seq2seq-translation-batched-注释.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](https://i.imgur.com/eBRPvWB.png)\n",
    "\n",
    "# Practical PyTorch: Translation with a Sequence to Sequence Network and Attention\n",
    "\n",
    "In this project we will be teaching a neural network to translate from French to English.\n",
    "\n",
    "```\n",
    "[KEY: > input, = target, < output]\n",
    "\n",
    "> il est en train de peindre un tableau .\n",
    "= he is painting a picture .\n",
    "< he is painting a picture .\n",
    "\n",
    "> pourquoi ne pas essayer ce vin delicieux ?\n",
    "= why not try that delicious wine ?\n",
    "< why not try that delicious wine ?\n",
    "\n",
    "> elle n est pas poete mais romanciere .\n",
    "= she is not a poet but a novelist .\n",
    "< she not not a poet but a novelist .\n",
    "\n",
    "> vous etes trop maigre .\n",
    "= you re too skinny .\n",
    "< you re all alone .\n",
    "```\n",
    "\n",
    "... to varying degrees of success.\n",
    "\n",
    "This is made possible by the simple but powerful idea of the [sequence to sequence network](http://arxiv.org/abs/1409.3215), in which two recurrent neural networks work together to transform one sequence to another. An encoder network condenses an input sequence into a single vector, and a decoder network unfolds that vector into a new sequence.\n",
    "\n",
    "To improve upon this model we'll use an [attention mechanism](https://arxiv.org/abs/1409.0473), which lets the decoder learn to focus over a specific range of the input sequence."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Sequence to Sequence Learning\n",
    "\n",
    "A [Sequence to Sequence network](http://arxiv.org/abs/1409.3215), or seq2seq network, or [Encoder Decoder network](https://arxiv.org/pdf/1406.1078v3.pdf), is a model consisting of two separate RNNs called the **encoder** and **decoder**. The encoder reads an input sequence one item at a time, and outputs a vector at each step. The final output of the encoder is kept as the **context** vector. The decoder uses this context vector to produce a sequence of outputs one step at a time.\n",
    "\n",
    "![](https://i.imgur.com/tVtHhNp.png)\n",
    "\n",
    "When using a single RNN, there is a one-to-one relationship between inputs and outputs. We would quickly run into problems with different sequence orders and lengths that are common during translation. Consider the simple sentence \"Je ne suis pas le chat noir\" &rarr; \"I am not the black cat\". Many of the words have a pretty direct translation, like \"chat\" &rarr; \"cat\". However the differing grammars cause words to be in different orders, e.g. \"chat noir\" and \"black cat\". There is also the \"ne ... pas\" &rarr; \"not\" construction that makes the two sentences have different lengths.\n",
    "\n",
    "With the seq2seq model, by encoding many inputs into one vector, and decoding from one vector into many outputs, we are freed from the constraints of sequence order and length. The encoded sequence is represented by a single vector, a single point in some N dimensional space of sequences. In an ideal case, this point can be considered the \"meaning\" of the sequence.\n",
    "\n",
    "This idea can be extended beyond sequences. Image captioning tasks take an [image as input, and output a description](https://arxiv.org/abs/1411.4555) of the image (img2seq). Some image generation tasks take a [description as input and output a generated image](https://arxiv.org/abs/1511.02793) (seq2img). These models can be referred to more generally as \"encoder decoder\" networks."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## The Attention Mechanism\n",
    "\n",
    "The fixed-length vector carries the burden of encoding the the entire \"meaning\" of the input sequence, no matter how long that may be. With all the variance in language, this is a very hard problem. Imagine two nearly identical sentences, twenty words long, with only one word different. Both the encoders and decoders must be nuanced enough to represent that change as a very slightly different point in space.\n",
