Traducao_com_Atencao.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KdkvAZYXjbxf"
      },
      "source": [
        "# Tradução utilizando Attention\n",
        "> Baseado no artigo [Attention is all you need](https://arxiv.org/abs/1706.03762)."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hNkD-nvL6UCe"
      },
      "source": [
        "O Transformador - para facilitar o entendimento daqueles com um conhecimento prévio em redes neurais, chamaremos de \"Transformer\" assim como na literatura original - do artigo \"Attention is all you need\" têm estado na mente de várias pessoas desde que foi escrito no ano de 2017. Além de produzir grandes melhorias na qualidade da tradução, fornece uma arquitetura para muitas outras tarefas de NLP ( Natural Language Processing ). O artigo em si está escrito de forma muito clara, mas os leitores comentam que é muito difícil de implementá-lo corretamente.\n",
        "\n",
        "Nesse notebook apresentamos uma implementação, tentando comentar linha por linha. Reorganizamos e excluímos algumas seções do artigo original e adicionamos comentários onde pensamos ser necessário.\n",
        "\n",
        "Vale mencionar que, no original, existem cerca 400 linhas de código de biblioteca que podem processar 27.000 tokens por segundo em 4 GPUs, mas como não contamos com esse poder computacional, implementamos uma versão que funcione com apenas 1 GPU, caso quem for rodar esse código possua uma, do contrário será usado a CPU.\n",
        "\n",
        "O código apresentado aqui é fortemente baseado nos pacotes [OpenNMT](https://opennmt.net/). Para outras implementações de serviço completas do modelo, vale conferir [Tensor2Tensor (Tensorflow)](https://github.com/tensorflow/tensor2tensor) e [Sockeye (mxnet)](https://github.com/awslabs/sockeye)."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Dyt8aemViaIk"
      },
      "source": [
        "## Bibliotecas"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "id": "HUhC_6ObibG9"
      },
      "outputs": [
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "[nltk_data] Downloading package punkt to\n",
            "[nltk_data]     C:\\Users\\User\\AppData\\Roaming\\nltk_data...\n",
            "[nltk_data]   Package punkt is already up-to-date!\n"
          ]
        }
      ],
      "source": [
        "import numpy as np\n",
        "\n",
        "import torch\n",
        "import torch.nn as nn\n",
        "import torch.nn.functional as F\n",
        "from torch.autograd import Variable\n",
        "\n",
        "import nltk\n",
        "from nltk.translate.bleu_score import sentence_bleu\n",
        "nltk.download('punkt')\n",
        "\n",
        "import math, copy, time\n",
        "\n",
        "import matplotlib.pyplot as plt\n",
        "import seaborn\n",
        "seaborn.set_context(context = \"talk\")\n",
        "\n",
        "from IPython.display import Image\n",
        "\n",
        "from statistics import mean\n",
        "\n",
        "%matplotlib inline"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "odf5cWZj-5f0"
      },
      "source": [
        "## Breve background\n",
        "O objetivo de reduzir a computação sequencial também forma a base do [Extended Neural GPU](https://www.researchgate.net/publication/309631810_Extensions_and_Limitations_of_the_Neural_GPU), [ByteNet](https://sh-tsang.medium.com/review-neural-machine-translation-in-linear-time-bytenet-e5a50bfb462b) e [ConvS2S](https://sh-tsang.medium.com/review-convolutional-sequence-to-sequence-learning-convs2s-510a9eddce05), que usam redes neurais convolucionais como bloco de construção básico, computando representações ocultas em paralelo para todas as posições de entrada e saída. Nesses modelos, o número de operações necessárias para relacionar sinais de duas posições arbitrárias de entrada ou saída cresce na distância entre as posições, linearmente para ConvS2S e logaritmicamente para ByteNet. Isso torna mais difícil aprender dependências entre posições distantes. No Transformer, isso é reduzido a um número constante de operações, embora ao custo de uma resolução efetiva reduzida devido à média das posições ponderadas de atenção, um efeito que é neutralizado com a Atenção de Múltiplas Cabeças ( Multi-Head ).\n",
        "\n",
        "A autoatenção, às vezes chamada de intra-atenção, é um mecanismo de atenção que relaciona diferentes posições de uma única sequência para computar uma representação da sequência. A autoatenção tem sido usada com sucesso em uma variedade de tarefas, incluindo compreensão de leitura, resumo de abstratos, vinculação textual e representações de sentenças independentes da tarefa de aprendizagem. As redes de memória de ponta a ponta são baseadas em um mecanismo de atenção recorrente em vez de recorrência alinhada por sequência e têm demonstrado um bom desempenho em respostas a perguntas de linguagem simples e tarefas de modelagem de linguagem.\n",
        "\n",
        "Até onde sabemos, no entanto, o Transformer é o primeiro modelo de tradução que depende inteiramente da autoatenção para calcular as representações de sua entrada e saída sem usar redes neurais recorrentes alinhadas à sequência ou convolução.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "T4Jq5J98jR-B"
      },
      "source": [
        "## Arquitetura do modelo\n",
        "A maioria dos modelos neurais competitivos de tradução de sequência tem uma estrutura [codificador-decodificador (Encoder-Decoder)](https://arxiv.org/abs/1409.0473). Aqui, o codificador mapeia uma sequência de entrada de representações de símbolos $ (x_1, ..., x_n) $ para uma sequência de representações contínuas $ z = (z_1, ..., z_n) $. Dado $ z $, o decodificador então gera uma sequência de saída $ (y_1, ..., y_m) $ de símbolos um elemento por vez. A cada passo, o modelo é [auto-regressivo](https://arxiv.org/abs/1308.0850), consumindo os símbolos gerados anteriormente como entrada adicional quando gerando o próximo."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "qw6PxsbPigcN"
      },
      "outputs": [],
      "source": [
        "class EncoderDecoder(nn.Module):\n",
        "    '''\n",
        "    A standard Encoder-Decoder architecture. \n",
        "    Base for this and many other models.\n",
        "    '''\n",
        "    def __init__(self, encoder, decoder, src_embed, tgt_embed, generator):\n",
        "        super(EncoderDecoder, self).__init__()\n",
        "        self.encoder = encoder\n",
        "        self.decoder = decoder\n",
        "        self.src_embed = src_embed\n",
        "        self.tgt_embed = tgt_embed\n",
        "        self.generator = generator\n",
        "        \n",
        "    def forward(self, src, tgt, src_mask, tgt_mask):\n",
        "        '''\n",
        "        Take in and process masked src and target sequences.\n",
        "        '''\n",
        "        return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask)\n",
        "    \n",
        "    def encode(self, src, src_mask):\n",
        "        return self.encoder(self.src_embed(src), src_mask)\n",
        "    \n",
        "    def decode(self, memory, src_mask, tgt, tgt_mask):\n",
        "        return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 3,
      "metadata": {
        "id": "fKtApiXLjTh9"
      },
      "outputs": [],
      "source": [
        "class Generator(nn.Module):\n",
        "    '''\n",
        "    Define standard linear + softmax generation step.\n",
        "    '''\n",
        "    def __init__(self, d_model, vocab):\n",
        "        super(Generator, self).__init__()\n",
        "        self.proj = nn.Linear(d_model, vocab)\n",
        "\n",
        "    def forward(self, x):\n",
        "        return F.log_softmax(self.proj(x), dim = -1)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Ly5DZF5NA9F8"
      },
      "source": [
        "O Transformer segue essa arquitetura geral usando camadas de autoatenção empilhadas e totalmente conectadas por pontos para o codificador e o decodificador, mostradas nas metades esquerda e direita da figura abaixo, respectivamente."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 577
        },
        "id": "qIqEZLPW1Z5z",
        "outputId": "aa4a7409-0787-4cb3-8eff-67275d518c07"
      },
      "outputs": [
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<IPython.core.display.Image object>"
            ]
          },
          "execution_count": 4,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "Image(filename='Imagens/architecture.png')"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "0rON22Oykxe-"
      },
      "source": [
        "## Pilhas de Encoder e Decoder"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "oM97bTuNk0fY"
      },
      "source": [
        "### Encoder\n",
        "O enconder é composto de uma pilha de $ N = 6 $ camadas idênticas."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 5,
      "metadata": {
        "id": "W4YB926Fkz2U"
      },
      "outputs": [],
      "source": [
        "def clones(module, N):\n",
        "    '''\n",
        "    Produce N identical layers.\n",
        "    '''\n",
        "    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 6,
      "metadata": {
        "id": "EcqcEC2mk8iV"
      },
      "outputs": [],
      "source": [
        "class Encoder(nn.Module):\n",
        "    '''\n",
        "    Core encoder is a stack of N layers\n",
        "    '''\n",
        "    def __init__(self, layer, N):\n",
        "        super(Encoder, self).__init__()\n",
        "        self.layers = clones(layer, N)\n",
        "        self.norm = LayerNorm(layer.size)\n",
        "        \n",
        "    def forward(self, x, mask):\n",
        "        '''\n",
        "        Pass the input (and mask) through each layer in turn.\n",
        "        '''\n",
        "        for layer in self.layers:\n",
        "            x = layer(x, mask)\n",
        "            \n",
        "        return self.norm(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "zXSbUBcMBX_Y"
      },
      "source": [
        "Empregamos uma [conexão residual](https://arxiv.org/abs/1512.03385) em torno de cada uma das duas subcamadas, seguida da [normalização de camada](https://arxiv.org/abs/1607.06450)."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 7,
      "metadata": {
        "id": "EJj1wBkck_i0"
      },
      "outputs": [],
      "source": [
        "class LayerNorm(nn.Module):\n",
        "    '''\n",
        "    Construct a layernorm module\n",
        "    '''\n",
        "    def __init__(self, features, eps = 1e-6):\n",
        "        super(LayerNorm, self).__init__()\n",
        "        self.a_2 = nn.Parameter(torch.ones(features))\n",
        "        self.b_2 = nn.Parameter(torch.zeros(features))\n",
        "        self.eps = eps\n",
        "\n",
        "    def forward(self, x):\n",
        "        mean = x.mean(-1, keepdim = True)\n",
        "        std = x.std(-1, keepdim = True)\n",
        "        \n",
        "        return self.a_2 * (x - mean)/ (std + self.eps) + self.b_2"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hKwqxTZCCqjv"
      },
      "source": [
        "Isto é, a saída de cada subcamada é $ LayerNorm(x + Sublayer(x)) $, em que $ Sublayer(x) $ é a função implementada na própria subcamada. Aplicamos [dropout](https://jmlr.org/papers/v15/srivastava14a.html) na saída de cada subcamada, antes de ser adicionada à entrada da subcamada e normalizada.\n",
        "\n",
        "Para facilitar essas conexões residuais, todas as subcamadas do modelo, tanto quanto as camadas para os embedding, produzem saídas de dimensão $ d_{model} = 512 $."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 8,
      "metadata": {
        "id": "NtuALmSilDfd"
      },
      "outputs": [],
      "source": [
        "class SublayerConnection(nn.Module):\n",
        "    '''\n",
        "    A residual connection followed by a layer norm.\n",
        "    Note for code simplicity the norm is first as opposed to last.\n",
        "    '''\n",
        "    def __init__(self, size, dropout):\n",
        "        super(SublayerConnection, self).__init__()\n",
        "        self.norm = LayerNorm(size)\n",
        "        self.dropout = nn.Dropout(dropout)\n",
        "\n",
        "    def forward(self, x, sublayer):\n",
        "        '''\n",
        "        Apply residual connection to any sublayer with the same size.\n",
        "        '''\n",
        "        return x + self.dropout(sublayer(self.norm(x)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "1POjfYUlDfiB"
      },
      "source": [
        "Cada camada tem duas subcamadas. A primeira é um mecanismo de autoatenção de múltiplas cabeças, e a segunda é uma rede feed-forward simples e totalmente conectada em termos de posição.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 9,
      "metadata": {
        "id": "7R97BL5Wlr11"
      },
      "outputs": [],
      "source": [
        "class EncoderLayer(nn.Module):\n",
        "    '''\n",
        "    Encoder is made up of self-attn and feed forward (defined below)\n",
        "    '''\n",
        "    def __init__(self, size, self_attn, feed_forward, dropout):\n",
        "        super(EncoderLayer, self).__init__()\n",
        "        self.self_attn = self_attn\n",
        "        self.feed_forward = feed_forward\n",
        "        self.sublayer = clones(SublayerConnection(size, dropout), 2)\n",
        "        self.size = size\n",
        "\n",
        "    def forward(self, x, mask):\n",
        "        '''\n",
        "        Follow figure above for connections\n",
        "        '''\n",
        "        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))\n",
        "        \n",
        "        return self.sublayer[1](x, self.feed_forward)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QGAEh2GPlxA4"
      },
      "source": [
        "### Decoder\n",
        "O decoder também será composto de uma pilha de $ N = 6 $ camadas idênticas."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {
        "id": "D1LyoQZqlwWE"
      },
      "outputs": [],
      "source": [
        "class Decoder(nn.Module):\n",
        "    '''\n",
        "    Generic N layer decoder with masking.\n",
        "    '''\n",
        "    def __init__(self, layer, N):\n",
        "        super(Decoder, self).__init__()\n",
        "        self.layers = clones(layer, N)\n",
        "        self.norm = LayerNorm(layer.size)\n",
        "        \n",
        "    def forward(self, x, memory, src_mask, tgt_mask):\n",
        "        for layer in self.layers:\n",
        "            x = layer(x, memory, src_mask, tgt_mask)\n",
        "        \n",
        "        return self.norm(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "oL04HLG2EFV0"
      },
      "source": [
        "Além das duas subcamadas em cada camada do codificador, o decodificador insere uma terceira subcamada, que realiza a atenção de múltiplas cabeças sobre a saída da pilha do codificador. Semelhante ao codificador, empregamos conexões residuais em torno de cada uma das subcamadas, seguidas de normalização de camada."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {
        "id": "EXgylmPJl4ZU"
      },
      "outputs": [],
      "source": [
        "class DecoderLayer(nn.Module):\n",
        "    '''\n",
        "    Decoder is made of self-attn, src-attn, and feed forward (defined below)\n",
        "    '''\n",
        "    def __init__(self, size, self_attn, src_attn, feed_forward, dropout):\n",
        "        super(DecoderLayer, self).__init__()\n",
        "        self.size = size\n",
        "        self.self_attn = self_attn\n",
        "        self.src_attn = src_attn\n",
        "        self.feed_forward = feed_forward\n",
        "        self.sublayer = clones(SublayerConnection(size, dropout), 3)\n",
        " \n",
        "    def forward(self, x, memory, src_mask, tgt_mask):\n",
        "        m = memory\n",
        "        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))\n",
        "        x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))\n",
        "        \n",
        "        return self.sublayer[2](x, self.feed_forward)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "glRYk2niEzbJ"
      },
      "source": [
        "Também modificamos a subcamada de autoatenção na pilha do decodificador para evitar que as posições atendam às posições subsequentes. Esse mascaramento, combinado com o fato de que os embeddings de saída são deslocados em uma posição, garante que as previsões de posição $ i $ pode depender apenas das saídas conhecidas em posições menores que $ i $.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {
        "id": "XLLzL2bBmcnr"
      },
      "outputs": [],
      "source": [
        "def subsequent_mask(size):\n",
        "    '''\n",
        "    Mask out subsequent positions.\n",
        "    '''\n",
        "    attn_shape = (1, size, size)\n",
        "    subsequent_mask = np.triu(np.ones(attn_shape), k = 1).astype('uint8')\n",
        "    \n",
