero [end]\n",
            "-\n",
            "I'm not a hero.\n",
            "[start] no soy un héroe [end]\n"
          ]
        }
      ],
      "source": [
        "import numpy as np\n",
        "spa_vocab = target_vectorization.get_vocabulary()\n",
        "spa_index_lookup = dict(zip(range(len(spa_vocab)), spa_vocab))\n",
        "max_decoded_sentence_length = 20\n",
        "\n",
        "def decode_sequence(input_sentence):\n",
        "    tokenized_input_sentence = source_vectorization([input_sentence])\n",
        "    decoded_sentence = \"[start]\"\n",
        "    for i in range(max_decoded_sentence_length):\n",
        "        tokenized_target_sentence = target_vectorization(\n",
        "            [decoded_sentence])[:, :-1]\n",
        "        predictions = transformer(\n",
        "            [tokenized_input_sentence, tokenized_target_sentence])\n",
        "        sampled_token_index = np.argmax(predictions[0, i, :])\n",
        "        sampled_token = spa_index_lookup[sampled_token_index]\n",
        "        decoded_sentence += \" \" + sampled_token\n",
        "        if sampled_token == \"[end]\":\n",
        "            break\n",
        "    return decoded_sentence\n",
        "\n",
        "test_eng_texts = [pair[0] for pair in test_pairs]\n",
        "for _ in range(20):\n",
        "    input_sentence = random.choice(test_eng_texts)\n",
        "    print(\"-\")\n",
        "    print(input_sentence)\n",
        "    print(decode_sequence(input_sentence))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "gz8VRatlC6Yq"
      },
      "source": [
        "## 요약"
      ]
    }
  ],
  "metadata": {
    "accelerator": "GPU",
    "colab": {
      "name": "chapter11_part04_sequence-to-sequence-learning.i",
      "provenance": [],
      "machine_shape": "hm",
      "gpuType": "V100",
      "gpuClass": "premium"
    },
    "kernelspec": {
      "display_name": "default:Python",
      "language": "python",
      "name": "conda-env-default-py"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.9.10"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}