    "\n",
    "The **attention mechanism** [introduced by Bahdanau et al.](https://arxiv.org/abs/1409.0473) addresses this by giving the decoder a way to \"pay attention\" to parts of the input, rather than relying on a single vector. For every step the decoder can select a different part of the input sentence to consider.\n",
    "\n",
    "![](https://i.imgur.com/5y6SCvU.png)\n",
    "\n",
    "Attention is calculated using the current hidden state and each encoder output, resulting in a vector the same size as the input sequence, called the *attention weights*. These weights are multiplied by the encoder outputs to create a weighted sum of encoder outputs, which is called the *context* vector. The context vector and hidden state are used to predict the next output element.\n",
    "\n",
    "![](https://i.imgur.com/K1qMPxs.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Requirements\n",
    "\n",
    "You will need [PyTorch](http://pytorch.org/) to build and train the models, and [matplotlib](https://matplotlib.org/) for plotting training and visualizing attention outputs later. The rest are builtin Python libraries."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import unicodedata\n",
    "import string\n",
    "import re\n",
    "import random\n",
    "import time\n",
    "import datetime\n",
    "import math\n",
    "import socket\n",
    "hostname = socket.gethostname()\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "from torch.autograd import Variable\n",
    "from torch import optim\n",
    "import torch.nn.functional as F\n",
    "from torch.nn.utils.rnn import pad_packed_sequence, pack_padded_sequence#, masked_cross_entropy\n",
    "from masked_cross_entropy import *             # 这个loss函数不是torch库里的\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib.ticker as ticker\n",
    "import numpy as np\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Here we will also define a constant to decide whether to use the GPU (with CUDA specifically) or the CPU. **If you don't have a GPU, set this to `False`**. Later when we create tensors, this variable will be used to decide whether we keep them on CPU or move them to GPU."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "USE_CUDA = False"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Loading data files\n",
    "\n",
    "The data for this project is a set of many thousands of English to French translation pairs.\n",
    "\n",
    "[This question on Open Data Stack Exchange](http://opendata.stackexchange.com/questions/3888/dataset-of-sentences-translated-into-many-languages) pointed me to the open translation site http://tatoeba.org/ which has downloads available at http://tatoeba.org/eng/downloads - and better yet, someone did the extra work of splitting language pairs into individual text files here: http://www.manythings.org/anki/\n",
    "\n",
    "The English to French pairs are too big to include in the repo, so download `fra-eng.zip`, extract the text file in there, and rename it to `data/eng-fra.txt` before continuing (for some reason the zipfile is named backwards). The file is a tab separated list of translation pairs:\n",
    "\n",
    "```\n",
    "I am cold.    Je suis froid.\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Similar to the character encoding used in the character-level RNN tutorials, we will be representing each word in a language as a one-hot vector, or giant vector of zeros except for a single one (at the index of the word). Compared to the dozens of characters that might exist in a language, there are many many more words, so the encoding vector is much larger. We will however cheat a bit and trim the data to only use a few thousand words per language."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Indexing words\n",
    "\n",