        "    return torch.from_numpy(subsequent_mask) == 0"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "b9M3FFlxFK_S"
      },
      "source": [
        "Abaixo a máscara de atenção mostra a posição que cada palavra alvo, por linha, pode olhar, representada pela coluna. As palavras são bloqueadas para atender a futuras palavras durante o treinamento."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 605
        },
        "id": "ORPGtoMum-Yd",
        "outputId": "da191990-aa1d-4004-dce6-4ccdef3bfa52"
      },
      "outputs": [
        {
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAfIAAAHfCAYAAACro1o4AAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMSwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/YYfK9AAAACXBIWXMAAAsTAAALEwEAmpwYAAAkIklEQVR4nO3de5SlVX3m8e+DIJcoTVpZoZvEC0FAJ3E1RsELxlYydpyVII4iowSjLu/RjEZQM14wgybo6IiKt8Qoy3gBxICa6GqDA16iyGhsEZWLLrlZaBBoLnIR6N/88b49ORanqk4V53TVPv39rHXWqd7v3vvdb++qes57rVQVkiSpTTss9wAkSdLSGeSSJDXMIJckqWEGuSRJDdtxuQcwSUnuoPuwcsNyj0WSpCXaHdhSVUMzO9N81XqSLUBW7T7+Aw+/uOEeY+9TkqTZ7uB2gKqqoWE21XvkwA2rdt9h1bUX7TP2jjesXTf2PiVJmu2c+jR3cPucR5Y9Ry5JUsMMckmSGmaQS5LUMINckqSGjT3Ik+yc5C1JZpLckuTcJIeO2HbvJKcl2ZzkhiRnJnnguMcoSdK0mMQe+cnAK4CPAv8d2AJ8Psmj5muU5F7A2cBjgTcDxwEPA85J8usTGKckSc0b6+1nSQ4C/hvwiqo6sS/7CHAB8Bbg9+dp/hJgX+D3qurbfdvP921fAbxhnGOVJGkajHuP/GnA7cAHtxZU1a3A3wOHJFmzQNtzt4Z43/ZC4IvA08c8TkmSpsK4HwhzIHBhVd00q/w8IMA64KrZjZLsADwU+NshfZ4H/Ocku1XVzbPabV5gPKtGG7YkSW0a9x75GoYE9UDZ2jnarQZ2nqdt+r4lSdKAce+R7wrcNqT81oHlc7VjsW2rao/5BtPvsbtXLkmaWuPeI7+Fbs96tl0Gls/VjiW2lSRpuzXuIL+K4YfAt5bNzNHuWrq98bnaFsMPu0uStF0bd5BvAg7o7wkfdHD//p1hjapqC/Bd4OFDFh8MXDL7QjdJkjT+ID8d2Al43taCJDsDzwH+tapm+rL7JTlgSNtHJjlwoO3+wBOAT455nJIkTYWxXuxWVd9I8kngrf094z8C/hS4P/DsgaofAR5HdzX6Vu8Fng98LsnbgTuAv6A7pP6OcY5TkqRpMe6r1gGeBRzfv/86cD7wX6rqX+drVFU3JllPF9qvpztacDbw8qq6ZgLjlCSpeWMP8v5Jbsf2r7nqrJ+j/ErgiHGPSZKkaeWfMZUkqWEGuSRJDZvEOfLtwsaZTWPvc8PadWPvU5I03dwjlySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSw3Zc7gHoP2yc2TSRfjesXTeRfiVJy889ckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSw8Ya5EkekeQ9Sb6f5BdJLk9ySpJ9R2j7xiQ15PXTcY5RkqRpsuOY+3s18Bjgk8D5wF7AS4FvJzmoqn4wQh8vBG4e+PctYx6jJElTY9xB/r+BZ1bVL7cWJDkV+C5dyD97hD5Oq6rNYx6XJElTaayH1qvqa4Mh3pddAnwPePCI3STJ7kkyzrFJkjSNJn6xWx/IvwH8fMQmlwPXA9cn+VCS1RMbnCRJjRv3ofVhjgL2Bl67QL3rgHcD5wK/BJ5Ad778YUkOrqrbZjdIsnmBPlcterRTaOPMpon0u2Htuon0K0ka3USDPMkBwHuArwL/MF/dqnrnrKLTk1zQt38W8HcTGaQkSQ2b2KH1JHsB/0y3p31EVW1ZQjfvp7uC/dBhC6tqj/ledIfoJUmaWhPZI0+yCvg83aHtx1TVku4Fr6otSX4CeJ5ckqQhxr5HnmQX4LPAfsAfVdVFd6OvnYDfAq4e0/AkSZoq436y2z2AU4FH0R1OP3eOevfrz58Plu05pOqxwC7AxnGOU5KkaTHuQ+tvBw6j2yNfneRPBpbdVFVn9l9/BHgcMHiv+GVJTgEuAG4DHg88le5CuY+PeZySJE2FcQf5uv79j/vXoMuAM+dp+zG6x7seAdwTuBQ4HvibqrpjnIOUJGlajDXIq2r9UutV1fPHORZJkrYH/hlTSZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWHj/jOm2o5snNk0kX43rF03kX4laRq5Ry5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNWzH5R6ANNvGmU1j73PD2nVj71OSVgL3yCVJaphBLklSwwxySZIaZpBLktQwg1ySpIaNNciTrE9Sc7wOGKH93klOS7I5yQ1JzkzywHGOUZKkaTKp289OBL41q2xmvgZJ7gWcDdwbeDNwB/AK4Jwk66rqugmMU5Kkpk0qyL9UVWcuss1LgH2B36uqbwMk+TxwAV2gv2GsI5QkaQpM7Bx5knsnWcwHhacB524NcYCquhD4IvD0cY9PkqRpMKkg/wfgBuCWJF9I8rvzVU6yA/BQ4JtDFp8H7JdktyHtNs/3Albd/U2RJGnlGveh9V8CpwOfB35OF87HAF9N8oiquniOdquBnYGrhiy7CgiwBvjRmMcrSVLTxhrkVfU14GsDRZ9J8lm6Pe3jgKPmaLpr/37bkGW3zqozuL495huPe+WSpGk38fvIq+o7wFnAofNUu6V/33nIsl1m1ZEkSb1t9UCYK+gOn8/lWrq98TVDlq0BiuGH3SVJ2q5tqyDfB7h6roVVtQX4LvDwIYsPBi6pqpsnNDZJkpo17ie77Tmk7BDg8cDGgbL7DXnS2+nAI5McOFBvf+AJwCfHOU5JkqbFuK9aPzXJzXQXvP0c+B3gBf3Xbxyo9xHgcXRXo2/1XuD5wOeSvJ3uyW5/QXdI/R1jHqckSVNh3EF+Jt2V6a8Edgf+Hfg48Maquny+hlV1Y5L1dKH9erqjBWcDL6+qa8Y8TkmSpsK4bz97F/CuEeqtn6P8SuCIcY5JkqRp5p8xlSSpYZP6oynSirJxZtNE+t2wdt1E+pWkUblHLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1bMflHoDUso0zmybS74a16ybSr6Tp4x65JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhYw3yJCcnqXlee8/T9o1ztPnpOMcoSdI02XHM/X0AOGtWWYD3A5dW1U9G6OOFwM0D/75lTGOTJGnqjDXIq+rrwNcHy5IcAuwGfGzEbk6rqs3jHJckSdNqW5wjfyZQwMdHrJ8kuyfJBMckSdJUGPeh9V+RZCfg6cDXqurSEZtdDtwLuDHJ6cAxVXXtHP1vXqCvVSOuU5KkJk00yIENwH0Y7bD6dcC7gXOBXwJPoDtf/rAkB1fVbRMbpbTCbJzZNJF+N6xdN5F+JS2fSQf5M4HbgdMWqlhV75xVdHqSC4D3AM8C/m5Imz3m67PfY3evXJI0tSZ2jjzJvYAnAxur6poldvN+uivYDx3bwCRJmiKTvNjtcBZ3tfpdVNUW4CfA6jGNSZKkqTLJID8KuAn4zFI76C+W+y3g6nENSpKkaTKRIE+yJ/AHwBlVdfOQ5fdLcsCQNrMdC+wCbJzEOCVJat2kLnY7su97rsPqHwEeR/fUt60uS3IKcAFwG/B44KnAVxn9HnRJkrYrkwryo4B/566Pa53Px4DHAEcA9wQuBY4H/qaq7hj3ACVJmgYTCfKqetQCy9cPKXv+JMYiSdI088+YSpLUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ2b1J8xlbQCbZzZNPY+N6xdN/Y+JY3OPXJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDVsx+UegKS2bZzZNJF+N6xdN5F+pWnjHrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIaNFORJ1iQ5IcnZSW5MUknWz1H3sCT/luTWJJcnOS7JSLe5JdkhyauS/Lhvf36SI0ffHEmSti+j7pHvD7wa+E3g/LkqJXkScCZwLfCy/us3AO8YcT1vBt4CfKFvfzlwSpKnjdhekqTtyqgPhPkWcN+quibJ4cAZc9R7G/BtYENV3QmQ5AbgL5O8q6oumWsFSfYGXgm8s6pe3pd9EPgS8LYk/1hVW0YcryRJ24WR9sir6saquma+OkkeAjwE+MDWEO+9t1/PUxdYzZOBnfr6W9dbwPuA+wMHjTJWSZK2J+N8ROuB/fs3BwuraibJlQPL52t/Q1VdPKv8vIHl5w4uSLJ5gT5XLbBckqSmjfOq9TX9+1VDll0FrB2h/U/naMsI7SVJ2u6Mc4981/79tiHLbgV2G6H9XG0H+///qmqP+Trs99jdK5ckTa1x7pHf0r/vPGTZLgPL52s/V9vB/iVJUm+cQb71EPiaIcvWADMjtN9rjraM0F6SpO3OOIN8U//+8MHCJGvp7j/fxPw2Absn2W9W+cGz+pckSb2xBXlVfQ+4EHhBknsMLHoxsAX41NaCJKuSHJBk8Pz1p4HbgZcM1AvwIroHw3xjXGOVJGlajHyxW5LX9V8+uH8/OskhwOaqOqkvOxb4DLAxyanA7wAvpbu3fPC2sqcAHwaeA5wMUFVXJjkROCbJLnS3sR0OPBY40ofBSJJ0V4u5av34Wf9+bv9+GXASQFX9U5L/ChwHvBu4GnjTkLZzeQ1wHfBCupC/GHhmVZ22iHFKkrTdSPfwtOmUZPOq3XdYde1F+yz3UCQt0oa165Z7CNKKcE59mju4/fq5brn2z5hKktSwcT4QRpLGZuPMpon0656+po175JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSw3Zc7gFI0ra0cWbTRPrdsHbdRPqVFuIeuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYSMFeZI1SU5IcnaSG5NUkvWz6twnybFJvpLk6iSbk3w9yREjruMBfb/DXn+4+E2TJGn67Thivf2BVwM/BM4HHj2kzqOANwOfA94E3AE8FTgtyRuq6vgR1/VRYOOssu+M2FaSpO3KqEH+LeC+VXVNksOBM4bU+R7woKq6bGtBkvcCZwF/meRtVXXLKOuqqo+OOC5JkrZrIx1ar6obq+qaBer8eDDE+7ICzgR2BR4w6qCS/FqSe45aX5Kk7dW2uNhtr/795yPWPx64Cbi1P8f++3NV7M/Dz/kCVt29oUuStLKNemh9SZKsBp4HnFNVVy9QfQvdufEzgBngQcAxwFlJDq2qr0xyrJJ0d2yc2TT2PjesXTf2PjV9JhbkSXYAPka3V/znC9WvqsuBX7k6PckpwPeBE4DHDGmzxwJj2Ix75ZKkKTbJQ+vvBjYAz6mq7y6lg6qaAT4BPDLJbuMcnCRJ02AiQZ7kOOAlwKuq6hN3s7sr6Ma5x90dlyRJ02bsQZ7kz4A3Au+oqreNoct9gDuB68bQlyRJU2WsQZ7kSOBddOfGXzlPvVVJDkiyaqBszyH19gWeAXx5xHvQJUnarox8sVuS1/VfPrh/PzrJIcDmqjopyUHAR4BrgC8CRyUZ7OJfqupn/ddPAT4MPAc4uS97a5J9+rZXAb8NvKhfdsxiNkqSpO3FYq5an/2I1ef275cBJwEPAe4J7Al8aEj7xwM/G1K+1RfogvtldOfDr+vL/qqqvreIcUqStN1I9/C16ZRk86rdd1h17UX7LPdQJGnRvI9cAOfUp7mD26+f65Zr/4ypJEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhq2mL9+JknahjbObJpIv/4xluniHrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1LAdl3sAkqRta+PMpon0u2Htuon0q/m5Ry5JUsMMckmSGmaQS5LUMINckqSGGeSSJDVspCBPsibJCUnOTnJjkkqyfki9S/tls18njLieHZK8KsmPk9ya5PwkRy5ukyRJ2n6MevvZ/sCrgR8C5wOPnqfut4ATZ5VdMOJ63gy8Bvhb4JvAk4FTktxZVaeP2IckSduNUYP8W8B9q+qaJIcDZ8xT98qq+uhiB5Jkb+CVwDur6uV92QeBLwFvS/KPVbVlsf1KkjTNRjq0XlU3VtU1o3aaZOckuy1yLE8GdgLeO7DeAt4H3B84aJH9SZI09SZxsdsTgV8Av0jyoyQvGLHdgcANVXXxrPLzBpb/iiSb53sBq5a6EZIktWDcj2g9H/gKcDGwJ/B84ANJVlfVQhe8rQF+OqT8qv597dhGKUnSlBhrkFfVYYP/TvJh4KvA65O8r6qun6f5rsBtQ8pvHVg+e317zDce98olSdNuoveRV9WddFew7wY8aoHqtwA7DynfZWC5JEkasC0eCHNF/756gXpXAXsNKV/Tv8+MbUSSJE2JbRHk+/TvVy9QbxOwe5L9ZpUfPLBckiQNGFuQJ1mdZIdZZbsAxwI3Al8fKF+V5IAkg+evPw3cDrxkoF6AFwGXA98Y11glSZoWI1/sluR1/ZcP7t+PTnIIsLmqTgIOA16b5HTgUuA+wJ8C+wEvrqqbBrp7CvBh4DnAyQBVdWWSE4Fj+g8A3wQOBx4LHOnDYCRJuqvFXLV+/Kx/P7d/vww4CfgucCFwNN2tZ7cB/wa8sqr+acR1vAa4DnghXchfDDyzqk5bxDglSdpupHt42nRKsnnV7jusuvaifRauLEm6WzasXbfcQ5hK59SnuYPbr5/rlmv/jKkkSQ0zyCVJati4H9EqSdpObZzZNJF+PWQ/P/fIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSG7bjcA5AkaT4bZzaNvc8Na9eNvc/l4h65JEkNM8glSWqYQS5JUsMMckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhIwV5kjVJTkhydpIbk1SS9bPqrO/L53q9doF1PGCetn+49E2UJGl67Thivf2BVwM/BM4HHj2kzg+Ao4eUHw08EfjCiOv6KLBxVtl3RmwrSdJ2ZdQg/xZw36q6JsnhwBmzK1TVz+hC+FckOQ64pKr+76jrqqq79CNJku5qpEPrVXVjVV2z2M6THATsC3xske1+Lck9F7s+SZK2N6PukS/VUf37YoL8eOAdQCX5BvDqqvrysIpJNi/Q16pFrFeStJ3YOLNpIv1uWLtuIv3OZ2JBnuQewJHAeVX1wxGabKE7N34GMAM8CDgGOCvJoVX1lUmNVZKkVk1yj/xQ4DeAvx6lclVdDvzK1elJTgG+D5wAPGZImz3m67PfY3evXJI0tSZ5H/lRwJ3AqUvtoKpmgE8Aj0yy27gGJknStJhIkCfZFXgKcFZ/NfvdcQXdOPe4u+OSJGnaTGqP/DDg3izyavU57EO3Z3/dGPqSJGmqTCrInwnczJD7zQGSrEpyQJJVA2V7Dqm3L/AM4MtVdcuExipJUrNGvtgtyev6Lx/cvx+d5BBgc1WdNFBvNfAk4FNVddMc3T0F+DDwHODkvuytSfYBvghcBfw28KJ+2TGjjlOSpO3JYq5aP37Wv5/bv18GnDRQfgSwE/DxRY7lC3TB/TK68+HX9WV/VVXfW2RfkiRtF1JVyz2GiUmyedXuO6y69qJ9lnsokqTtwCQeCHNOfZo7uP36uW659s+YSpLUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhi/nrZ5IkaR4bZzaNvc/V+9/J9TfMvdw9ckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWEGuSRJDTPIJUlqmEEuSVLDDHJJkhpmkEuS1DCDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktQwg1ySpIYZ5JIkNSxVtdxjmJgkW4Cs2t3PK5KkNl1/wxaAqqqhYTbtQX4H3VGHG0aovqp/v35yI1o207ptbld7pnXb3K72tLRtuwNbqmrHYQunOsgXI8lmgKraY3lHMn7Tum1uV3umddvcrvZM07Z5zFmSpIYZ5JIkNcwglySpYQa5JEkNM8glSWqYQS5JUsMMckmSGuZ95JIkNcw9ckmSGmaQS5LUMINckqSGGeSSJDXMIJckqWFTH+RJdk7yliQzSW5Jcm6SQ0dsu3eS05JsTnJDkjOTPHDSYx5FkkckeU+S7yf5RZLLk5ySZN8R2r4xSQ15/XRbjH2Bsa2fY2yV5IAR2q/kOTt5nm2rJHvP03ZFzFmSNUlOSHJ2khv7Mayfo+5hSf4tya399+dxSYb+GcYhbXdI8qokP+7bn5/kyHFuy6z1LbhdSe6T5NgkX0lydf899vUkR4y4jgfMM/d/OInt6tc70pwluXSOsZ0w4npW4pzN9/ukkrx2gXUsy5wt1kg/VI07GXgqcCLwQ+DZwOeTPK6qvj5XoyT3As4G7g28GbgDeAVwTpJ1VXXdZIe9oFcDjwE+CZwP7AW8FPh2koOq6gcj9PFC4OaBf98y9lEu3YnAt2aVzczXoIE5+wBw1qyyAO8HLq2qn4zQx3LP2f5033s/pPu+e/SwSkmeBJwJ/B/gZcDvAm8A7tv/eyFvBl4D/C3wTeDJwClJ7qyq0+/eJgw1ynY9qh/X54A30X1/PRU4Lckbqur4Edf1UWDjrLLvLGXQIxppznrfovvZG3TBiOtZiXP2A+DoIeVHA08EvjDiurb1nC1OVU3tCzgIKODlA2W70E38lxdo+ypgC3DgQNkBdD+8/3MFbNujgXvOKnsQcCtw8gJt39j/v+yx3NsxZGzr+7EdvoS2K3rO5hjzIf32/o8W5ozuQ9J9+q8P78e0fki979GFwj0Gyt4E3Ak8aIF17A38EjhxoCzAl4FLgR2WY7uABwL3n1UW4It0H652XWAdD5j9+2iFzdmlwJlLXMeKnLN52l4CXDxCvWWZs8W+pv3Q+tOA24EPbi2oqluBvwcOSbJmgbbnVtW3B9peSPdD+/TJDHd0VfW1qvrlrLJL6H6BPnjEbpJk9yQZ+wDHIMm9Rz0U21vRczaHZ9L9ovj4iPWXdc6q6saquma+OkkeAjwE+EBV3Tmw6L10p/OeusBqngzs1Nffut4C3gfcn+4D+liNsl1V9eOqumxWWdEdediV7pf+SJL8WpJ7LmGoizbKtg1Kdzpyt0WuZkXO2TBJDgL2BT62yHbbbM4Wa9qD/EDgwqq6aVb5eXSfFtcNa5RkB+ChdIeHZjsP2G8J3+gT1/9y/w3g5yM2uRy4Hrg+yYeSrJ7Y4BbvH4AbgFuSfCHJ785XucU5S7IT3QeMr1XVpSM2W8lzttWB/fuvzEVVzQBXDiyfr/0NVXXxrPLzZvW/UuzVv4/6c3c8cBNwa3+O/fcnM6wleSLwC+AXSX6U5AUjtmtpzo7q3xcT5Ct5zqb+HPkaYNh5x6v697VztFsN7DxQb3bb9H3/6O4OcMyOojvENe8FHMB1wLuBc+kOhz2B7tzrw5IcXFW3TXSU8/slcDrwebpfjA8FjgG+muQRQ35RbNXinG0A7sNov1BW8pzNtvVI11xzMdfP3WD7YRfxLfRzu831H6SeB5xTVVcvUH0L3XnWM+iu93gQ3ff2WUkOraqvTHSwCzsf+ApwMbAn8HzgA0lWV9VCF7w1MWdJ7gEcCZxXVT8coclKnzNg+oN8V2DYL7hbB5bP1Y4ltl0W6a7ofg/wVbq92TlV1TtnFZ2e5IK+/bOAv5vIIEdQVV8DvjZQ9Jkkn6XbuzuO//g0PVtzc0Z3WP124LSFKq7kORtioblY6MjIUn9ut6n+KNDHgFXAny9Uv6ouB37lSuckpwDfB06gu3h12VTVYYP/TvJhut8nr0/yvqq6fp7mTcwZcCjdUcu/HqXySp+zrab90PotdHtps+0ysHyudiyx7TaXZC/gn+n22o6oqi1L6Ob9dBfsjHRr3rZUVd+hu9p7vrG1Nmf3ojuvuHEp5/l6K3XOFpqLheZhqT+329q76Y6qPKeqvruUDvrTDZ8AHrnSTv301zecSPfB61ELVG9lzo6iu+Dy1KV2sBLnbNqD/Cr+4zDfoK1lc93OdC3dp8u52hbDDxtuc0lW0R2GXgVsqKol3Vfch/9P6A5Rr0RXMP/Ympmz3uF0vyAXdcHNoBU8Z1v/n+eai3lvI+zb7zWkfKGf220myXHAS4BXVdUn7mZ3V9D9Lt7j7o5rAq7o3xf6HmthznYFngKcVVU/u5vdrag5m/Yg3wQc0O/9DDq4fx96H2D/C/K7wMOHLD4YuKSqbh6ybJtKsgvwWWA/4I+q6qK70ddOwG8BC53nWy77MM/YWpmzAUfRXTzzmaV2sILnbFP//itzkWQt8JsDy+drv3uS/WaVHzywfNkk+TO62wHfUVVvG0OX+9DtJS73cw6G2ad/X+h7bBMreM56h9HdsrbkD88DVtScTXuQn053S8TzthYk2Rl4DvCv/SESktwvd31q2Ol0h04OHGi7P91FRp+c9MAX0l+0cSrdIa8jqurcOerdZduS7Dmk6rF0h8FmP/Rgmxo2tiSHAI9nYGwtztlW/Tb+AXDGsA8Xrc3ZbFX1PeBC4AX99+lWL6a7eOhTWwuSrEpyQH9kaatP01078JKBegFeRHfV/jcmOPx5pXtS2bvowuCV89S7y3bN8b29L/AMuudaLNvh5ySr+3P+g2W70H2P3Qh8faC8qTkb8Ey6U1FnDFvY2pwNmuqL3arqG0k+Cby1v2f8R8Cf0t3X+OyBqh8BHkd3ZfNW76W7avNzSd5O91CRv6A7hPSOyY9+QW+n+4T5WWB1kj8ZWHZTVZ3Zfz1s2y7rL9i4gO5w9OPp7u39KqPfzzwppya5me6Ct58DvwO8oP/6jQP1WpyzrY6k+9mba89gRc9Zktf1X259XsHR/YetzVV1Ul92LN3Rho1JTqWbx5fS3Vs+eOfBU4AP0324Phmgqq5MciJwTB8m36Q7FfFY4MglXgNyt7erv//4I8A1dM8mOGrW7fz/MnDI9i7bRfd7aJ++7VXAb9MFHXRXQk/MCHN2GPDaJKfTPcDlPnS/K/cDXjzrFt5m5myg3mrgScCnhtyOvNWKmrNFWe4n0kz6RbfH8r/oJuFWuvsa/2BWnXPon18wq/w36fbkrqf7VPoZYJ/l3qbBMc/xunS+baO7wvn7/TbdBlwE/E8WeDLVNtquP6f79H4N3Sf8nwAfAu7X+pwNjPHrwM8YeOrZQtu2kuZslO+7vt7hwLf7n7srgL8CdpxV59l922fPKt8B+Eu6ULmN7rTJM5ZzuwbGOtdr/XzbRbcX9yW6w9S3A/9Od8fCf1ruOQN+j26n4Mr+//uG/vvwj4b01cycDdR7YV/+x/P0taLmbDGv9IOVJEkNmvZz5JIkTTWDXJKkhhnkkiQ1zCCXJKlhBrkkSQ0zyCVJaphBLklSwwxySZIaZpBLktSw/wcnijM1QUqfeAAAAABJRU5ErkJggg==",
            "text/plain": [
              "<Figure size 576x576 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ],
      "source": [
        "plt.figure(figsize = (8, 8))\n",
        "plt.imshow(subsequent_mask(20)[0])\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SgtUKWGknE97"
      },
      "source": [
        "## Attention\n",
        "Uma função de atenção pode ser descrita como o mapeamento de uma consulta e um conjunto de pares chave-valor para uma saída, onde a consulta, as chaves, os valores e a saída são todos vetores. A saída é calculada como uma soma ponderada dos valores, em que o peso atribuído a cada valor é calculado por uma função de compatibilidade da consulta com a chave correspondente.\n",
        "\n",
        "Chamamos nossa atenção particular de “Atenção de produto interno escalonado” (Scaled Dot-Product Attention). A entrada consiste em consultas e chaves de dimensão $ d_k $, e valores de dimensão $ d_v $. Calculamos os produtos internos da consulta com todas as chaves, dividimos cada um por $ \\sqrt{d_k} $, e aplicamos a [função softmax](https://developers.google.com/machine-learning/crash-course/multi-class-neural-networks/softmax) para obter os pesos dos valores.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 14,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 238
        },
        "id": "dw6hGazPGLDK",
        "outputId": "cac89c8c-9708-40db-a40f-76f2640d8a91"
      },
      "outputs": [
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<IPython.core.display.Image object>"
            ]
          },
          "execution_count": 14,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "Image(filename='Imagens/attention.png')"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "7sFkjUMyGC78"
      },
      "source": [
        "Na prática, calculamos a função de atenção num conjunto de consultas simultaneamente, agrupadas numa matriz $ Q $. As chaves e valores também são colocadas juntas em matrizes $ K $ e $ V $. Calculamos a matriz de saídas da seguinte forma:\n",
        "\n",
        "$ \n",
        "\\mathrm{Attention}(Q, K, V) = \\mathrm{softmax}(\\frac{QK^T}{\\sqrt{d_k}})V \n",
        "$"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 15,
      "metadata": {
        "id": "0gERvYadnCaV"
      },
      "outputs": [],
      "source": [
        "def attention(query, key, value, mask = None, dropout = None):\n",
        "    '''\n",
        "    Compute 'Scaled Dot Product Attention'\n",
        "    '''\n",
        "    d_k = query.size(-1)\n",
        "    scores = torch.matmul(query, key.transpose(-2, -1))/ math.sqrt(d_k)\n",
        "\n",
        "    if mask is not None:\n",
        "        scores = scores.masked_fill(mask == 0, -1e9)\n",
        "    \n",
        "    p_attn = F.softmax(scores, dim = -1)\n",
        "    \n",
        "    if dropout is not None:\n",
        "        p_attn = dropout(p_attn)\n",
        "    \n",
        "    return torch.matmul(p_attn, value), p_attn"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9KHPsCxLIJ9X"
      },
      "source": [
        "As duas funções de atenção mais utilizadas são a [atenção aditiva](https://arxiv.org/abs/1409.0473), e a atenção por produto interno (ou atenção multiplicativa). A atenção por produto interno é idêntica ao algoritmo apresentado, exceto pelo fator de escalonamento de $ \\frac{1}{\\sqrt{d_k}} $. Atenção aditiva calcula a função de compatibilidade utilizando a rede feed-forward com uma única camada oculta (hidden layer). Enquanto as duas são similares em complexidade teórica, atenção por produto interno é muito mais rápida e mais eficiente em termos de espaço na prática, pois pode ser implementada utilizando código de multiplicação de matriz altamente otimizado.\n",
        "\n",
        "Enquanto que para valores pequenos de $ d_k $ os dois mecanismos performam de maneira similar, [atenção aditiva supera a atenção por produto interno sem escalar para maiores valores de $ d_k $](https://arxiv.org/abs/1703.03906). A suspeita dos autores é de que para maiores valores de $ d_k $, os produtos internos crescem em magnitude, empurrando a função softmax para regiões em que ela possui gradientes extremamente pequenos (para ilustrar o porquê dos produtos internos cresceram, assuma que os componentes de $ q $ e $ k $ sejam variáveis independetes aleatórias com média 0 e variância 1. Então, o produto interno deles, $ q \\cdot k = \\sum_{i=1}^{d_k} q_ik_i $, tem média 0 e variância $ d_k $ ). Para contra-atacar esse efeito, escalonamos o produto interno por $ \\frac{1}{\\sqrt{d_k}} $."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 16,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 658
        },
        "id": "Yc6EdBPCIJpp",
        "outputId": "79110359-8cdc-40d7-9482-6aeaae6da1a0"
      },
      "outputs": [
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<IPython.core.display.Image object>"
            ]
          },
