{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "This is a companion notebook for the book [Deep Learning with Python, Second Edition](https://www.manning.com/books/deep-learning-with-python-second-edition?a_aid=keras&a_bid=76564dff). For readability, it only contains runnable code blocks and section titles, and omits everything else in the book: text paragraphs, figures, and pseudocode.\n\n**If you want to be able to follow what's going on, I recommend reading the notebook side by side with your copy of the book.**\n\nThis notebook was generated for TensorFlow 2.6."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "## The Transformer architecture"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "### Understanding self-attention"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "#### Generalized self-attention: the query-key-value model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "### Multi-Head attention"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "### The Transformer encoder"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Getting the data**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "!curl -O https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz\n",
    "!tar -xf aclImdb_v1.tar.gz\n",
    "!rm -r aclImdb/train/unsup"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Preparing the data**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "import os, pathlib, shutil, random\n",
    "from tensorflow import keras\n",
    "batch_size = 32\n",
    "base_dir = pathlib.Path(\"aclImdb\")\n",
    "val_dir = base_dir / \"val\"\n",
    "train_dir = base_dir / \"train\"\n",
    "for category in (\"neg\", \"pos\"):\n",
    "    os.makedirs(val_dir / category)\n",
    "    files = os.listdir(train_dir / category)\n",
    "    random.Random(1337).shuffle(files)\n",
    "    num_val_samples = int(0.2 * len(files))\n",
    "    val_files = files[-num_val_samples:]\n",
    "    for fname in val_files:\n",
    "        shutil.move(train_dir / category / fname,\n",
    "                    val_dir / category / fname)\n",
    "\n",
    "train_ds = keras.preprocessing.text_dataset_from_directory(\n",
    "    \"aclImdb/train\", batch_size=batch_size\n",
    ")\n",
    "val_ds = keras.preprocessing.text_dataset_from_directory(\n",
    "    \"aclImdb/val\", batch_size=batch_size\n",
    ")\n",
    "test_ds = keras.preprocessing.text_dataset_from_directory(\n",
    "    \"aclImdb/test\", batch_size=batch_size\n",
    ")\n",
    "text_only_train_ds = train_ds.map(lambda x, y: x)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Vectorizing the data**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "from tensorflow.keras.layers.experimental.preprocessing import TextVectorization\n",
    "max_length = 600\n",
    "max_tokens = 20000\n",
    "text_vectorization = TextVectorization(\n",
    "    max_tokens=max_tokens,\n",
    "    output_mode=\"int\",\n",
    "    output_sequence_length=max_length,\n",
    ")\n",
    "text_vectorization.adapt(text_only_train_ds)\n",
    "\n",
    "int_train_ds = train_ds.map(lambda x, y: (text_vectorization(x), y))\n",
    "int_val_ds = val_ds.map(lambda x, y: (text_vectorization(x), y))\n",
    "int_test_ds = test_ds.map(lambda x, y: (text_vectorization(x), y))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Transformer encoder implemented as a subclassed Layer**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "from tensorflow import keras\n",
    "from tensorflow.keras import layers\n",
    "\n",
    "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(TransformerEncoder, self).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": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Text classification model that combines the Transformer encoder and a pooling layer**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "vocab_size = 20000\n",
    "embed_dim = 256\n",
    "num_heads = 2\n",
    "dense_dim = 32\n",
    "\n",
    "inputs = keras.Input(shape=(None,), dtype=\"int64\")\n",
    "x = layers.Embedding(vocab_size, embed_dim)(inputs)\n",
    "x = TransformerEncoder(embed_dim, dense_dim, num_heads)(x)\n",
    "x = layers.GlobalMaxPooling1D()(x)\n",
    "x = layers.Dropout(0.5)(x)\n",
    "outputs = layers.Dense(1, activation=\"sigmoid\")(x)\n",
    "model = keras.Model(inputs, outputs)\n",
    "model.compile(optimizer=\"rmsprop\",\n",
    "              loss=\"binary_crossentropy\",\n",
    "              metrics=[\"accuracy\"])\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Training and evaluating the Transformer encoder based model**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "callbacks = [\n",
    "    keras.callbacks.ModelCheckpoint(\"transformer_encoder.keras\",\n",
    "                                    save_best_only=True)\n",
    "]\n",
    "model.fit(int_train_ds, validation_data=int_val_ds, epochs=20, callbacks=callbacks)\n",
    "model = keras.models.load_model(\n",
    "    \"transformer_encoder.keras\",\n",
    "    custom_objects={\"TransformerEncoder\": TransformerEncoder})\n",
    "print(f\"Test acc: {model.evaluate(int_test_ds)[1]:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "#### Using positional encoding to reinject order information"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Implementing positional embedding as a subclassed layer**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "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": {
    "colab_type": "text"
   },
   "source": [
    "#### Putting it all together: a text-classification Transformer"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "**Text classification model that combines positional embedding, the Transformer encoder, and a pooling layer**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab_type": "code"
   },
   "outputs": [],
   "source": [
    "vocab_size = 20000\n",
    "sequence_length = 600\n",
    "embed_dim = 256\n",
    "num_heads = 2\n",
    "dense_dim = 32\n",
    "\n",
    "inputs = keras.Input(shape=(None,), dtype=\"int64\")\n",
    "x = PositionalEmbedding(sequence_length, vocab_size, embed_dim)(inputs)\n",
    "x = TransformerEncoder(embed_dim, dense_dim, num_heads)(x)\n",
    "x = layers.GlobalMaxPooling1D()(x)\n",
    "x = layers.Dropout(0.5)(x)\n",
    "outputs = layers.Dense(1, activation=\"sigmoid\")(x)\n",
    "model = keras.Model(inputs, outputs)\n",
    "model.compile(optimizer=\"rmsprop\",\n",
    "              loss=\"binary_crossentropy\",\n",
    "              metrics=[\"accuracy\"])\n",
    "model.summary()\n",
    "\n",
    "callbacks = [\n",
    "    keras.callbacks.ModelCheckpoint(\"full_transformer_encoder.keras\",\n",
    "                                    save_best_only=True)\n",
    "]\n",
    "model.fit(int_train_ds, validation_data=int_val_ds, epochs=20, callbacks=callbacks)\n",
    "model = keras.models.load_model(\n",
    "    \"full_transformer_encoder.keras\",\n",
    "    custom_objects={\"TransformerEncoder\": TransformerEncoder,\n",
    "                    \"PositionalEmbedding\": PositionalEmbedding})\n",
    "print(f\"Test acc: {model.evaluate(int_test_ds)[1]:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text"
   },
   "source": [
    "### When to use sequence models over bag-of-words models?"
   ]
  }
 ],
 "metadata": {
  "colab": {
   "collapsed_sections": [],
   "name": "chapter11_part03_transformer.i",
   "private_outputs": false,
   "provenance": [],
   "toc_visible": true
  },
  "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": "3.7.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}