    "We'll need a unique index per word to use as the inputs and targets of the networks later. To keep track of all this we will use a helper class called `Lang` which has word &rarr; index (`word2index`) and index &rarr; word (`index2word`) dictionaries, as well as a count of each word (`word2count`). This class includes a function `trim(min_count)` to remove rare words once they are all counted."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 一些标记 \n",
    "\n",
    "PAD_token = 0              # Padding. 由于并行训练，句子的长度应该一样; 当句子长度不够，则用PAD_token填充.\n",
    "SOS_token = 1              # Start Of Sentence 句子开头\n",
    "EOS_token = 2              # End Of Sentence   句子结尾\n",
    "\n",
    "# 单词表\n",
    "class Lang:\n",
    "    def __init__(self, name):\n",
    "        self.name = name\n",
    "        self.trimmed = False\n",
    "        self.word2index = {}\n",
    "        self.word2count = {}\n",
    "        self.index2word = {0: \"PAD\", 1: \"SOS\", 2: \"EOS\"}\n",
    "        self.n_words = 3 # Count default tokens\n",
    "\n",
    "    def index_words(self, sentence):\n",
    "        for word in sentence.split(' '):\n",
    "            self.index_word(word)\n",
    "\n",
    "    def index_word(self, word):\n",
    "        if word not in self.word2index:\n",
    "            self.word2index[word] = self.n_words\n",
    "            self.word2count[word] = 1\n",
    "            self.index2word[self.n_words] = word\n",
    "            self.n_words += 1\n",
    "        else:\n",
    "            self.word2count[word] += 1\n",
    "\n",
    "    # Remove words below a certain count threshold\n",
    "    def trim(self, min_count):\n",
    "        if self.trimmed: return\n",
    "        self.trimmed = True\n",
    "        \n",
    "        keep_words = []\n",
    "        \n",
    "        for k, v in self.word2count.items():\n",
    "            if v >= min_count:\n",
    "                keep_words.append(k)\n",
    "\n",
    "        print('keep_words %s / %s = %.4f' % (\n",
    "            len(keep_words), len(self.word2index), len(keep_words) / len(self.word2index)\n",
    "        ))\n",
    "\n",
    "        # Reinitialize dictionaries\n",
    "        self.word2index = {}\n",
    "        self.word2count = {}\n",
    "        self.index2word = {0: \"PAD\", 1: \"SOS\", 2: \"EOS\"}\n",
    "        self.n_words = 3 # Count default tokens\n",
    "\n",
    "        for word in keep_words:\n",
    "            self.index_word(word)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Reading and decoding files\n",
    "\n",
    "The files are all in Unicode, to simplify we will turn Unicode characters to ASCII, make everything lowercase, and trim most punctuation."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Turn a Unicode string to plain ASCII, thanks to http://stackoverflow.com/a/518232/2809427\n",
    "def unicode_to_ascii(s):\n",
    "    return ''.join(\n",
    "        c for c in unicodedata.normalize('NFD', s)\n",
    "        if unicodedata.category(c) != 'Mn'\n",
    "    )\n",
    "\n",
    "# Lowercase, trim, and remove non-letter characters .保留, . ! ? 四个字符和英文字符\n",
    "def normalize_string(s):\n",
    "    s = unicode_to_ascii(s.lower().strip())\n",
    "    s = re.sub(r\"([,.!?])\", r\" \\1 \", s)\n",
    "    s = re.sub(r\"[^a-zA-Z,.!?]+\", r\" \", s)\n",
    "    s = re.sub(r\"\\s+\", r\" \", s).strip()\n",
    "    return s"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To read the data file we will split the file into lines, and then split lines into pairs. The files are all English &rarr; Other Language, so if we want to translate from Other Language &rarr; English I added the `reverse` flag to reverse the pairs."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 读取文件，并转化为Pair\n",
    "def read_langs(lang1, lang2, reverse=False):  # 若reverse为true，则表示翻译方向为 lang2 -> lang1\n",
    "    print(\"Reading lines...\")\n",
    "\n",
    "    # Read the file and split into lines\n",
    "    filename = '../data/%s-%s.txt' % (lang1, lang2)\n",
    "    # filename = '../%s-%s.txt' % (lang1, lang2)\n",
    "    lines = open(filename).read().strip().split('\\n')\n",
    "\n",
    "    # Split every line into pairs and normalize\n",