          "execution_count": 16,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "Image(filename='Imagens/multi-head.png')"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "kII0btnuIJX7"
      },
      "source": [
        "Atenção de Múltiplicas Cabeças (Multi-head attention) permite que o modelo atenda conjuntamente a informações de diferentes subespaços de representação em diferentes posições. Com atenção de uma única cabeça, o cálculo da média inibe isso.\n",
        "\n",
        "$\n",
        "\\mathrm{MultiHead}(Q, K, V) = \\mathrm{Concat}(\\mathrm{head_1}, ...,\n",
        "\\mathrm{head_h})W^O    \\\\\n",
        "    \\text{em que}~\\mathrm{head_i} = \\mathrm{Attention}(QW^Q_i, KW^K_i, VW^V_i)\n",
        "$\n",
        "\n",
        "Em que as projeções são matrizes de parâmetros $ W^Q_i \\in \\mathbb{R}^{d_{\\text{model}} \\times d_k} $, $ W^K_i \\in \\mathbb{R}^{d_{\\text{model}} \\times d_k} $, $ W^V_i \\in \\mathbb{R}^{d_{\\text{model}} \\times d_v} $ e $ W^O \\in \\mathbb{R}^{hd_v \\times d_{\\text{model}}} $. Nesse trabalho, são utilizadas $ h = 8 $ camadas de atenção paralelas, ou \"cabeças\". Para cada uma dessas é usado $ d_k = d_v = d_{model}/ h = 64 $. Devido à dimensão reduzida de cada cabeça, o custo computacional total é semelhante ao da atenção de cabeça única com dimensionalidade total. "
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 17,
      "metadata": {
        "id": "Opjy34lwnfBt"
      },
      "outputs": [],
      "source": [
        "class MultiHeadedAttention(nn.Module):\n",
        "    def __init__(self, h, d_model, dropout):\n",
        "        '''\n",
        "        Take in model size and number of heads.\n",
        "        '''\n",
        "        super(MultiHeadedAttention, self).__init__()\n",
        "        \n",
        "        assert d_model % h == 0\n",
        "        \n",
        "        # We assume d_v always equals d_k\n",
        "        self.d_k = d_model // h\n",
        "        self.h = h\n",
        "        self.linears = clones(nn.Linear(d_model, d_model), 4)\n",
        "        self.attn = None\n",
        "        self.dropout = nn.Dropout(p=dropout)\n",
        "        \n",
        "    def forward(self, query, key, value, mask = None):\n",
        "        '''\n",
        "        Implements figure above\n",
        "        '''\n",
        "        if mask is not None:\n",
        "            # Same mask applied to all h heads.\n",
        "            mask = mask.unsqueeze(1)\n",
        "\n",
        "        nbatches = query.size(0)\n",
        "        \n",
        "        # 1) Do all the linear projections in batch from d_model => h x d_k \n",
        "        query, key, value = [\n",
        "            l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)\n",
        "            for l, x in zip(self.linears, (query, key, value))\n",
        "        ]\n",
        "        \n",
        "        # 2) Apply attention on all the projected vectors in batch. \n",
        "        x, self.attn = attention(query, key, value, mask = mask, dropout = self.dropout)\n",
        "        \n",
        "        # 3) \"Concat\" using a view and apply a final linear. \n",
        "        x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)\n",
        "        \n",
        "        return self.linears[-1](x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uVNBFkfeXa-F"
      },
      "source": [
        "### Aplicações de Atenção no nosso Modelo\n",
        "O Transformer usa Multi-Head attention em três maneiras diferentes:\n",
        "\n",
        "1. nas camadas de atenção \"encoder-decoder\", as consultas vêm da camada do decodificador anterior e as chaves de memória e os valores vêm da saída do codificador. Isso permite que cada posição no decodificador atenda a todas as posições na sequência de entrada. Isso imita [os mecanismos típicos de atenção do codificador-decodificador em modelos de sequência para sequência (sequence-to-sequence)](https://arxiv.org/abs/1609.08144);\n",
        "2. o codificador contém camadas de autoatenção. Em uma camada de autoatenção todas as chaves, valores e consultas vêm do mesmo local, neste caso, a saída da camada anterior no codificador. Cada posição no codificador pode atender a todas as posições na camada anterior do codificador;\n",
        "3. da mesma forma, as camadas de autoatenção no decodificador permitem que cada \n",
        "posição no decodificador atenda a todas as posições no decodificador até essa posição, inclusive. Precisamos evitar o fluxo de informações para a esquerda no decodificador para preservar a propriedade auto-regressiva. Implementamos isso dentro da atenção de produto escalar escalonado, mascarando (definindo como $- \\infty $) todos os valores na entrada do softmax que correspondem a conexões ilegais.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "NOhIifYcoMch"
      },
      "source": [
        "### Position-wise Feed-Forward Networks\n",
        "Além das subcamadas de atenção, cada uma das camadas em nosso codificador e decodificador contém uma rede feed-forward totalmente conectada, que é aplicada a cada posição separadamente e de forma idêntica. Ela consiste em duas transformações lineares com uma [ativação ReLU](https://machinelearningmastery.com/rectified-linear-activation-function-for-deep-learning-neural-networks/) no meio.\n",
        "\n",
        "$ \n",
        "\\mathrm{FFN}(x)=\\max(0, xW_1 + b_1) W_2 + b_2 \n",
        "$\n",
        "\n",
        "Embora as transformações lineares sejam as mesmas em diferentes posições, elas usam parâmetros diferentes de camada para camada. Outra maneira de descrever isso é como duas convoluções com tamanho de kernel igual à 1. A dimensionalidade de entrada e saída é $ d_{model}= 512 $, e a camada interna tem dimensionalidade $ d_{ff} = 2048 $.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 18,
      "metadata": {
        "id": "prtJAefZoLqk"
      },
      "outputs": [],
      "source": [
        "class PositionwiseFeedForward(nn.Module):\n",
        "    '''\n",
        "    Implements FFN equation.\n",
        "    '''\n",
        "    def __init__(self, d_model, d_ff, dropout):\n",
        "        super(PositionwiseFeedForward, self).__init__()\n",
        "        self.w_1 = nn.Linear(d_model, d_ff)\n",
        "        self.w_2 = nn.Linear(d_ff, d_model)\n",
        "        self.dropout = nn.Dropout(dropout)\n",
        "\n",
        "    def forward(self, x):\n",
        "        return self.w_2(self.dropout(F.relu(self.w_1(x))))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "vQvdtHPpo9ou"
      },
      "source": [
        "### Embeddings e Softmax\n",
        "Da mesma forma que outros modelos de tradução de sequência, usamos embeddings pré-treinados para converter os tokens de entrada e os tokens de saída em vetores de dimensão $ d_{model} $. Também usamos a transformação linear pré-treinada usual e a função softmax para converter a saída do decodificador em probabilidades previstas do próximo token. Em nosso modelo, [compartilhamos a mesma matriz de peso entre as duas camadas de embedding e a transformação linear pré-softmax](https://arxiv.org/abs/1608.05859). Nas camadas de embedding, multiplicamos esses pesos por $ \\sqrt{d_{model}} $. "
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 19,
      "metadata": {
        "id": "bbO6PQzCo8rV"
      },
      "outputs": [],
      "source": [
        "class Embeddings(nn.Module):\n",
        "    def __init__(self, d_model, vocab):\n",
        "        super(Embeddings, self).__init__()\n",
        "        self.lut = nn.Embedding(vocab, d_model)\n",
        "        self.d_model = d_model\n",
        "\n",
        "    def forward(self, x):\n",
        "        return self.lut(x) * math.sqrt(self.d_model)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "OnuOqMjYpNNY"
      },
      "source": [
        "### Positional Encoding\n",
        "Como nosso modelo não contém recorrência e nem convolução, para que o modelo faça uso da ordem da sequência, devemos injetar algumas informações sobre a posição relativa ou absoluta dos tokens na sequência. Para isso, adicionamos “codificações posicionais” aos embeddings de entrada na parte inferior das pilhas do codificador e do decodificador. As codificações posicionais têm a mesma dimensão $d_{model} $ dos embeddings, para que os dois possam ser somados. [Há muitas escolhas de codificação posicional, pré-treinada e fixa](https://arxiv.org/pdf/1705.03122.pdf).\n",
        "\n",
        "Neste trabalho, usamos funções seno e cosseno de diferentes frequências: \n",
        "\n",
        "$ \n",
        "PE_{(pos,2i)} = sin(pos / 10000^{2i/d_{\\text{model}}}) \n",
        "$ \n",
        "$ \n",
        "PE_{(pos,2i+1)} = cos(pos / 10000^{2i/d_{\\text{model}}}) \n",
        "$ em que $ pos $ é a posição e $ i $ é a dimensão. Isto é, cada dimensão de codificação posicional corresponde à uma sinusóide. Os comprimentos de onda formam uma progressão geométrica de $ 2 \\pi $ até $ 10.000 \\times 2 \\pi$. Essa função foi escolhida pois os autores do artigo original hipotetizaram que ela permitiria o modelo aprender facilmente a atender por posições relativas, já que para qualquer deslocamento fixo $ k $, $ PE_{pos+k} $ pode ser representado como uma função linear de $ PE_{pos} $.\n",
        "\n",
        "Além disso, aplicamos dropout às somas dos embeddings e das codificações posicionais nas pilhas do codificador e do decodificador. Para o modelo base, usamos uma taxa de $ P_{drop} = 0.1 $."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 20,
      "metadata": {
        "id": "NtioNbA1pBUd"
      },
      "outputs": [],
      "source": [
        "class PositionalEncoding(nn.Module):\n",
        "    '''\n",
        "    Implement the PE function.\n",
        "    '''\n",
        "    def __init__(self, d_model, dropout, max_len = 5000):\n",
        "        super(PositionalEncoding, self).__init__()\n",
        "        self.dropout = nn.Dropout(p=dropout)\n",
        "        \n",
        "        # Compute the positional encodings once in log space.\n",
        "        pe = torch.zeros(max_len, d_model)\n",
        "        \n",
        "        position = torch.arange(0, max_len).unsqueeze(1)\n",
        "        div_term = torch.exp(torch.arange(0, d_model, 2) * -(math.log(10000.0) / d_model))\n",
        "        \n",
        "        pe[:, 0::2] = torch.sin(position * div_term)\n",
        "        pe[:, 1::2] = torch.cos(position * div_term)\n",
        "        pe = pe.unsqueeze(0)\n",
        "        \n",
        "        self.register_buffer('pe', pe)\n",
        "        \n",
        "    def forward(self, x):\n",
        "        x = x + Variable(self.pe[:, :x.size(1)], requires_grad = False)\n",
        "        return self.dropout(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-JlUbCl5amur"
      },
      "source": [
        "Os autores originais também experimentaram o uso de [incorporações posicionais pré-treinadas](https://arxiv.org/pdf/1705.03122.pdf) e descobriram que as duas versões produziram resultados quase idênticos. Para esse trabalho foi escolhida a versão senoidal porque ela pode permitir que o modelo extrapole para comprimentos de sequência maiores do que os encontrados durante o treinamento.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9S8fuqhipo6J"
      },
      "source": [
        "## Modelo completo\n",
        "Aqui definimos uma função que recebe hiperparâmetros e produz um modelo completo."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Hiperparâmetros"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 21,
      "metadata": {},
      "outputs": [],
      "source": [
        "n_epochs = 2\n",
        "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
        "BATCH_SIZE = 256\n",
        "\n",
        "# modelo\n",
        "N = 6\n",
        "d_model = 512\n",
        "d_ff = 2048\n",
        "h = 8\n",
        "dropout = 0.1\n",
        "V = 11\n",
        "\n",
        "# otimizador\n",
        "lr = 0 \n",
        "betas = (0.9, 0.98)\n",
        "eps = 1e-9\n",
        "factor = 1\n",
        "warmup_steps = 4000\n",
        "\n",
        "# suavização\n",
        "smoothing = 0.0"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 22,
      "metadata": {
        "id": "84dI4luJpjWt"
      },
      "outputs": [],
      "source": [
        "def make_model(src_vocab, tgt_vocab, N, d_model, d_ff, h, dropout):\n",
        "    '''\n",
        "    Helper: Construct a model from hyperparameters.\n",
        "    '''\n",
        "    c = copy.deepcopy\n",
        "    \n",
        "    attn = MultiHeadedAttention(h, d_model, dropout)\n",
        "    ff = PositionwiseFeedForward(d_model, d_ff, dropout)\n",
        "    position = PositionalEncoding(d_model, dropout)\n",
        "    \n",
        "    model = EncoderDecoder(\n",
        "        Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),\n",
        "        Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),\n",
        "        nn.Sequential(Embeddings(d_model, src_vocab), c(position)),\n",
        "        nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),\n",
        "        Generator(d_model, tgt_vocab)\n",
        "    )\n",
        "    \n",
        "    # This was important from their code. \n",
        "    # Initialize parameters with Glorot / fan_avg.\n",
        "    for p in model.parameters():\n",
        "        if p.dim() > 1:\n",
        "            nn.init.xavier_uniform_(p)\n",
        "    \n",
        "    return model"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 23,
      "metadata": {
        "id": "U20F5RXeprvN"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "EncoderDecoder(\n",
              "  (encoder): Encoder(\n",
              "    (layers): ModuleList(\n",
              "      (0): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (1): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (2): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (3): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (4): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (5): EncoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "    )\n",
              "    (norm): LayerNorm()\n",
              "  )\n",
              "  (decoder): Decoder(\n",
              "    (layers): ModuleList(\n",
              "      (0): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (1): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (2): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (3): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (4): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "      (5): DecoderLayer(\n",
              "        (self_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (src_attn): MultiHeadedAttention(\n",
              "          (linears): ModuleList(\n",
              "            (0): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (1): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (2): Linear(in_features=512, out_features=512, bias=True)\n",
              "            (3): Linear(in_features=512, out_features=512, bias=True)\n",
              "          )\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (feed_forward): PositionwiseFeedForward(\n",
              "          (w_1): Linear(in_features=512, out_features=2048, bias=True)\n",
              "          (w_2): Linear(in_features=2048, out_features=512, bias=True)\n",
              "          (dropout): Dropout(p=0.1, inplace=False)\n",
              "        )\n",
              "        (sublayer): ModuleList(\n",
              "          (0): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (1): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "          (2): SublayerConnection(\n",
              "            (norm): LayerNorm()\n",
              "            (dropout): Dropout(p=0.1, inplace=False)\n",
              "          )\n",
              "        )\n",
              "      )\n",
              "    )\n",
              "    (norm): LayerNorm()\n",
              "  )\n",
              "  (src_embed): Sequential(\n",
              "    (0): Embeddings(\n",
              "      (lut): Embedding(2, 512)\n",
              "    )\n",
              "    (1): PositionalEncoding(\n",
              "      (dropout): Dropout(p=0.1, inplace=False)\n",
              "    )\n",
              "  )\n",
              "  (tgt_embed): Sequential(\n",
              "    (0): Embeddings(\n",
              "      (lut): Embedding(2, 512)\n",
              "    )\n",
              "    (1): PositionalEncoding(\n",
              "      (dropout): Dropout(p=0.1, inplace=False)\n",
              "    )\n",
              "  )\n",
              "  (generator): Generator(\n",
              "    (proj): Linear(in_features=512, out_features=2, bias=True)\n",
              "  )\n",
              ")"
            ]
          },
          "execution_count": 23,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "# Small example model.\n",