    "    pairs = [[normalize_string(s) for s in l.split('\\t')] for l in lines]\n",
    "\n",
    "    # Reverse pairs, make Lang instances\n",
    "    if reverse:   #\n",
    "        pairs = [list(reversed(p)) for p in pairs]          # reversed 产生一个list 的迭代器，迭代器逆向访问list\n",
    "        input_lang = Lang(lang2)\n",
    "        output_lang = Lang(lang1)\n",
    "    else:\n",
    "        input_lang = Lang(lang1)\n",
    "        output_lang = Lang(lang2)\n",
    "\n",
    "    return input_lang, output_lang, pairs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 限制句子的token数量. 不满足条件的过滤掉。\n",
    "MIN_LENGTH = 3\n",
    "MAX_LENGTH = 25\n",
    "\n",
    "def filter_pairs(pairs):\n",
    "    filtered_pairs = []\n",
    "    for pair in pairs:\n",
    "        if len(pair[0]) >= MIN_LENGTH and len(pair[0]) <= MAX_LENGTH \\\n",
    "            and len(pair[1]) >= MIN_LENGTH and len(pair[1]) <= MAX_LENGTH:\n",
    "                filtered_pairs.append(pair)\n",
    "    return filtered_pairs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The full process for preparing the data is:\n",
    "\n",
    "* Read text file and split into lines\n",
    "* Split lines into pairs and normalize\n",
    "* Filter to pairs of a certain length\n",
    "* Make word lists from sentences in pairs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Reading lines...\n",
      "Read 160872 sentence pairs\n",
      "Filtered to 30262 pairs\n",
      "Indexing words...\n",
      "Indexed 7645 words in input language, 4755 words in output\n"
     ]
    }
   ],
   "source": [
    "def prepare_data(lang1_name, lang2_name, reverse=False):\n",
    "    input_lang, output_lang, pairs = read_langs(lang1_name, lang2_name, reverse)\n",
    "    print(\"Read %d sentence pairs\" % len(pairs))\n",
    "    \n",
    "    pairs = filter_pairs(pairs)\n",
    "    print(\"Filtered to %d pairs\" % len(pairs))\n",
    "    \n",
    "    print(\"Indexing words...\")\n",
    "    for pair in pairs:\n",
    "        input_lang.index_words(pair[0])\n",
    "        output_lang.index_words(pair[1])\n",
    "    \n",
    "    print('Indexed %d words in input language, %d words in output' % (input_lang.n_words, output_lang.n_words))\n",
    "    return input_lang, output_lang, pairs\n",
    "\n",
    "input_lang, output_lang, pairs = prepare_data('eng', 'fra', True)   # 参数3为True，表示翻译方向为： 法文 -> english"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "['va !', 'go .']\n",
      "法文单词数 7645\n",
      "英文单词书 4755\n"
     ]
    }
   ],
   "source": [
    "# 测试\n",
    "print(pairs[0])\n",
    "print(\"法文单词数 %s\" % len(input_lang.index2word))\n",
    "print(\"英文单词书 %s\" % len(output_lang.index2word))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Filtering vocabularies\n",
    "\n",
    "To get something that trains in under an hour, we'll trim the data set a bit. First we will use the `trim` function on each language (defined above) to only include words that are repeated a certain amount of times through the dataset (this softens the difficulty of learning a correct translation for words that don't appear often)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "keep_words 1945 / 7642 = 0.2545\n",
      "keep_words 1708 / 4752 = 0.3594\n"
     ]
    }
   ],
   "source": [
    "# 将出现频次<5的单词都去掉. (MIN_COUNT=5时，只有约30% 被留下)\n",
    "MIN_COUNT = 5\n",
    "\n",
    "input_lang.trim(MIN_COUNT)\n",
    "output_lang.trim(MIN_COUNT)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Filtering pairs\n",
    "\n",
    "Now we will go back to the set of all sentence pairs and remove those with unknown words."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Trimmed from 30262 pairs to 19592, 0.6474 of total\n"
     ]
    }
   ],
   "source": [
    "# 只有如下句子会被保留： 句子中所有单词都在单词表中(MIN_COUNT=5)\n",
    "keep_pairs = []\n",
    "\n",
    "for pair in pairs:\n",
    "    input_sentence = pair[0]\n",
    "    output_sentence = pair[1]\n",
    "    keep_input = True\n",
    "    keep_output = True\n",
    "    \n",
    "    for word in input_sentence.split(' '):\n",
    "        if word not in input_lang.word2index:\n",
    "            keep_input = False\n",
    "            break\n",
    "\n",
    "    for word in output_sentence.split(' '):\n",
    "        if word not in output_lang.word2index:\n",
    "            keep_output = False\n",
    "            break\n",
    "\n",
    "    # Remove if pair doesn't match input and output conditions\n",
    "    if keep_input and keep_output:\n",
    "        keep_pairs.append(pair)\n",
    "\n",
    "print(\"Trimmed from %d pairs to %d, %.4f of total\" % (len(pairs), len(keep_pairs), len(keep_pairs) / len(pairs)))\n",
    "pairs = keep_pairs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "['va !', 'go .']\n"
     ]
    }
   ],
   "source": [
    "# 测试\n",
    "print(pairs[0])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Turning training data into Tensors\n",
    "\n",
    "To train we need to turn the sentences into something the neural network can understand, which of course means numbers. Each sentence will be split into words and turned into a `LongTensor` which represents the index (from the Lang indexes made earlier) of each word. While creating these tensors we will also append the EOS token to signal that the sentence is over.\n",
    "\n",
    "![](https://i.imgur.com/LzocpGH.png)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Return a list of indexes, one for each word in the sentence, plus EOS\n",
    "# 将句子转化为 index，句子结尾添加 EOS 标记。\n",
    "def indexes_from_sentence(lang, sentence):\n",
    "    return [lang.word2index[word] for word in sentence.split(' ')] + [EOS_token]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can make better use of the GPU by training on batches of many sequences at once, but doing so brings up the question of how to deal with sequences of varying lengths. The simple solution is to \"pad\" the shorter sentences with some padding symbol (in this case `0`), and ignore these padded spots when calculating the loss.\n",
    "# 通过PAD_token ，将长度不一的句子填充成长度一致的。计算LOSS时，pad标记不考虑。\n",
    "![](https://i.imgur.com/gGlkEEF.png)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Pad a with the PAD symbol\n",
    "# 添加Padding\n",
    "def pad_seq(seq, max_length):\n",
    "    seq += [PAD_token for i in range(max_length - len(seq))]\n",
    "    return seq"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To create a Variable for a full batch of inputs (and targets) we get a random sample of sequences and pad them all to the length of the longest sequence. We'll keep track of the lengths of each batch in order to un-pad later.\n",
    "\n",
    "Initializing a `LongTensor` with an array (batches) of arrays (sequences) gives us a `(batch_size x max_len)` tensor - selecting the first dimension gives you a single batch, which is a full sequence. When training the model we'll want a single time step at once, so we'll transpose to `(max_len x batch_size)`. Now selecting along the first dimension returns a single time step across batches.\n",
    "# transpose后保证每次执行batch中的一个时间点\n",
    "![](https://i.imgur.com/nBxTG3v.png)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 随机选取batch_size个 pair对，将单词转为数字index, 按照其中最长的句子，用PAD_token填充。\n",
    "def random_batch(batch_size):\n",
    "    input_seqs = []\n",
    "    target_seqs = []\n",
    "\n",
    "    # Choose random pairs\n",
    "    for i in range(batch_size):\n",
    "        pair = random.choice(pairs)                                        # 句子对：[\"fra sentence\",\"english sentence!\"]\n",
    "        input_seqs.append(indexes_from_sentence(input_lang, pair[0]))      # input index，如 [1879, 1735, 1262, 592, 2]\n",
    "        target_seqs.append(indexes_from_sentence(output_lang, pair[1]))    # output index, 如 [1141, 1645, 352, 1114, 2] \n",
    "\n",
    "    # Zip into pairs, sort by length (descending), unzip\n",
    "    seq_pairs = sorted(zip(input_seqs, target_seqs), key=lambda p: len(p[0]), reverse=True)# 按照input sentence长度，从大到小排序\n",
    "    input_seqs, target_seqs = zip(*seq_pairs)                             # 输入句子的index和输出句子的index\n",