        "tmp_model = make_model(2, 2, N = N, d_model = d_model, d_ff = d_ff, h = h, dropout = dropout).to(device)\n",
        "tmp_model"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SHVClNozpxJj"
      },
      "source": [
        "## Treinamento\n",
        "Esta seção descreve o regime de treinamento para nossos modelos."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wmZie2BipzK9"
      },
      "source": [
        "### Batches e Masking\n",
        "Paramos para um breve interlúdio para apresentar algumas das ferramentas necessárias para treinar um modelo de decodificador de codificador padrão. Primeiro, definimos um objeto batch que contém as sentenças src e target para treinamento, além de construir as máscaras."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 24,
      "metadata": {
        "id": "rcAXs12JpunN"
      },
      "outputs": [],
      "source": [
        "class Batch:\n",
        "    '''\n",
        "    Object for holding a batch of data with mask during training.\"\n",
        "    '''\n",
        "    def __init__(self, src, trg = None, pad = 0):\n",
        "        self.src = src\n",
        "        self.src_mask = (src != pad).unsqueeze(-2)\n",
        "        \n",
        "        if trg is not None:\n",
        "            self.trg = trg[:, :-1]\n",
        "            self.trg_y = trg[:, 1:]\n",
        "            self.trg_mask = self.make_std_mask(self.trg, pad)\n",
        "            self.ntokens = (self.trg_y != pad).data.sum()\n",
        "    \n",
        "    @staticmethod\n",
        "    def make_std_mask(tgt, pad):\n",
        "        '''\n",
        "        Create a mask to hide padding and future words.\n",
        "        '''\n",
        "        tgt_mask = (tgt != pad).unsqueeze(-2)\n",
        "        tgt_mask = tgt_mask & Variable(subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data))\n",
        "        \n",
        "        return tgt_mask"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hLvgYk3Dp5-a"
      },
      "source": [
        "### Loop de treinamento\n",
        "Em seguida, criamos uma função genérica de treinamento e pontuação para acompanhar a perda. Passamos em uma função de cálculo de perda (loss) genérica que também lida com atualizações de parâmetros."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 25,
      "metadata": {
        "id": "NQMkj2G7p2Md"
      },
      "outputs": [],
      "source": [
        "def run_epoch(data_iter, model, loss_compute):\n",
        "    '''\n",
        "    Standard Training\n",
        "    '''\n",
        "    start = time.time()\n",
        "    \n",
        "    total_tokens = 0\n",
        "    total_loss = 0\n",
        "    tokens = 0\n",
        "    \n",
        "    for i, batch in enumerate(data_iter):\n",
        "        out = model.forward(\n",
        "            batch.src, \n",
        "            batch.trg, \n",
        "            batch.src_mask, \n",
        "            batch.trg_mask\n",
        "        )\n",
        "        \n",
        "        loss = loss_compute(out, batch.trg_y, batch.ntokens)\n",
        "        total_loss += loss\n",
        "        total_tokens += batch.ntokens\n",
        "        tokens += batch.ntokens\n",
        "        \n",
        "        if i % 200 == 0:\n",
        "            elapsed = time.time() - start\n",
        "\n",
        "            print(\"Epoch Step: %d Loss: %f Tokens per Sec: %f\" %\n",
        "                (i, loss / batch.ntokens, tokens / elapsed)\n",
        "            )\n",
        "\n",
        "            start = time.time()\n",
        "            tokens = 0\n",
        "    \n",
        "    return total_loss/ total_tokens"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "PbooV3COqO4f"
      },
      "source": [
        "### Batching\n",
        "Os pares de sentenças foram agrupados pelo comprimento aproximado da sequência. Cada lote de treinamento continha um conjunto de pares de frases contendo aproximadamente 25.000 tokens de origem e 25.000 tokens de destino.\n",
        "\n",
        "Usaremos a biblioteca torchtext para lotes. Isso é discutido com mais detalhes abaixo. Aqui, criamos lotes em uma função torchtext que garante que nosso tamanho de lote preenchido até o tamanho máximo de lote não ultrapasse um limite (25.000, no caso dos autores originais que tinham 8 GPUs).\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 26,
      "metadata": {
        "id": "CmOEQ6-2p_7N"
      },
      "outputs": [],
      "source": [
        "global max_src_in_batch, max_tgt_in_batch\n",
        "\n",
        "def batch_size_fn(new, count, sofar):\n",
        "    '''\n",
        "    Keep augmenting batch and calculate total number of tokens + padding.\n",
        "    '''\n",
        "    global max_src_in_batch, max_tgt_in_batch\n",
        "    \n",
        "    if count == 1:\n",
        "        max_src_in_batch = 0\n",
        "        max_tgt_in_batch = 0\n",
        "    \n",
        "    max_src_in_batch = max(max_src_in_batch,  len(new.src))\n",
        "    max_tgt_in_batch = max(max_tgt_in_batch,  len(new.trg) + 2)\n",
        "    \n",
        "    src_elements = count * max_src_in_batch\n",
        "    tgt_elements = count * max_tgt_in_batch\n",
        "    \n",
        "    return max(src_elements, tgt_elements)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Y3r1mm5ShQ1-"
      },
      "source": [
        "### Hardware e cronograma\n",
        "Os autores originais treinaram os modelos em uma máquina com 8 GPUs NVIDIA P100. Para os modelos de base usando os hiperparâmetros descritos ao longo do artigo, cada etapa de treinamento levou cerca de 0,4 segundos. Treinaram os modelos básicos para um total de 100.000 passos ou 12 horas. Para os modelos grandes, o tempo de passo foi de 1,0 segundos. Os grandes modelos foram treinados para 300.000 passos (3,5 dias).\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "EUTMEMBdqWXP"
      },
      "source": [
        "### Otimizador\n",
        "Foi utilizado o otimizador [Adam](https://arxiv.org/abs/1412.6980) com $ \\beta_1=0.9 $, $ \\beta_2=0.98 $ e $ \\epsilon=10^{-9} $. Variamos a taxa de aprendizado ao longo do treinamento, de acordo com a fórmula: $ lrate = d_{\\text{model}}^{-0.5} \\cdot \\min({step\\_num}^{-0.5}, {step\\_num} \\cdot {warmup\\_steps}^{-1.5}) $. Isso corresponde a aumentar a taxa de aprendizado linearmente para o primeiro $ {warmup\\_steps} $ dos passos de treinamento, e diminuindo a partir daí proporcionalmente à raiz quadrada inversa do número do passo. Utilizamos $ {warmup\\_steps} = 4.000 $.\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 27,
      "metadata": {
        "id": "X-cQ3iynqRod"
      },
      "outputs": [],
      "source": [
        "class NoamOpt:\n",
        "    '''\n",
        "    Optim wrapper that implements rate.\n",
        "    '''    \n",
        "    def __init__(self, model_size, factor, warmup, optimizer):\n",
        "        self.optimizer = optimizer\n",
        "        self._step = 0\n",
        "        self.warmup = warmup\n",
        "        self.factor = factor\n",
        "        self.model_size = model_size\n",
        "        self._rate = 0\n",
        "        \n",
        "    def step(self):\n",
        "        '''\n",
        "        Update parameters and rate\n",
        "        '''\n",
        "        self._step += 1\n",
        "        \n",
        "        rate = self.rate()\n",
        "        \n",
        "        for p in self.optimizer.param_groups:\n",
        "            p['lr'] = rate\n",
        "        \n",
        "        self._rate = rate\n",
        "        self.optimizer.step()\n",
        "        \n",
        "    def rate(self, step = None):\n",
        "        '''\n",
        "        Implement 'lrate' above\n",
        "        '''\n",
        "        if step is None:\n",
        "            step = self._step\n",
        "        \n",
        "        return self.factor * ( \n",
        "            self.model_size ** (-0.5) * min(step ** (-0.5), \n",
        "            step * self.warmup ** (-1.5))\n",
        "        )"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 28,
      "metadata": {},
      "outputs": [],
      "source": [
        "def get_std_opt(model):\n",
        "    return NoamOpt(\n",
        "        model.src_embed[0].d_model, 2, warmup_steps,\n",
        "        torch.optim.Adam(model.parameters(), lr = lr, betas = betas, eps = eps)\n",
        "    )"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3r53pvIQq902"
      },
      "source": [
        "### Suavização de rótulos\n",
        "Durante o treinamento, empregamos a [suavização de rótulos](https://arxiv.org/abs/1512.00567) com valor $ \\epsilon_{ls} = 0.1 $. Isso prejudica a perplexidade, pois o modelo aprende a ser mais inseguro, mas melhora a precisão e a pontuação BLEU. \n",
        "\n",
        "Implementamos a suavização de rótulos usando a [perda de div KL](https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html). Em vez de usar uma distribuição de destino único ( one-hot target distribution ), criamos uma distribuição que tem *confiança* na palavra correta e no restante da massa de *suavização* distribuída por todo o vocabulário."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 29,
      "metadata": {
        "id": "TQ1aXAWAqjLt"
      },
      "outputs": [],
      "source": [
        "class LabelSmoothing(nn.Module):\n",
        "    '''\n",
        "    Implement label smoothing.\n",
        "    '''\n",
        "    def __init__(self, size, padding_idx, smoothing):\n",
        "        super(LabelSmoothing, self).__init__()\n",
        "        \n",
        "        self.criterion = nn.KLDivLoss(size_average = False)\n",
        "        self.padding_idx = padding_idx\n",
        "        self.confidence = 1.0 - smoothing\n",
        "        self.smoothing = smoothing\n",
        "        self.size = size\n",
        "        self.true_dist = None\n",
        "        \n",
        "    def forward(self, x, target):\n",
        "        \n",
        "        true_dist = x.data.clone()\n",
        "        true_dist.fill_(self.smoothing / (self.size - 2))\n",
        "        true_dist.scatter_(1, target.data.unsqueeze(1).long(), self.confidence)\n",
        "        true_dist[:, self.padding_idx] = 0\n",
        "        \n",
        "        mask = torch.nonzero(target.data == self.padding_idx)\n",
        "        \n",
        "        if mask.dim() > 0:\n",
        "            true_dist.index_fill_(0, mask.squeeze(), 0.0)\n",
        "        \n",
        "        self.true_dist = true_dist\n",
        "        \n",
        "        return self.criterion(x, Variable(true_dist, requires_grad = False))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "XegldZg3lZ2W"
      },
      "source": [
        "A suavização de rótulos na verdade começa a penalizar o modelo se ele ficar muito confiante sobre uma determinada escolha."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 30,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 314
        },
        "id": "keVODnJ7rCt9",
        "outputId": "2ac5d931-1963-4d09-f436-4e783f5a8a0c"
      },
      "outputs": [
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "C:\\Users\\User\\AppData\\Roaming\\Python\\Python39\\site-packages\\torch\\nn\\_reduction.py:42: UserWarning: size_average and reduce args will be deprecated, please use reduction='sum' instead.\n",
            "  warnings.warn(warning.format(ret))\n"
          ]
        },
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "needs_background": "light"
          },
          "output_type": "display_data"
        }
      ],
      "source": [
        "crit = LabelSmoothing(5, 0, 0.1)\n",
        "\n",
        "def loss(x):\n",
        "    d = x + 3\n",
        "    \n",
        "    predict = torch.FloatTensor([[0, x /d, 1/ d, 1/ d, 1/ d],])\n",
        "    \n",
        "    return crit(Variable(predict.log()),Variable(torch.LongTensor([1]))).data\n",
        "\n",
        "plt.plot(np.arange(1, 1000), [loss(x) for x in range(1, 1000)])\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2Mhv_0YksA3p"
      },
      "source": [
        "## Teste\n",
        "Podemos começar experimentando uma tarefa de cópia simples. Dado um conjunto aleatório de símbolos de entrada de um pequeno vocabulário, o objetivo é gerar de volta esses mesmos símbolos."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "WJoIfDE_sIeH"
      },
      "source": [
        "### Perda computacional"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 31,
      "metadata": {
        "id": "vYIsQks7sHg2"
      },
      "outputs": [],
      "source": [
        "class SimpleLossCompute:\n",
        "    '''\n",
        "    A simple loss compute and train function.\n",
        "    '''\n",
        "    def __init__(self, generator, criterion, opt = None):\n",
        "        self.generator = generator\n",
        "        self.criterion = criterion\n",
        "        self.opt = opt\n",
        "        \n",
        "    def __call__(self, x, y, norm):\n",
        "        x = self.generator(x)\n",
        "        \n",
        "        loss = self.criterion(\n",
        "            x.contiguous().view(-1, x.size(-1)), \n",
        "            y.contiguous().view(-1)\n",
        "        )/ norm\n",
        "        \n",
        "        loss.backward()\n",
        "        \n",
        "        if self.opt is not None:\n",
        "            self.opt.step()\n",
        "            self.opt.optimizer.zero_grad()\n",
        "        \n",
        "        return loss.data * norm"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "MVxlMJh3sLRC"
      },
      "source": [
        "### Decodificação gulosa"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FLTQ3xwjmXzT"
      },
      "source": [
        "Este código prevê uma tradução usando decodificação gulosa para simplificar."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 32,
      "metadata": {
        "id": "cj10YjrIsPX-"
      },
      "outputs": [],
      "source": [
        "def greedy_decode(model, src, src_mask, max_len, start_symbol):\n",
        "    memory = model.encode(src, src_mask)\n",
        "    \n",
        "    ys = torch.ones(1, 1).fill_(start_symbol).type_as(src.data)\n",
        "    \n",
        "    for i in range(max_len - 1):\n",
        "        out = model.decode(\n",
        "            memory, src_mask, \n",
        "            Variable(ys), \n",
        "            Variable(subsequent_mask(ys.size(1)).type_as(src.data))\n",
        "        )\n",
        "        \n",
        "        prob = model.generator(out[:, -1])\n",
        "        \n",
        "        _, next_word = torch.max(prob, dim = 1)\n",
        "        next_word = next_word.data[0]\n",
        "        \n",
        "        ys = torch.cat([\n",
        "            ys, \n",
        "            torch.ones(1, 1).type_as(src.data).fill_(next_word)\n",
        "        ], dim = 1)\n",
        "    \n",
        "    return ys"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 33,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "Zhsrf1VHnEPT",
        "outputId": "44dea46f-444d-4e5d-d622-6c855952abc5"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "tensor([[1, 3, 3, 3, 3, 3, 3, 3, 3, 3]])\n"
          ]
        }
      ],
      "source": [
        "temp_model = make_model(V, V, N = 2, d_model = d_model, d_ff = d_ff, h = h, dropout = dropout).to(device)\n",
        "temp_model.eval()\n",
        "temp_src = Variable(torch.LongTensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]]))\n",
        "temp_src_mask = Variable(torch.ones(1, 1, 10))\n",
        "\n",
        "print(greedy_decode(temp_model, temp_src, temp_src_mask, max_len = 10, start_symbol = 1))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "blAgmw1Us_BN"
      },
      "source": [
        "## Exemplo real\n",
        "Agora consideramos um exemplo do mundo real usando a [tarefa de tradução inglês-alemão Multi30k](https://paperswithcode.com/dataset/multi30k-1). Esta tarefa é muito menor do que a tarefa WMT considerada no artigo, mas ilustra todo o sistema.\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 34,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "TvNDOnUosd3T",
        "outputId": "b1cb9134-be8b-4791-a6a1-8aceb72c0ebd"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Requirement already satisfied: torchtext in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (0.9.0)\n",
            "Requirement already satisfied: torch==1.8.0 in c:\\users\\user\\appdata\\roaming\\python\\python39\\site-packages (from torchtext) (1.8.0)\n",
            "Requirement already satisfied: tqdm in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torchtext) (4.63.0)\n",
            "Requirement already satisfied: numpy in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torchtext) (1.22.3)\n",
            "Requirement already satisfied: requests in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torchtext) (2.27.1)\n",
            "Requirement already satisfied: typing-extensions in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torch==1.8.0->torchtext) (4.1.1)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext) (1.26.8)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext) (2.0.12)\n",
            "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext) (3.3)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext) (2021.10.8)\n",
            "Requirement already satisfied: colorama in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from tqdm->torchtext) (0.4.4)\n",
            "Note: you may need to restart the kernel to use updated packages.\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "WARNING: You are using pip version 21.2.4; however, version 22.0.4 is available.\n",
            "You should consider upgrading via the 'C:\\Users\\User\\AppData\\Local\\Programs\\Python\\Python39\\python.exe -m pip install --upgrade pip' command.\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Requirement already satisfied: spacy in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (3.2.3)\n",
            "Requirement already satisfied: typer<0.5.0,>=0.3.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (0.4.0)\n",
            "Requirement already satisfied: wasabi<1.1.0,>=0.8.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (0.9.0)\n",
            "Requirement already satisfied: numpy>=1.15.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (1.22.3)\n",
            "Requirement already satisfied: cymem<2.1.0,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (2.0.6)\n",
            "Requirement already satisfied: pathy>=0.3.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (0.6.1)\n",
            "Requirement already satisfied: langcodes<4.0.0,>=3.2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (3.3.0)\n",
            "Requirement already satisfied: packaging>=20.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (21.3)\n",
            "Requirement already satisfied: thinc<8.1.0,>=8.0.12 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (8.0.15)\n",
            "Requirement already satisfied: requests<3.0.0,>=2.13.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (2.27.1)\n",
            "Requirement already satisfied: preshed<3.1.0,>=3.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (3.0.6)\n",
            "Requirement already satisfied: spacy-loggers<2.0.0,>=1.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (1.0.1)\n",
            "Requirement already satisfied: setuptools in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (58.1.0)\n",
            "Requirement already satisfied: murmurhash<1.1.0,>=0.28.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (1.0.6)\n",
            "Requirement already satisfied: jinja2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (3.0.3)\n",
            "Requirement already satisfied: catalogue<2.1.0,>=2.0.6 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (2.0.6)\n",
            "Requirement already satisfied: pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (1.8.2)\n",
            "Requirement already satisfied: spacy-legacy<3.1.0,>=3.0.8 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (3.0.9)\n",
            "Requirement already satisfied: blis<0.8.0,>=0.4.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (0.7.6)\n",
            "Requirement already satisfied: srsly<3.0.0,>=2.4.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (2.4.2)\n",
            "Requirement already satisfied: tqdm<5.0.0,>=4.38.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy) (4.63.0)\n",
            "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from packaging>=20.0->spacy) (3.0.7)\n",
            "Requirement already satisfied: smart-open<6.0.0,>=5.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pathy>=0.3.5->spacy) (5.2.1)\n",
            "Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4->spacy) (4.1.1)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy) (2021.10.8)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy) (1.26.8)\n",
            "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy) (3.3)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy) (2.0.12)\n",
            "Requirement already satisfied: colorama in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from tqdm<5.0.0,>=4.38.0->spacy) (0.4.4)\n",
            "Requirement already satisfied: click<9.0.0,>=7.1.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from typer<0.5.0,>=0.3.0->spacy) (8.0.4)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from jinja2->spacy) (2.1.0)\n",
            "Note: you may need to restart the kernel to use updated packages.\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "WARNING: You are using pip version 21.2.4; however, version 22.0.4 is available.\n",
            "You should consider upgrading via the 'C:\\Users\\User\\AppData\\Local\\Programs\\Python\\Python39\\python.exe -m pip install --upgrade pip' command.\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "⚠ As of spaCy v3.0, shortcuts like 'en' are deprecated. Please use the full\n",
            "pipeline package name 'en_core_web_sm' instead.\n",
            "Collecting en-core-web-sm==3.2.0\n",
            "  Downloading https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.2.0/en_core_web_sm-3.2.0-py3-none-any.whl (13.9 MB)\n",
            "Requirement already satisfied: spacy<3.3.0,>=3.2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from en-core-web-sm==3.2.0) (3.2.3)\n",
            "Requirement already satisfied: srsly<3.0.0,>=2.4.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.4.2)\n",
            "Requirement already satisfied: tqdm<5.0.0,>=4.38.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (4.63.0)\n",
            "Requirement already satisfied: spacy-loggers<2.0.0,>=1.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (1.0.1)\n",
            "Requirement already satisfied: packaging>=20.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (21.3)\n",
            "Requirement already satisfied: pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (1.8.2)\n",
            "Requirement already satisfied: preshed<3.1.0,>=3.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.0.6)\n",
            "Requirement already satisfied: cymem<2.1.0,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.0.6)\n",
            "Requirement already satisfied: thinc<8.1.0,>=8.0.12 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (8.0.15)\n",
            "Requirement already satisfied: setuptools in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (58.1.0)\n",
            "Requirement already satisfied: wasabi<1.1.0,>=0.8.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (0.9.0)\n",
            "Requirement already satisfied: catalogue<2.1.0,>=2.0.6 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.0.6)\n",
            "Requirement already satisfied: jinja2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.0.3)\n",
            "Requirement already satisfied: spacy-legacy<3.1.0,>=3.0.8 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.0.9)\n",
            "Requirement already satisfied: murmurhash<1.1.0,>=0.28.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (1.0.6)\n",
            "Requirement already satisfied: requests<3.0.0,>=2.13.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.27.1)\n",
            "Requirement already satisfied: langcodes<4.0.0,>=3.2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.3.0)\n",
            "Requirement already satisfied: pathy>=0.3.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (0.6.1)\n",
            "Requirement already satisfied: typer<0.5.0,>=0.3.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (0.4.0)\n",
            "Requirement already satisfied: blis<0.8.0,>=0.4.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (0.7.6)\n",
            "Requirement already satisfied: numpy>=1.15.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (1.22.3)\n",
            "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from packaging>=20.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.0.7)\n",
            "Requirement already satisfied: smart-open<6.0.0,>=5.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pathy>=0.3.5->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (5.2.1)\n",
            "Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (4.1.1)\n",
            "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (3.3)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2021.10.8)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (1.26.8)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.0.12)\n",
            "Requirement already satisfied: colorama in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from tqdm<5.0.0,>=4.38.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (0.4.4)\n",
            "Requirement already satisfied: click<9.0.0,>=7.1.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from typer<0.5.0,>=0.3.0->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (8.0.4)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from jinja2->spacy<3.3.0,>=3.2.0->en-core-web-sm==3.2.0) (2.1.0)\n",
            "✔ Download and installation successful\n",
            "You can now load the package via spacy.load('en_core_web_sm')\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "WARNING: You are using pip version 21.2.4; however, version 22.0.4 is available.\n",
            "You should consider upgrading via the 'C:\\Users\\User\\AppData\\Local\\Programs\\Python\\Python39\\python.exe -m pip install --upgrade pip' command.\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "⚠ As of spaCy v3.0, shortcuts like 'de' are deprecated. Please use the full\n",
            "pipeline package name 'de_core_news_sm' instead.\n",
            "Collecting de-core-news-sm==3.2.0\n",
            "  Downloading https://github.com/explosion/spacy-models/releases/download/de_core_news_sm-3.2.0/de_core_news_sm-3.2.0-py3-none-any.whl (19.1 MB)\n",
            "Requirement already satisfied: spacy<3.3.0,>=3.2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from de-core-news-sm==3.2.0) (3.2.3)\n",
            "Requirement already satisfied: typer<0.5.0,>=0.3.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (0.4.0)\n",
            "Requirement already satisfied: pathy>=0.3.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (0.6.1)\n",
            "Requirement already satisfied: wasabi<1.1.0,>=0.8.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (0.9.0)\n",
            "Requirement already satisfied: langcodes<4.0.0,>=3.2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.3.0)\n",
            "Requirement already satisfied: jinja2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.0.3)\n",
            "Requirement already satisfied: tqdm<5.0.0,>=4.38.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (4.63.0)\n",
            "Requirement already satisfied: preshed<3.1.0,>=3.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.0.6)\n",
            "Requirement already satisfied: numpy>=1.15.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (1.22.3)\n",
            "Requirement already satisfied: setuptools in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (58.1.0)\n",
            "Requirement already satisfied: spacy-loggers<2.0.0,>=1.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (1.0.1)\n",
            "Requirement already satisfied: spacy-legacy<3.1.0,>=3.0.8 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.0.9)\n",
            "Requirement already satisfied: thinc<8.1.0,>=8.0.12 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (8.0.15)\n",
            "Requirement already satisfied: srsly<3.0.0,>=2.4.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.4.2)\n",
            "Requirement already satisfied: requests<3.0.0,>=2.13.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.27.1)\n",
            "Requirement already satisfied: pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (1.8.2)\n",
            "Requirement already satisfied: cymem<2.1.0,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.0.6)\n",
            "Requirement already satisfied: blis<0.8.0,>=0.4.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (0.7.6)\n",
            "Requirement already satisfied: murmurhash<1.1.0,>=0.28.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (1.0.6)\n",
            "Requirement already satisfied: catalogue<2.1.0,>=2.0.6 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.0.6)\n",
            "Requirement already satisfied: packaging>=20.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (21.3)\n",
            "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from packaging>=20.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.0.7)\n",
            "Requirement already satisfied: smart-open<6.0.0,>=5.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pathy>=0.3.5->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (5.2.1)\n",
            "Requirement already satisfied: typing-extensions>=3.7.4.3 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from pydantic!=1.8,!=1.8.1,<1.9.0,>=1.7.4->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (4.1.1)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (1.26.8)\n",
            "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (3.3)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2021.10.8)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests<3.0.0,>=2.13.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.0.12)\n",
            "Requirement already satisfied: colorama in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from tqdm<5.0.0,>=4.38.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (0.4.4)\n",
            "Requirement already satisfied: click<9.0.0,>=7.1.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from typer<0.5.0,>=0.3.0->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (8.0.4)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from jinja2->spacy<3.3.0,>=3.2.0->de-core-news-sm==3.2.0) (2.1.0)\n",
            "✔ Download and installation successful\n",
            "You can now load the package via spacy.load('de_core_news_sm')\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "WARNING: You are using pip version 21.2.4; however, version 22.0.4 is available.\n",
            "You should consider upgrading via the 'C:\\Users\\User\\AppData\\Local\\Programs\\Python\\Python39\\python.exe -m pip install --upgrade pip' command.\n"
          ]
        }
      ],
      "source": [
        "%pip install torchtext --user\n",
        "%pip install spacy --user\n",
        "!python -m spacy download en\n",
        "!python -m spacy download de"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RJh3CfrCtNEM"
      },
      "source": [
        "### Carregamento de dados\n",
        "Vamos carregar o conjunto de dados usando torchtext e spacy para tokenização."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 35,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Requirement already satisfied: torch==1.8.0 in c:\\users\\user\\appdata\\roaming\\python\\python39\\site-packages (1.8.0)\n",
            "Requirement already satisfied: torchtext==0.9.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (0.9.0)\n",
            "Requirement already satisfied: numpy in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torch==1.8.0) (1.22.3)\n",
            "Requirement already satisfied: typing-extensions in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torch==1.8.0) (4.1.1)\n",
            "Requirement already satisfied: requests in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torchtext==0.9.0) (2.27.1)\n",
            "Requirement already satisfied: tqdm in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from torchtext==0.9.0) (4.63.0)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext==0.9.0) (2021.10.8)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext==0.9.0) (2.0.12)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext==0.9.0) (1.26.8)\n",
            "Requirement already satisfied: idna<4,>=2.5 in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from requests->torchtext==0.9.0) (3.3)\n",
            "Requirement already satisfied: colorama in c:\\users\\user\\appdata\\local\\programs\\python\\python39\\lib\\site-packages (from tqdm->torchtext==0.9.0) (0.4.4)\n",
            "Note: you may need to restart the kernel to use updated packages.\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "WARNING: You are using pip version 21.2.4; however, version 22.0.4 is available.\n",
            "You should consider upgrading via the 'C:\\Users\\User\\AppData\\Local\\Programs\\Python\\Python39\\python.exe -m pip install --upgrade pip' command.\n"
          ]
        }
      ],
      "source": [
        "%pip install -U torch==1.8.0 torchtext==0.9.0"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 36,
      "metadata": {
        "id": "w7yiSeNHqQAE"
      },
      "outputs": [],
      "source": [
        "import torchtext\n",