    "    \n",
    "    # For input and target sequences, get array of lengths and pad with 0s to max length\n",
    "    input_lengths = [len(s) for s in input_seqs]                          # batch中每一个input句子的长度\n",
    "    input_padded = [pad_seq(s, max(input_lengths)) for s in input_seqs]   # 句子长度不足 max_input_length 时，填充PAD_token\n",
    "    target_lengths = [len(s) for s in target_seqs]                        # 同上\n",
    "    target_padded = [pad_seq(s, max(target_lengths)) for s in target_seqs]\n",
    "\n",
    "    # Turn padded arrays into (batch_size x max_len) tensors, transpose into (max_len x batch_size)\n",
    "    # LongTensor: Long类型的Tensor\n",
    "    input_var = Variable(torch.LongTensor(input_padded)).transpose(0, 1)  \n",
    "    target_var = Variable(torch.LongTensor(target_padded)).transpose(0, 1)\n",
    "    \n",
    "    if USE_CUDA:\n",
    "        input_var = input_var.cuda()\n",
    "        target_var = target_var.cuda()\n",
    "        \n",
    "    #  注意：返回的input_var和target_var并不进行one-hot编码\n",
    "    return input_var, input_lengths, target_var, target_lengths "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can test this to see that it will return a `(max_len x batch_size)` tensor for input and target sentences, along with a corresponding list of batch lenghts for each (which we will use for masking later).\n",
    "\n",
    "# random_batch 返回的Tensor，shape= max_length * batch_size"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([7, 2])\n"
     ]
    }
   ],
   "source": [
    "a = random_batch(2)\n",
    "print(a[0].shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Building the models\n",
    "* GRU的理解\n",
    "    * update gate : 上一个Hidden State，有多大的比例并入下一个hidden state. 用于长时记忆\n",
    "    * reset  gate : 上一个Hidden State，和当前word embeded vector 有多大的关联"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## The Encoder\n",
    "\n",
    "<img src=\"images/encoder-network.png\" style=\"float: right\" />\n",
    "\n",
    "The encoder will take a batch of word sequences, a `LongTensor` of size `(max_len x batch_size)`, and output an encoding for each word, a `FloatTensor` of size `(max_len x batch_size x hidden_size)`.\n",
    "\n",
    "The word inputs are fed through an [embedding layer `nn.Embedding`](http://pytorch.org/docs/nn.html#embedding) to create an embedding for each word, with size `seq_len x hidden_size` (as if it was a batch of words). This is resized to `seq_len x 1 x hidden_size` to fit the expected input of the [GRU layer `nn.GRU`](http://pytorch.org/docs/nn.html#gru). The GRU will return both an output sequence of size `seq_len x hidden_size`.\n",
    "* 图中各个Tensor 的具体信息\n",
    "    * Input  \n",
    "        * Shape:(max_length,batch_size)\n",
    "        * 类型： Long-Tensor \n",
    "    * Embedding\n",
    "        * 类型:  nn.Embedding\n",
    "        * shape: (input_size,hidden_size)，input_size即单词数量，hidden_size是每个单词会映射到的空间维度\n",
    "        * torch.nn.Embedding(num_embeddings, embedding_dim)\n",
    "            * num_embeddings, 单词数量 (size of the dictionary of embeddings)\n",
    "            * embedding_dim，每个单词的维度 (the size of each embedding vector) \n",
    "    * Embeded  \n",
    "        * Input在Embedding查表后的值\n",
    "        * shape： (max_length,batch_size,dim_of_word) \n",
    "    * prev_hidden: RNN上一个time step的hidden state\n",
    "    \n",
    "* class EncoderRNN 构造函数参数意义 [nn.GRU 官网](https://pytorch.org/docs/stable/nn.html)\n",
    "    * input_size: 单词个数\n",
    "    * hidden_size : 每个单词映射到的维度\n",
    "    * n_layers: 要叠加多少个gru层. (前一个GRU的输出，会作为下一个GRU的输入,此处为1) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "class EncoderRNN(nn.Module):\n",
    "    def __init__(self, input_size, hidden_size, n_layers=1, dropout=0.1):\n",
    "        # input_size 单词库数量\n",
    "        # hidden_size : word embedding 后的特征维度\n",
    "        super(EncoderRNN, self).__init__()\n",
    "        \n",
    "        self.input_size = input_size\n",
    "        self.hidden_size = hidden_size\n",