        "from torchtext.legacy import data, datasets\n",
        "    \n",
        "import spacy\n",
        "\n",
        "spacy_de = spacy.load('de_core_news_sm')\n",
        "spacy_en = spacy.load('en_core_web_sm')"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 37,
      "metadata": {
        "id": "9VzInINmtD21"
      },
      "outputs": [],
      "source": [
        "def tokenize_de(text):\n",
        "    return [tok.text for tok in spacy_de.tokenizer(text)]\n",
        "\n",
        "def tokenize_en(text):\n",
        "    return [tok.text for tok in spacy_en.tokenizer(text)]"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 38,
      "metadata": {},
      "outputs": [],
      "source": [
        "BOS_WORD = '<start>'\n",
        "EOS_WORD = '</end>'\n",
        "BLANK_WORD = \"<blank>\"\n",
        "\n",
        "MAX_LEN = 100\n",
        "MIN_FREQ = 2"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Tokenização"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 39,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "o8Y80P4qqT3B",
        "outputId": "0d0b0eb0-86c3-4a31-f83b-17e7b3ded121"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "downloading training.tar.gz\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "c:\\Users\\User\\Desktop\\TCC\\.data\\multi30k\\training.tar.gz: 100%|██████████| 1.21M/1.21M [00:02<00:00, 502kB/s] \n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "downloading validation.tar.gz\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "c:\\Users\\User\\Desktop\\TCC\\.data\\multi30k\\validation.tar.gz: 100%|██████████| 46.3k/46.3k [00:00<00:00, 109kB/s] \n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "downloading mmt_task1_test2016.tar.gz\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "c:\\Users\\User\\Desktop\\TCC\\.data\\multi30k\\mmt_task1_test2016.tar.gz: 100%|██████████| 66.2k/66.2k [00:00<00:00, 109kB/s] \n"
          ]
        }
      ],
      "source": [
        "SRC_DE2EN = torchtext.legacy.data.Field(\n",
        "    tokenize = tokenize_de, \n",
        "    pad_token = BLANK_WORD\n",
        ")\n",
        "\n",
        "TGT_DE2EN = torchtext.legacy.data.Field(\n",
        "    tokenize = tokenize_en, \n",
        "    init_token = BOS_WORD, \n",
        "    eos_token = EOS_WORD, \n",
        "    pad_token = BLANK_WORD\n",
        ")\n",
        "\n",
        "train_DE2EN, val_DE2EN, test_DE2EN = torchtext.legacy.datasets.Multi30k.splits(\n",
        "    exts = ('.de', '.en'), \n",
        "    fields = (SRC_DE2EN, TGT_DE2EN), \n",
        "    filter_pred = lambda x: len(vars(x)['src']) <= MAX_LEN and \n",
        "        len(vars(x)['trg']) <= MAX_LEN\n",
        ")\n",
        "\n",
        "SRC_DE2EN.build_vocab(train_DE2EN.src, min_freq = MIN_FREQ)\n",
        "TGT_DE2EN.build_vocab(train_DE2EN.trg, min_freq = MIN_FREQ)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QtdFYOfoy9ZX"
      },
      "source": [
        "### Iteradores\n",
        "Batching é muito importante para a velocidade. Queremos ter lotes muito bem divididos, com preenchimento (padding) absolutamente mínimo. Para fazer isso, temos que hackear um pouco o lote de torchtext padrão. Este código corrige seu lote padrão para garantir que pesquisamos frases suficientes para encontrar lotes apertados."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 40,
      "metadata": {
        "id": "pMnC27pftVxE"
      },
      "outputs": [],
      "source": [
        "class MyIterator(data.Iterator):\n",
        "    def create_batches(self):\n",
        "\n",
        "        if self.train:\n",
        "            \n",
        "            def pool(d, random_shuffler):\n",
        "                for p in data.batch(d, self.batch_size * 100):\n",
        "                    p_batch = data.batch(\n",
        "                        sorted(p, key = self.sort_key),\n",
        "                        self.batch_size, \n",
        "                        self.batch_size_fn\n",
        "                    )\n",
        "                    \n",
        "                    for b in random_shuffler(list(p_batch)):\n",
        "                        yield b\n",
        "            \n",
        "            self.batches = pool(self.data(), self.random_shuffler)\n",
        "            \n",
        "        else:\n",
        "            self.batches = []\n",
        "            \n",
        "            for b in data.batch(self.data(), self.batch_size, self.batch_size_fn):\n",
        "                self.batches.append(sorted(b, key = self.sort_key))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 41,
      "metadata": {
        "id": "v-jSXVWurIYf"
      },
      "outputs": [],
      "source": [
        "def rebatch(pad_idx, batch):\n",
        "    '''\n",
        "    Fix order in torchtext to match ours\n",
        "    '''\n",
        "    src, trg = batch.src.transpose(0, 1), batch.trg.transpose(0, 1)\n",
        "    \n",
        "    return Batch(src, trg, pad_idx)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Treinamento"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Modelo completo"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 42,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ig-hqRpQzNZ3",
        "outputId": "3d4ecaef-4c61-44b7-ce17-d33d8c1c94f9"
      },
      "outputs": [],
      "source": [
        "def full_model(src, tgt, train, val):\n",
        "\n",
        "    pad_idx = tgt.vocab.stoi[\"<blank>\"]\n",
        "    \n",
        "    model = make_model(len(src.vocab), len(tgt.vocab), \n",
        "        N = N, d_model = d_model, d_ff = d_ff, h = h, dropout = dropout).to(device)\n",
        "\n",
        "    criterion = LabelSmoothing(\n",
        "        size = len(tgt.vocab), \n",
        "        padding_idx = pad_idx, \n",
        "        smoothing = 0.1\n",
        "    ).to(device)\n",
        "\n",
        "    train_iter = MyIterator(\n",
        "        train, \n",
        "        batch_size = BATCH_SIZE, \n",
        "        device = device,\n",
        "        repeat = False, \n",
        "        sort_key = lambda x: (len(x.src), len(x.trg)),\n",
        "        batch_size_fn = batch_size_fn, \n",
        "        train = True\n",
        "    )\n",
        "\n",
        "    valid_iter = MyIterator(\n",
        "        val, \n",
        "        batch_size = BATCH_SIZE, \n",
        "        device = device,\n",
        "        repeat = False, \n",
        "        sort_key = lambda x: (len(x.src), len(x.trg)),\n",
        "        batch_size_fn = batch_size_fn, \n",
        "        train = False\n",
        "    )\n",
        "\n",
        "    return pad_idx, model, criterion, train_iter, valid_iter"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Fluxo de treinamento"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 43,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ZnGiHrs_y-m-",
        "outputId": "43adf534-7c71-42e6-a6f1-9c62e044912d"
      },
      "outputs": [],
      "source": [
        "def train_model(pad_idx, model, criterion, train_iter, valid_iter):\n",
        "\n",
        "    model_opt = NoamOpt(\n",
        "        model.src_embed[0].d_model, \n",
        "        factor,\n",
        "        warmup_steps,\n",
        "        torch.optim.Adam(model.parameters(), lr = lr, betas = betas, eps = eps)\n",
        "    )\n",
        "\n",
        "    for epoch in range(n_epochs):\n",
        "        model.train()\n",
        "\n",
        "        run_epoch(\n",
        "            (rebatch(pad_idx, b) for b in train_iter), \n",
        "            model, \n",
        "            SimpleLossCompute(model.generator, criterion, opt = model_opt)\n",
        "        )\n",
        "\n",
        "        model.eval()\n",
        "        \n",
        "        loss = run_epoch(\n",
        "            (rebatch(pad_idx, b) for b in valid_iter), \n",
        "            model, \n",
        "            SimpleLossCompute(model.generator, criterion, opt = None)\n",
        "        )\n",
        "        \n",
        "        print(loss)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Teste"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 44,
      "metadata": {},
      "outputs": [
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "C:\\Users\\User\\AppData\\Roaming\\Python\\Python39\\site-packages\\torch\\nn\\_reduction.py:42: UserWarning: size_average and reduce args will be deprecated, please use reduction='sum' instead.\n",
            "  warnings.warn(warning.format(ret))\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Epoch Step: 0 Loss: 7.673217 Tokens per Sec: 287.623596\n",
            "Epoch Step: 200 Loss: 4.887909 Tokens per Sec: 359.071991\n",
            "Epoch Step: 400 Loss: 3.645025 Tokens per Sec: 360.434906\n",
            "Epoch Step: 600 Loss: 3.232066 Tokens per Sec: 365.299255\n",
            "Epoch Step: 800 Loss: 3.351522 Tokens per Sec: 368.099121\n",
            "Epoch Step: 1000 Loss: 2.774139 Tokens per Sec: 360.030273\n",
            "Epoch Step: 1200 Loss: 3.012085 Tokens per Sec: 360.965271\n",
            "Epoch Step: 1400 Loss: 2.844360 Tokens per Sec: 360.415527\n",
            "Epoch Step: 1600 Loss: 2.689275 Tokens per Sec: 357.222351\n",
            "Epoch Step: 0 Loss: 2.284020 Tokens per Sec: 449.474609\n",
            "tensor(2.6080)\n",
            "Epoch Step: 0 Loss: 3.301101 Tokens per Sec: 313.296844\n",
            "Epoch Step: 200 Loss: 2.502338 Tokens per Sec: 354.649628\n",
            "Epoch Step: 400 Loss: 2.203788 Tokens per Sec: 358.586029\n",
            "Epoch Step: 600 Loss: 2.289674 Tokens per Sec: 356.617767\n",
            "Epoch Step: 800 Loss: 2.725626 Tokens per Sec: 360.745178\n",
            "Epoch Step: 1000 Loss: 2.878083 Tokens per Sec: 355.031677\n",
            "Epoch Step: 1200 Loss: 2.119563 Tokens per Sec: 360.488037\n",
            "Epoch Step: 1400 Loss: 2.813358 Tokens per Sec: 357.958954\n",
            "Epoch Step: 1600 Loss: 1.985573 Tokens per Sec: 358.698761\n",
            "Epoch Step: 0 Loss: 1.657634 Tokens per Sec: 445.633972\n",
            "tensor(2.1852)\n"
          ]
        }
      ],
      "source": [
        "pad_idx_DE2EN, model_DE2EN, criterion_DE2EN, train_iter_DE2EN, valid_iter_DE2EN = full_model(SRC_DE2EN, TGT_DE2EN, train_DE2EN, val_DE2EN)\n",
        "train_model(pad_idx_DE2EN, model_DE2EN, criterion_DE2EN, train_iter_DE2EN, valid_iter_DE2EN)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Tradução"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 45,
      "metadata": {},
      "outputs": [],
      "source": [
        "def translation(valid_iter, model, SRC, TGT):\n",
        "\n",
        "    bleu = []\n",
        "    \n",
        "    for i, batch in enumerate(valid_iter):\n",
        "\n",
        "        src = batch.src.transpose(0, 1)[:1]\n",
        "        \n",
        "        src_mask = (src != SRC.vocab.stoi[\"<blank>\"]).unsqueeze(-2)\n",
        "        \n",
        "        out = greedy_decode(\n",
        "            model, \n",
        "            src, \n",
        "            src_mask, \n",
        "            max_len = 60, \n",
        "            start_symbol = TGT.vocab.stoi[\"<start>\"]\n",
        "        )\n",
        "\n",
        "        print(\"Translation:\", end = \"\\t\")\n",
        "\n",
        "        hypothesis = ''\n",
        "\n",
        "        for i in range(1, out.size(1)):\n",
        "            trans = TGT.vocab.itos[out[0, i]]\n",
        "            \n",
        "            if trans == \"</end>\": break\n",
        "            \n",
        "            print(trans, end = \" \")\n",
        "            \n",
        "            hypothesis += trans\n",
        "            hypothesis += ' '\n",
        "\n",
        "        \n",
        "        print(\"\\nTarget:\", end = \"\\t\")\n",
        "\n",
        "        reference = ''\n",
        "\n",
        "        for i in range(1, batch.trg.size(0)):\n",
        "            target = TGT.vocab.itos[batch.trg.data[i, 0]]\n",
        "            \n",
        "            if target == \"</end>\": break\n",
        "            print(target, end = \" \")\n",
        "\n",
        "            reference += target\n",
        "            reference += ' '\n",
        "\n",
        "        references = [reference]\n",
        "\n",
        "\n",
        "        sent_bleu = sentence_bleu(references, hypothesis)\n",
        "        bleu.append(sent_bleu)\n",
        "\n",
        "        print(\"\\nBLEU score = \" + str(sent_bleu) + \"\\n\")\n",
        "\n",
        "    print(\"Mean BLEU score = \" + str(mean(bleu)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Alemão para Inglês"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 46,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Translation:\tTwo police officers are <unk> their <unk> . \n",
            "Target:\tFarmers are performing their <unk> during the day . \n",
            "BLEU score = 0.41754201443818345\n",
            "\n",
            "Translation:\tA few people stand on a <unk> . \n",
            "Target:\tSeveral people standing on a subway platform . \n",
            "BLEU score = 0.3744643450102179\n",
            "\n",
            "Translation:\tChildren are talking to a machine . \n",
            "Target:\tKids conversing and learning in class . \n",
            "BLEU score = 0.13484538548234365\n",
            "\n",
            "Translation:\tTwo men are walking down a dirt road . \n",
            "Target:\tTwo men walking down a dirt path . \n",
            "BLEU score = 0.7281555816542453\n",
            "\n",
            "Translation:\tPeople are on the street in a city street . \n",
            "Target:\tPeople on the street at a block party . \n",
            "BLEU score = 0.5800274664320602\n",
            "\n",
            "Translation:\tA woman is standing in front of a trees . \n",
            "Target:\tA woman standing in front of trees and smiling . \n",
            "BLEU score = 0.718237218225714\n",
            "\n",
            "Translation:\tA man is eating a restaurant . \n",
            "Target:\tA man in a restaurant having lunch . \n",
            "BLEU score = 0.6019536773275824\n",
            "\n",
            "Translation:\tA group of people are walking by a wooden structure . \n",
            "Target:\tA group of people running a <unk> marathon . \n",
            "BLEU score = 0.45605114189661233\n",
            "\n",
            "Translation:\tThis man is wearing a white hat is gathered . \n",
            "Target:\tThis man is wearing a white cap or hat . \n",
            "BLEU score = 0.7136294587034953\n",
            "\n",
            "Translation:\tA man is holding a camera while holding a camera . \n",
            "Target:\tA man holding a drinking glass at the camera . \n",
            "BLEU score = 0.5471447830538266\n",
            "\n",
            "Translation:\tTwo dogs are lying in the snow with his mouth . \n",
            "Target:\tTwo dogs are laying in the snow with their mouths open . \n",
            "BLEU score = 0.7063890726406052\n",
            "\n",
            "Translation:\tFour people are gathered in a public area . \n",
            "Target:\t4 people are strapped to an amusement park ride \n",
            "BLEU score = 0.34711324842911034\n",
            "\n",
            "Translation:\tA boy is showing a boy on the back . \n",
            "Target:\tOne boy <unk> another boy up on his back . \n",
            "BLEU score = 0.4555264577424794\n",
            "\n",
            "Translation:\tThree men are walking on a street in the street . \n",
            "Target:\tThree men are walking on a road in the mountains . \n",
            "BLEU score = 0.6867327904066065\n",
            "\n",
            "Translation:\tA man is doing a nap in a fire . \n",
            "Target:\tA man doing some type of public show involving fire . \n",
            "BLEU score = 0.30971139712230483\n",
            "\n",
            "Translation:\tMany Asian children are gathered under a train . \n",
            "Target:\tMany asian children made a train under a <unk> <unk> sign . \n",
            "BLEU score = 0.5505351505255455\n",
            "\n",
            "Translation:\tA woman with blond - hair is drinking out of a large large large window . \n",
            "Target:\tA woman with blond - hair is drinking out of a glass . \n",
            "BLEU score = 0.6605173931324406\n",
            "\n",
            "Translation:\tA man is taking a picture of a lake with trees in the background . \n",
            "Target:\tA man takes a photograph of a lake with mountains in the background . \n",
            "BLEU score = 0.6824546854487492\n",
            "\n",
            "Translation:\tA black and white dog swimming dog swimming in the water . \n",
            "Target:\tA black and white dog swimming in clear water . \n",
            "BLEU score = 0.6986987947556846\n",
            "\n",
            "Translation:\tA very brown dog is putting on the water . \n",
            "Target:\tA very wet brown dog is emerging from the water . \n",
            "BLEU score = 0.6532358806866065\n",
            "\n",
            "Translation:\tA <unk> is being <unk> by a person . \n",
            "Target:\tA <unk> car with many firefighters cutting into the car . \n",
            "BLEU score = 0.21131294241067478\n",
            "\n",
            "Translation:\tFour people are standing outside outside and holding a <unk> . \n",
            "Target:\tFour people in casual clothing are standing outside holding garbage bags . \n",
            "BLEU score = 0.5749007842822687\n",
            "\n",
            "Translation:\tA group of friends are sitting in a house . \n",
            "Target:\tA group of three friends are conversing inside of a home . \n",
            "BLEU score = 0.5323388190049984\n",
            "\n",
            "Translation:\tA man is performing a break in a park . \n",
            "Target:\tA man is leading two small <unk> on a walk at a park . \n",
            "BLEU score = 0.390252596504697\n",
            "\n",
            "Translation:\tPeople are gathered around and <unk> are gathered in the <unk> . \n",
            "Target:\tPeople <unk> and gardening the areas around the walkway . \n",
            "BLEU score = 0.529536295181678\n",
            "\n",
            "Translation:\tA young man is doing a soccer ball . \n",
            "Target:\tA young man gets ready to kick a soccer ball . \n",
            "BLEU score = 0.5892964115758363\n",
            "\n",
            "Translation:\tA man and woman are standing on the street . \n",
            "Target:\tMan and woman in black stand close together on the street . \n",
            "BLEU score = 0.5475134827664\n",
            "\n",
            "Translation:\tThree men are talking on the street in front of a city street . \n",
            "Target:\tThree men are talking in front of old stairs on the street . \n",
            "BLEU score = 0.7878477958527111\n",
            "\n",
            "Translation:\tA woman and a child are enjoying a child . \n",
            "Target:\tA woman and a young child has fun while playing a board game . \n",
            "BLEU score = 0.4210530692518689\n",
            "\n",
            "Translation:\tA man and woman sitting on the ground . \n",
            "Target:\tA man and a woman are sitting on the floor in front of luggage . \n",
            "BLEU score = 0.42415581369621125\n",
            "\n",
            "Translation:\tA child is at a lake at a lake . \n",
            "Target:\tA child that is on the bank of a lake pointing something out in the sky . \n",
            "BLEU score = 0.17118895083537514\n",
            "\n",
            "Translation:\tA large dog running on a fence . \n",
            "Target:\tA large black dog runs along a fence in the grass . \n",
            "BLEU score = 0.4220722102814102\n",
            "\n",
            "Translation:\tThe little brown dog is on the side of the mountains . \n",
            "Target:\tThe small brown dog is amidst the potted plants and fallen leaves . \n",
            "BLEU score = 0.3865507227568219\n",
            "\n",
            "Translation:\tA man stands on a boat near a boat . \n",
            "Target:\tMan standing on a pier next to a boat on a still lake . \n",
            "BLEU score = 0.3561474293019452\n",
            "\n",
            "Translation:\tA man in a large blue <unk> is walking by a large building . \n",
            "Target:\tA man in a long blue cape walks into an old stone building . \n",
            "BLEU score = 0.5816211823994077\n",
            "\n",
            "Translation:\tA boy in a sweatshirt is throwing a baby in a pool . \n",
            "Target:\tA boy in a hoodie is throwing an object into a dirty swimming pool . \n",
            "BLEU score = 0.5061423447434705\n",
            "\n",
            "Translation:\tThree children are watching a man playing a guitar . \n",
            "Target:\tThree children watch a man on stage playing guitar . \n",
            "BLEU score = 0.7757873472289065\n",
            "\n",
            "Translation:\tA group of people are dancing at a restaurant . \n",
            "Target:\tGroup of people dancing at a club or party in a costume . \n",
            "BLEU score = 0.5408588906904415\n",
            "\n",
            "Translation:\tA child in a green shirt is holding a red ball . \n",
            "Target:\tA child with a green shirt is holding a red ball against its mouth . \n",
            "BLEU score = 0.6170433174037093\n",
            "\n",
            "Translation:\tA man in a white t - shirt and white shirt and white t - shirt and a tree . \n",
            "Target:\tA man in a white shirt and khaki pants crouches on a fallen tree trunk . \n",
            "BLEU score = 0.49321948078296235\n",
            "\n",
            "Translation:\tA group of young adults are playing a game . \n",
            "Target:\tA bunch of young adults stare in <unk> at their computer monitors as they competitively game . \n",
            "BLEU score = 0.21235504273740183\n",
            "\n",
            "Translation:\tA man is laying in his room eating something . \n",
            "Target:\tA man lounges in his room without eating for <unk> . \n",
            "BLEU score = 0.508038984512192\n",
            "\n",
            "Translation:\tA couple is standing at the subway and waiting to be waiting at something . \n",
            "Target:\tA well built black man stands in the subway , listening to headphones . \n",
            "BLEU score = 0.38186208810377026\n",
            "\n",
            "Translation:\tTwo men are watching two men are watching a fence . \n",
            "Target:\tTwo men are observing another as he puts the finishing touches on wet cement . \n",
            "BLEU score = 0.29637966941825106\n",
            "\n",
            "Translation:\tA man in a black shirt is eating a kitchen . \n",
            "Target:\tA man in a black shirt is cooking out of a <unk> in a cluttered kitchen . \n",
            "BLEU score = 0.4580963652262192\n",
            "\n",
            "Translation:\tA man in a bright red dress stands on a hill . \n",
            "Target:\tAn old man dressed in <unk> - evil clothing holds an axe while standing on a hill . \n",
            "BLEU score = 0.29613559980769705\n",
            "\n",
            "Translation:\tA small girl in a woman in a chair in a classroom . \n",
            "Target:\tA little blond girl with a polka dot shirt is giving a stuffed animal a \" bath \" in a sink . \n",
            "BLEU score = 0.15652695373543973\n",
            "\n",
            "Translation:\tTwo men are using a basketball game while a basketball game . \n",
            "Target:\tTwo men guard the man with the basketball during a game at dusk . \n",
            "BLEU score = 0.49345479732750586\n",
            "\n",
            "Translation:\tA child in an orange shirt is jumping by a wooden table . \n",
            "Target:\tA child in an orange shirt jumps off bales of hay while other children watch . \n",
            "BLEU score = 0.4497606805696045\n",
            "\n",
            "Translation:\tA black dog jumping through the water wearing a stick . \n",
            "Target:\tA black dog is jumping through some water , while holding a stick in his mouth . \n",
            "BLEU score = 0.5288392272549498\n",
            "\n",
            "Translation:\tA woman in black is kneeling on a fence with a large fence . \n",
            "Target:\tA warmly dressed woman in black kneels with a small tan dog near a crowd of onlookers . \n",
            "BLEU score = 0.35786538513003047\n",
            "\n",
            "Translation:\t<unk> <unk> <unk> a Frisbee are being <unk> by a race . \n",
            "Target:\tA frisbee is being thrown to the girl while the other girl appears to be asking for it . \n",
            "BLEU score = 0.163913112779124\n",
            "\n",
            "Translation:\tPeople are sitting at a table while sitting in a room . \n",
            "Target:\tSeveral children are raising their hands while sitting on a colorful rug in a classroom . \n",
            "BLEU score = 0.33330398912544856\n",
            "\n",
            "Translation:\tA woman is throwing a Frisbee in a large Frisbee . \n",
            "Target:\tA woman <unk> a Frisbee to a dog as a baseball player in uniform runs behind her . \n",
            "BLEU score = 0.2704589845293785\n",
            "\n",
            "Translation:\tThere are many man in a man in a classroom and a man in the snow . \n",
            "Target:\tThere is a man wrapped in a blanket of some sort sliding down a hill that is covered in snow . \n",
            "BLEU score = 0.31680717592082364\n",
            "\n",
            "Translation:\tA man and woman sitting on a bench are sitting on a bench . \n",
            "Target:\tA man and woman sit on a bench , watching a boat go by . \n",
            "BLEU score = 0.6119843384057435\n",
            "\n",
            "Translation:\tA boy in a striped shirt and orange shirt is holding a large flag . \n",
            "Target:\tA boy in a t - shirt and shorts is holding a snowball and facing a snowy mountain . \n",
            "BLEU score = 0.4807557649449535\n",
            "\n",
            "Translation:\tA person in a red <unk> on a wall in front of a building . \n",
            "Target:\tA person wearing a red long - sleeved shirt is lying down on a wall in front of a lamp post in a very unusual manner . \n",
            "BLEU score = 0.27806453337944603\n",
            "\n",
            "Translation:\tA man and woman walking by a tree with several people in a baby . \n",
            "Target:\tA man and woman pushing strollers are walking by some people who are selling items in tents . \n",
            "BLEU score = 0.4179773553781119\n",
            "\n",
            "Translation:\tA very small child with a child sitting on the ground near a dock , and <unk> . \n",
            "Target:\tA very young child is sitting in the sink with paint on his body and face and is playing with the kitchen <unk> . \n",
            "BLEU score = 0.3809470171146004\n",
            "\n",
            "Translation:\tTwo men are gathered around a sunny day in a public area . \n",
            "Target:\tTwo men are out on a bright , sunny day attempting to catch some fish on the lake . \n",
            "BLEU score = 0.32804211875184425\n",
            "\n",
            "Translation:\tYoung adults , in a <unk> , wearing a purple striped striped striped striped striped and blue striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped striped \n",
            "Target:\tYoung adults wearing red , yellow , and black shirts in a straight line perform a sketch . \n",
            "BLEU score = 0.09475172537353811\n",
            "\n",
            "Translation:\tA large dog jumping into a large man in a black room . \n",
            "Target:\tA large dog is jumping in a fountain while a man in a black T - shirt and vest stands nearby . \n",
            "BLEU score = 0.35192290566718426\n",
            "\n",
            "Translation:\tA woman in a pink shirt is looking at a man in a room . \n",
            "Target:\tA woman wearing a pink shirt showing a man with a striped sweater how to do some work with yarn . \n",
            "BLEU score = 0.2900750217788543\n",
            "\n",
            "Translation:\tAn older man is sitting outside outside of a large building with a large building . \n",
            "Target:\tAn older man is sitting outside on a bench in front a large banner that says , \" <unk> <unk> <unk> <unk> . \" \n",
            "BLEU score = 0.42838999104763237\n",
            "\n",
            "Translation:\tA dark - haired woman wearing a blue shirt and a young boy with her hair . \n",
            "Target:\tA dark - haired woman wearing a blue blouse and sunglasses holds a banner while a young girl with braided hair writes on the banner . \n",
            "BLEU score = 0.3777018460802431\n",
            "\n",
            "Translation:\tTwo people hold a large <unk> <unk> with a large <unk> and <unk> . \n",
            "Target:\tTwo people are holding a large upside - down earth <unk> , about 4 ' in <unk> , and a child appears to be jumping over <unk> . \n",
            "BLEU score = 0.29812314378756366\n",
            "\n",
            "Translation:\tA girl is sitting in a public area holding a book and holding a book . \n",
            "Target:\tSitting casually in a public place , a girl reads holding the book open with her hand on which is a butterfly ring . \n",
            "BLEU score = 0.3507583747572476\n",
            "\n",
            "Mean BLEU score = 0.45286908856927954\n"
          ]
        }
      ],
      "source": [
        "translation(valid_iter_DE2EN, model_DE2EN, SRC_DE2EN, TGT_DE2EN)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### Inglês para Alemão"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 47,
      "metadata": {},
      "outputs": [],
      "source": [
        "SRC_EN2DE = torchtext.legacy.data.Field(\n",
        "    tokenize = tokenize_en, \n",
        "    pad_token = BLANK_WORD\n",
        ")\n",
        "\n",
        "TGT_EN2DE = torchtext.legacy.data.Field(\n",
        "    tokenize = tokenize_de, \n",
        "    init_token = BOS_WORD, \n",
        "    eos_token = EOS_WORD, \n",
        "    pad_token = BLANK_WORD\n",
        ")\n",
        "\n",
        "train_EN2DE, val_EN2DE, test_EN2DE = torchtext.legacy.datasets.Multi30k.splits(\n",
        "    exts = ('.en', '.de'), \n",
        "    fields = (SRC_EN2DE, TGT_EN2DE), \n",
        "    filter_pred = lambda x: len(vars(x)['src']) <= MAX_LEN and \n",
        "        len(vars(x)['trg']) <= MAX_LEN\n",
        ")\n",
        "\n",
        "SRC_EN2DE.build_vocab(train_EN2DE.src, min_freq = MIN_FREQ)\n",
        "TGT_EN2DE.build_vocab(train_EN2DE.trg, min_freq = MIN_FREQ)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 48,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Epoch Step: 0 Loss: 7.775187 Tokens per Sec: 302.325012\n",
            "Epoch Step: 200 Loss: 5.813664 Tokens per Sec: 346.148651\n",
            "Epoch Step: 400 Loss: 3.930289 Tokens per Sec: 347.672791\n",
            "Epoch Step: 600 Loss: 3.825217 Tokens per Sec: 349.095062\n",
            "Epoch Step: 800 Loss: 3.435733 Tokens per Sec: 350.259705\n",
            "Epoch Step: 1000 Loss: 3.430660 Tokens per Sec: 347.187836\n",
            "Epoch Step: 1200 Loss: 3.311951 Tokens per Sec: 346.352875\n",
            "Epoch Step: 1400 Loss: 3.269810 Tokens per Sec: 341.160126\n",
            "Epoch Step: 1600 Loss: 2.000238 Tokens per Sec: 333.685425\n",
            "Epoch Step: 0 Loss: 2.041561 Tokens per Sec: 466.892151\n",
            "tensor(2.6087)\n",
            "Epoch Step: 0 Loss: 2.813402 Tokens per Sec: 312.126953\n",
            "Epoch Step: 200 Loss: 2.181512 Tokens per Sec: 340.518799\n",
            "Epoch Step: 400 Loss: 2.433208 Tokens per Sec: 343.391052\n",
            "Epoch Step: 600 Loss: 2.079488 Tokens per Sec: 339.565704\n",
            "Epoch Step: 800 Loss: 2.560645 Tokens per Sec: 342.445709\n",
            "Epoch Step: 1000 Loss: 1.610133 Tokens per Sec: 341.861084\n",
            "Epoch Step: 1200 Loss: 3.236626 Tokens per Sec: 343.106506\n",
            "Epoch Step: 1400 Loss: 2.809208 Tokens per Sec: 339.667450\n",
            "Epoch Step: 1600 Loss: 2.956203 Tokens per Sec: 340.494537\n",
            "Epoch Step: 0 Loss: 1.534855 Tokens per Sec: 495.915161\n",
            "tensor(2.1273)\n"
          ]
        }
      ],
      "source": [
        "pad_idx_EN2DE, model_EN2DE, criterion_EN2DE, train_iter_EN2DE, valid_iter_EN2DE = full_model(SRC_EN2DE, TGT_EN2DE, train_EN2DE, val_EN2DE)\n",
        "train_model(pad_idx_EN2DE, model_EN2DE, criterion_EN2DE, train_iter_EN2DE, valid_iter_EN2DE)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 49,
      "metadata": {},
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Translation:\tEin Junge springt ins Wasser . \n",
            "Target:\tEin kleiner Junge springt in das Wasser . \n",
            "BLEU score = 0.644170720445744\n",
            "\n",
            "Translation:\tKinder halten sich auf dem Weg . \n",
            "Target:\tZwei Kinder graben Löcher in die Erde . \n",
            "BLEU score = 0.24840959555301853\n",
            "\n",
            "Translation:\tJunge in der Mitte . \n",
            "Target:\tJungen treten in einer <unk> gegeneinander an . \n",
            "BLEU score = 0.1356230050172979\n",
            "\n",
            "Translation:\tKinder werden um einen <unk> herum , das in der Luft . \n",
            "Target:\tKinder werden in einem <unk> <unk> . \n",
            "BLEU score = 0.4725637436859827\n",
            "\n",
            "Translation:\tZwei Männer stehen draußen auf einem <unk> . \n",
            "Target:\tZwei Männer stehen an Telefonzellen im Freien . \n",
            "BLEU score = 0.5059780268428081\n",
            "\n",
            "Translation:\tEin kleines Mädchen geht an einem kleinen Weg entlang . \n",