    "        self.n_layers = n_layers\n",
    "        self.dropout = dropout\n",
    "        \n",
    "        self.embedding = nn.Embedding(input_size, hidden_size)  # 似乎有问题，应该为 fra.words_num * input_size\n",
    "        # 此处：word embedding 维度和gru 的隐藏层维度大小相同\n",
    "        self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=self.dropout, bidirectional=True)\n",
    "        \n",
    "        \n",
    "    def forward(self, input_seqs, input_lengths, hidden=None):\n",
    "        # input_seqs : mini-batch，不做one-hot处理。一次forword只处理所有句子的第k个单词。\n",
    "        # input_lengths ：一个batch中，每一个句子的未padding前的长度，用来训练时，屏蔽掉前面PAD_token填充的词。\n",
    "        # hidden 上一次执行gru的hidden state tensor\n",
    "        \n",
    "        print(\"Encoder forword\")\n",
    "        # Note: we run this all at once (over multiple batches of multiple sequences)\n",
    "        \n",
    "        embedded = self.embedding(input_seqs)      # 先做embedding\n",
    "        packed = torch.nn.utils.rnn.pack_padded_sequence(embedded, input_lengths) \n",
    "        outputs, hidden = self.gru(packed, hidden) # 执行完所有time-step的RNN\n",
    "        outputs, output_lengths = torch.nn.utils.rnn.pad_packed_sequence(outputs) # unpack (back to padded)\n",
    "        \n",
    "        # 双向gru 后，其shape 为 [max_input_length, batch_size, hidden_state*2]\n",
    "        # 此处直接进行element-wise 相加\n",
    "        # 论文中： bi-d 执行的是concat 操作，此处直接相加！[diff]\n",
    "        outputs = outputs[:, :, :self.hidden_size] + outputs[:, : ,self.hidden_size:] # Sum bidirectional outputs\n",
    "        # 返回值,RNN输出和hidden-state. \n",
    "        return outputs, hidden"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Attention Decoder"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Interpreting the Bahdanau et al. model\n",
    "\n",
    "[Neural Machine Translation by Jointly Learning to Align and Translate](https://arxiv.org/abs/1409.0473) (Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio) introduced the idea of using attention for seq2seq translation.\n",
    "\n",
    "Each decoder output is conditioned on the previous outputs and some $\\mathbf x$, where $\\mathbf x$ consists of the current hidden state (which takes into account previous outputs) and the attention \"context\", which is calculated below. The function $g$ is a fully-connected layer with a nonlinear activation, which takes as input the values $y_{i-1}$, $s_i$, and $c_i$ concatenated.\n",
    "\n",
    "$$\n",
    "p(y_i \\mid \\{y_1,...,y_{i-1}\\},\\mathbf{x}) = g(y_{i-1}, s_i, c_i)\n",
    "$$\n",
    "\n",
    "The current hidden state $s_i$ is calculated by an RNN $f$ with the last hidden state $s_{i-1}$, last decoder output value $y_{i-1}$, and context vector $c_i$.\n",
    "\n",
    "In the code, the RNN will be a `nn.GRU` layer, the hidden state $s_i$ will be called `hidden`, the output $y_i$ called `output`, and context $c_i$ called `context`.\n",
    "# (1)\n",
    "$$\n",
    "s_i = f(s_{i-1}, y_{i-1}, c_i)\n",
    "$$\n",
    "\n",
    "The context vector $c_i$ is a weighted sum of all encoder outputs, where each weight $a_{ij}$ is the amount of \"attention\" paid to the corresponding encoder output $h_j$.\n",
    "# (2)\n",
    "$$\n",
    "c_i = \\sum_{j=1}^{T_x} a_{ij} h_j\n",
    "$$\n",
    "\n",
    "... where each weight $a_{ij}$ is a normalized (over all steps) attention \"energy\" $e_{ij}$ ...\n",
    "# (3)\n",
    "$$\n",
    "a_{ij} = \\dfrac{exp(e_{ij})}{\\sum_{k=1}^{T} exp(e_{ik})}\n",
    "$$\n",
    "\n",
    "... where each attention energy is calculated with some function $a$ (such as another linear layer) using the last hidden state $s_{i-1}$ and that particular encoder output $h_j$:\n",
    "# (4) \n",
    "$$\n",
    "e_{ij} = a(s_{i-1}, h_j)\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Interpreting the Luong et al. models"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[Effective Approaches to Attention-based Neural Machine Translation](https://arxiv.org/abs/1508.04025) (Minh-Thang Luong, Hieu Pham, Christopher D. Manning) describe a few more attention models that offer improvements and simplifications. They describe a few \"global attention\" models, the distinction between them being the way the attention scores are calculated.\n",