            "Target:\tEin kleines Mädchen geht an einem kleinen Bach entlang \n",
            "BLEU score = 0.879962308706789\n",
            "\n",
            "Translation:\tEin Mann , der ein anderer Mann und ein anderer gekleidet , spielt . \n",
            "Target:\tEin Mann greift bei einem Footballspiel einen anderen Mann an . \n",
            "BLEU score = 0.47546688515161956\n",
            "\n",
            "Translation:\tEine <unk> Frau arbeitet an einem <unk> . \n",
            "Target:\tEine Ureinwohnerin arbeitet an einer Handarbeit . \n",
            "BLEU score = 0.43210787870967254\n",
            "\n",
            "Translation:\tDiese <unk> , die einen roten roten und roten Shorts trägt , <unk> . \n",
            "Target:\tDieses Footballteam trägt rote Trikots und rote Helme . \n",
            "BLEU score = 0.36110870574919174\n",
            "\n",
            "Translation:\tEin Mann hält ein <unk> an der Kamera . \n",
            "Target:\tEin Mann hält der Kamera ein <unk> entgegen . \n",
            "BLEU score = 0.7541746547365834\n",
            "\n",
            "Translation:\tZwei Mädchen spielen auf einem Baum . \n",
            "Target:\tZwei Mädchen spielen mit ihrem Hund auf einem Baum . \n",
            "BLEU score = 0.6593573131391627\n",
            "\n",
            "Translation:\tEin kleines Mädchen mit einem roten Eimer und ein Mädchen trägt ein <unk> . \n",
            "Target:\tEin kleines asiatisches Mädchen mit schmutzigem Gesicht trägt ein Kissen . \n",
            "BLEU score = 0.5197608402198394\n",
            "\n",
            "Translation:\tArbeiter stehen auf einem Stuhl . \n",
            "Target:\tBauarbeiter stehen auf einer Maschine \n",
            "BLEU score = 0.6117012009078837\n",
            "\n",
            "Translation:\tEin Bauarbeiter in einer orangefarbenen Weste , die in einer orangefarbenen <unk> . \n",
            "Target:\tEin Bauarbeiter in einer orangen Weste verlegt <unk> . \n",
            "BLEU score = 0.5489001502926218\n",
            "\n",
            "Translation:\tEine Frau und zwei Jungen schauen sich an einem Stand an . \n",
            "Target:\tEine Frau und zwei Jungen blicken auf einen <unk> . \n",
            "BLEU score = 0.5733401818074073\n",
            "\n",
            "Translation:\tEin Junge mit einem Hut sitzt auf dem Kopf . \n",
            "Target:\tEin Junge mit Kopfhörern sitzt auf den Schultern einer Frau . \n",
            "BLEU score = 0.5107191131571326\n",
            "\n",
            "Translation:\tEine Gruppe von Leuten geht in der Mitte . \n",
            "Target:\tEine Gruppe von Menschen läuft bei einem Marathon im Winter . \n",
            "BLEU score = 0.3403432964255604\n",
            "\n",
            "Translation:\tEin Mann singt in ein Mikrofon und hält ein Mikrofon . \n",
            "Target:\tEin Mann singt in ein Mikrofon und hält dabei eine Gitarre . \n",
            "BLEU score = 0.7176580438295255\n",
            "\n",
            "Translation:\tEin kleiner Junge sitzt mit seinem <unk> . \n",
            "Target:\tEin kleiner indischer Junge sitzt da und denkt über seine Zukunft nach . \n",
            "BLEU score = 0.3625783259953076\n",
            "\n",
            "Translation:\tKind in Kostümen , die sich . \n",
            "Target:\tEin als Spiderman verkleidetes Kind <unk> an der Tür . \n",
            "BLEU score = 0.11051235739752421\n",
            "\n",
            "Translation:\tEin Kind steht an einem Kind und hält ein Spielzeug . \n",
            "Target:\tEin Kind steht neben einem Kinderwagen und hält ein <unk> . \n",
            "BLEU score = 0.6608629548896526\n",
            "\n",
            "Translation:\tEin kleiner Spieler wird von einem Ball . \n",
            "Target:\tEin junger Footballspieler bringt sich für ein <unk> in <unk> . \n",
            "BLEU score = 0.21599419825090801\n",
            "\n",
            "Translation:\tEin schwarzer Personen wird in der Mitte , das an . \n",
            "Target:\tSchwarze Menschen sind in einem Lebensmittelgeschäft , um Essen zu kaufen . \n",
            "BLEU score = 0.24368413081004306\n",
            "\n",
            "Translation:\tEin kleines Mädchen rennt in die Luft . \n",
            "Target:\tEin kleines asiatisches Mädchen rennt entlang einer gepflasterten Straße in den <unk> . \n",
            "BLEU score = 0.23927498534196914\n",
            "\n",
            "Translation:\tEine Frau , die eine große Kamera , und steht in der Nähe . \n",
            "Target:\tEine Frau fotografiert mit einer großen Kamera , während sie auf der Straße steht . \n",
            "BLEU score = 0.45361020186980106\n",
            "\n",
            "Translation:\tEin Mann wartet auf die <unk> . \n",
            "Target:\tEin Mann wartet am Bahnsteig auf den Zug . \n",
            "BLEU score = 0.5020863789382245\n",
            "\n",
            "Translation:\tPersonen , die an einem Marathon Mann , der ein Mann <unk> . \n",
            "Target:\tBei einer Art Jahrmarkt stellt ein Mann Zuckerwatte her . \n",
            "BLEU score = 0.29575164454446873\n",
            "\n",
            "Translation:\tEine Frau in einem grünen Hemd geht auf einem belebten Platz . \n",
            "Target:\tEine Frau in einem grünen Shirt geht eine belebte Straße entlang . \n",
            "BLEU score = 0.6406054085483649\n",
            "\n",
            "Translation:\tEine große Menschenmenge , die ein paar Personen im Hintergrund ist . \n",
            "Target:\tEine große Menschenmenge sieht mit großen Gebäuden im Hintergrund bei etwas zu . \n",
            "BLEU score = 0.5761581962715584\n",
            "\n",
            "Translation:\tEin kleines Kind und eine Frau stehen um einen <unk> herum . \n",
            "Target:\tEin kleines Kind und eine Frau stehen um einen Hahn im Käfig herum . \n",
            "BLEU score = 0.7841760332776697\n",
            "\n",
            "Translation:\tEin Mann , die eine Frau , während eine Frau an . \n",
            "Target:\tEin Mann schiebt Blumen , während eine Frau eine Geste in seine Richtung macht . \n",
            "BLEU score = 0.39033469584544567\n",
            "\n",
            "Translation:\tDer Mann telefoniert vor einem <unk> . \n",
            "Target:\tDer Mann telefoniert vor einem Sportgeschäft mit seinem Handy . \n",
            "BLEU score = 0.43766550777691166\n",
            "\n",
            "Translation:\tEine Frau und ein kleines Kind halten ein <unk> . \n",
            "Target:\tEine Frau und ein kleines Kind haben Spaß bei einem Brettspiel . \n",
            "BLEU score = 0.5675603858117972\n",
            "\n",
            "Translation:\tEine Gruppe Leute wartet hinter einem <unk> und blickt auf die <unk> . \n",
            "Target:\tEine Gruppe von Menschen wartet hinter einer Absperrung und blickt nach rechts . \n",
            "BLEU score = 0.5272961753882396\n",
            "\n",
            "Translation:\tEin Kind und ein Bild von einer Gruppe von ihnen . \n",
            "Target:\tEin Sohn und seine Eltern nehmen ein Gruppenfoto in einer Kirche auf . \n",
            "BLEU score = 0.3648352781297289\n",
            "\n",
            "Translation:\tEin <unk> in schwarzer Kleidung steht in der Mitte , während ein Mann in der Hand ist . \n",
            "Target:\tEin gut <unk> Schwarzer steht in der U-Bahn und hört dabei etwas über Kopfhörer . \n",
            "BLEU score = 0.4316588107349429\n",
            "\n",
            "Translation:\tZwei Männer halten ein <unk> auf den Basketball in der Nähe eines <unk> . \n",
            "Target:\tZwei Männer decken den Mann mit dem Basketball während eines Spiels in der Dämmerung ab . \n",
            "BLEU score = 0.4677962292625119\n",
            "\n",
            "Translation:\tDrei Kinder stehen neben einem <unk> und einem blauen Trikot . \n",
            "Target:\tDrei Kinder stehen nahe beieinander an einem hohen blauen Holzpfahl . \n",
            "BLEU score = 0.5420315800137393\n",
            "\n",
            "Translation:\tEin Junge in einem <unk> wirft einen Gegenstand in einem Pool . \n",
            "Target:\tEin Junge in einem Kapuzenpullover wirft einen Gegenstand in einen schmutzigen Swimmingpool . \n",
            "BLEU score = 0.5235856016860171\n",
            "\n",
            "Translation:\tZwei Kinder halten ihre <unk> in der Hand und eine <unk> . \n",
            "Target:\tZwei Kinder stecken ihre Gesichter in das Bild eines <unk> und einer <unk> Dame . \n",
            "BLEU score = 0.4997413326025607\n",
            "\n",
            "Translation:\tEin Mann hält ein Kind auf seinem Weg . \n",
            "Target:\tEin Mann trägt ein Kind auf seinen Schultern , während um ihn herum Menschen gehen . \n",
            "BLEU score = 0.2503039910919944\n",
            "\n",
            "Translation:\tEin Mädchen liest ein Buch mit ihrem <unk> auf ihrem Schreibtisch . \n",
            "Target:\tEin Teenager mit Notizen auf seinem Tisch liest ein Buch im Klassenzimmer und sieht dabei gelangweilt aus . \n",
            "BLEU score = 0.28123047701017795\n",
            "\n",
            "Translation:\tEine Frau in einem schwarzen Hemd wird von einem <unk> . \n",
            "Target:\tEine Frau in einem schwarzen T-Shirt scheint eine Rolltreppe <unk> . \n",
            "BLEU score = 0.5850909428128345\n",
            "\n",
            "Translation:\tEine Frau in einem blauen Oberteil und Jeans wirft etwas in einem blauen Trikot . \n",
            "Target:\tEine Frau in einer blauen Bluse und Jeans wirft etwas in einen Müllcontainer . \n",
            "BLEU score = 0.6896223302202278\n",
            "\n",
            "Translation:\tEine blonde Frau in einem braunen Hut wird von einem <unk> in der Kamera . \n",
            "Target:\tEine blonde Frau mit braunem Cowboyhut macht eine <unk> Geste in Richtung der Kamera . \n",
            "BLEU score = 0.5690479904744572\n",
            "\n",
            "Translation:\tEin kleines Mädchen mit einem roten Helm geht einen <unk> entlang . \n",
            "Target:\tEin kleines barfüßiges Mädchen mit einem Helm geht über einen mit <unk> bedeckten umgefallenen Baum . \n",
            "BLEU score = 0.49111770617222705\n",
            "\n",
            "Translation:\tEine Person in <unk> Kleidung posiert mit einem <unk> . \n",
            "Target:\tEine Person in Winterkleidung posiert mit einem Schneemann und ist dabei von einer schneebedeckten Landschaft umgeben . \n",
            "BLEU score = 0.24045075199320823\n",
            "\n",
            "Translation:\tEin Mann in <unk> Kleidung hält sich mit seinem Boden in der Hand . \n",
            "Target:\tEin <unk> gekleideter Mann hält sich mit seinen Händen über dem Boden . \n",
            "BLEU score = 0.6603983270843637\n",
            "\n",
            "Translation:\tZwei Frauen in roten und ein Mann machen ein Mann in einem roten <unk> . \n",
            "Target:\tZwei Frauen in Rot und ein Mann , der aus einer <unk> Toilette kommt . \n",
            "BLEU score = 0.5520139706197733\n",
            "\n",
            "Translation:\tEine Gruppe von Personen in lila Kleidung und <unk> geht in die Luft . \n",
            "Target:\tEine Gruppe von Menschen in lilafarbenen T-Shirts und hellbraunen Hosen geht gemeinsam in dieselbe Richtung . \n",
            "BLEU score = 0.36095000105065633\n",
            "\n",
            "Translation:\tEin kleines Mädchen in einem <unk> schaut auf eine Frau in einem schwarzen Kleid . \n",
            "Target:\tEin kleines Mädchen in einem gepunkteten Kleid blickt nach hinten zu einer Frau in einem schwarzen Kleid . \n",
            "BLEU score = 0.6053533711284104\n",
            "\n",
            "Translation:\tEine Gruppe von Personen in einem <unk> und einem <unk> macht ein Bild von einem <unk> . \n",
            "Target:\tEine Gruppe von Asiaten betrachtet und fotografiert ein neues Fahrzeug bei einer <unk> . \n",
            "BLEU score = 0.45829617993054067\n",
            "\n",
            "Translation:\tEine Frau in einem roten Kleid steht vor einem roten Trikot vor einem roten Trikot . \n",
            "Target:\tEine Frau mit einem roten Kleid und rotem Haar steht vor einem leuchtend bunten Haus . \n",
            "BLEU score = 0.6154634943747401\n",
            "\n",
            "Translation:\tZwei Männer in schwarzer Kleidung und blauer Kleidung vor einem roten Trikot . \n",
            "Target:\tZwei schwarz gekleidete Männer mit einer grünen und einer roten Fliege treten vor einer Menschenmenge auf . \n",
            "BLEU score = 0.3758270263907132\n",
            "\n",
            "Translation:\tKleine Erwachsene in roten Kleidung und schwarzen Shorts und schwarzen Hemden sind in einer <unk> . \n",
            "Target:\tJunge Erwachsene in einer <unk> Linie , die rote , gelbe und schwarze T-Shirts tragen , führen einen Sketch auf . \n",
            "BLEU score = 0.5034460980211569\n",
            "\n",
            "Translation:\tEine <unk> wird von einem Mädchen , während ein Mädchen zu sehen . \n",
            "Target:\t<unk> Mädchen wird eine Frisbeescheibe <unk> , während das andere Mädchen darum zu <unk> scheint . \n",
            "BLEU score = 0.4389190086193597\n",
            "\n",
            "Translation:\tZwei kleine Mädchen spielen in weißen Kleidung , während ein anderer Spieler in einem roten Trikot in der Hand . \n",
            "Target:\tZwei junge Mädchen spielen Basketball , wobei die eine in Weiß einen Korbleger versucht und die andere in Rot verteidigt . \n",
            "BLEU score = 0.45323101974388247\n",
            "\n",
            "Translation:\tEine Frau in weißen Shorts und einem gelben Oberteil und einem weißen Shorts ist einen Stock . \n",
            "Target:\tEine Frau in weißen Shorts , gelbem Oberteil und weißen Sandalen scheint einen Stock zu werfen . \n",
            "BLEU score = 0.7267206434017633\n",
            "\n",
            "Translation:\tEin Mann , der sich vor einem <unk> liegt , das auf dem Boden . \n",
            "Target:\tEin Mann , der aussieht , als ob er geschlagen <unk> wäre , liegt vor <unk> <unk> auf dem Boden . \n",
            "BLEU score = 0.44588348210964174\n",
            "\n",
            "Translation:\tZwei Frauen in blauen Kleidung sitzen draußen und einem <unk> . \n",
            "Target:\tZwei Frauen in blauen Anzügen sitzen draußen auf dem Beton , während sich hinter ihnen Beton und <unk> <unk> befinden . \n",
            "BLEU score = 0.32690230763900546\n",
            "\n",
            "Translation:\tEine <unk> in einem Raum und <unk> steht unter einer <unk> und <unk> . \n",
            "Target:\tEin Raum voll mit sitzenden und stehenden Erwachsenen und Kindern , in dem acht pinkfarbene und schwarze Luftballons von der Decke hängen . \n",
            "BLEU score = 0.14462367322991612\n",
            "\n",
            "Translation:\tEine alte Frau sitzt an einem Tisch und trägt eine schwarze Kleidung und eine schwarze Kleidung . \n",
            "Target:\tEine afroamerikanische Frau sitzt an einem braunen Tisch und trägt ein lilafarbenes Kleid , pinkfarbene Schuhe und eine schwarze Sonnenbrille . \n",
            "BLEU score = 0.4679250625144918\n",
            "\n",
            "Translation:\tEin Mann in einem blauen Hemd hält ein Schild auf der Hand , das auf dem „ <unk> <unk> <unk> <unk> <unk> . \n",
            "Target:\tEin Mann in einem blauen Hemd hält ein Schild , auf dem steht : „ <unk> <unk> <unk> <unk> <unk> <unk> <unk> <unk> . “ \n",
            "BLEU score = 0.757455162472565\n",
            "\n",
            "Translation:\tMehrere Arbeiter in der Mitte , die in der Hand wird , die sich an einem <unk> vorbei . \n",
            "Target:\tEin paar Bauarbeiter , in Schubkarren wird Beton <unk> , sie bauen oder reparieren eine Wand und die Wand ist noch nass . \n",
            "BLEU score = 0.3354653524787597\n",
            "\n",
            "Translation:\tEine junge Frau und eine junge Frau tragen <unk> und <unk> in der Hand , während andere Personen in der andere Personen . \n",
            "Target:\tEine junge Frau und eine ältere Frau in traditionellen Saris <unk> <unk> , während drei weitere Personen in moderner Kleidung nur von der Taille abwärts auf dem Bild zu sehen sind . \n",
            "BLEU score = 0.4432002518897664\n",
            "\n",
            "Translation:\tEin Mann mit einem <unk> , der in der Hand in der Hand hält , und ein <unk> . \n",
            "Target:\tEin Metallarbeiter in einer Gegend der mittleren <unk> oder <unk> bedient mit der rechten Hand ein <unk> , während er in der linken Hand die Maske hält . \n",
            "BLEU score = 0.29616759594672815\n",
            "\n",
            "Mean BLEU score = 0.47430685297252406\n"
          ]
        }
      ],
      "source": [
        "translation(valid_iter_EN2DE, model_EN2DE, SRC_EN2DE, TGT_EN2DE)"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "name": "Traducao com Atencao.ipynb",
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "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": "3.9.9"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}