    "\n",
    "The general form of the attention calculation relies on the target (decoder) side hidden state and corresponding source (encoder) side state, normalized over all states to get values summing to 1:\n",
    "\n",
    "$$\n",
    "a_t(s) = align(h_t, \\bar h_s)  = \\dfrac{exp(score(h_t, \\bar h_s))}{\\sum_{s'} exp(score(h_t, \\bar h_{s'}))}\n",
    "$$\n",
    "\n",
    "The specific \"score\" function that compares two states is either *dot*, a simple dot product between the states; *general*, a a dot product between the decoder hidden state and a linear transform of the encoder state; or *concat*, a dot product between a new parameter $v_a$ and a linear transform of the states concatenated together.\n",
    "* score三种计算方法, hs和ht，一个是decoder的hidden-state,另一个是encoder所有output\n",
    "   \n",
    "$$\n",
    "score(h_t, \\bar h_s) =\n",
    "\\begin{cases}\n",
    "h_t ^\\top \\bar h_s & dot \\\\\n",
    "h_t ^\\top \\textbf{W}_a \\bar h_s & general \\\\\n",
    "v_a ^\\top \\textbf{W}_a [ h_t ; \\bar h_s ] & concat\n",
    "\\end{cases}\n",
    "$$\n",
    "\n",
    "The modular definition of these scoring functions gives us an opportunity to build specific attention module that can switch between the different score methods. The input to this module is always the hidden state (of the decoder RNN) and set of encoder outputs."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Implementing an attention module"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Attn 类: Attention机制获取每个encoder输出的权重; 输入为encoder-outputs和 Decoder的Hidden-State\n",
    "# Decoder 的一部分\n",
    "class Attn(nn.Module):\n",
    "    def __init__(self, method, hidden_size):\n",
    "        # method  score获取的方法：(1) dot; (2) general; (3)concat\n",
    "        super(Attn, self).__init__()\n",
    "        \n",
    "        self.method = method\n",
    "        self.hidden_size = hidden_size\n",
    "        \n",
    "        if self.method == 'general':\n",
    "            self.attn = nn.Linear(self.hidden_size, hidden_size)     # 类似于没有激活函数的全连接。 输入和输出维度一样\n",
    "\n",
    "        elif self.method == 'concat':\n",
    "            self.attn = nn.Linear(self.hidden_size * 2, hidden_size)\n",
    "            self.v = nn.Parameter(torch.FloatTensor(1, hidden_size))  # 额外定义一个1*hidden_size的tensor,将score映射到一个浮点数\n",
    "\n",
    "    def forward(self, hidden, encoder_outputs):\n",
    "        # hidden ： RNN 的输出，shape = (1,batch_size,hidden_size) 【1：是因为hidden 只取一步的输出】\n",
    "        # 【注意：这篇论文中，RNN的输入为上一个预测的单词，RNN的输出再去获取context，得到最终的Decoder输出】\n",
    "        # encoder_outputs: Encoder 所有的输出 shape = (sentence_length, batch_size, hidden_size)\n",
    "        # 返回值： attention权重\n",
    "        \n",
    "        max_len = encoder_outputs.size(0)                # 句子长度\n",
    "        this_batch_size = encoder_outputs.size(1)        #  batch_size\n",
    "\n",
    "        # Create variable to store attention energies    # 临时创建一个tensor，存储energy(给定encoder_outputs,energy值是固定的)\n",
    "        attn_energies = Variable(torch.zeros(this_batch_size, max_len)) # B x S ,max_len: input句子的长度\n",
    "\n",
    "        if USE_CUDA:\n",
    "            attn_energies = attn_energies.cuda()\n",
    "\n",
    "        # 枚举每个batch的每个encoder outputs\n",
    "        # For each batch of encoder outputs\n",
    "        for b in range(this_batch_size):\n",
    "            # Calculate energy for each encoder output\n",
    "            for i in range(max_len):\n",
    "                # Energy 赋值\n",
    "                # print(\"b=%s i=%s\" %(b,i))\n",
    "                # print(\"hidden[:,b] shape\",hidden[:,b].shape)\n",
    "                # print(\"en.unsqueeze shape\", encoder_outputs[i, b].unsqueeze(0).shape)\n",
    "                attn_energies[b, i] = self.score(hidden[:, b], encoder_outputs[i, b].unsqueeze(0))\n",
    "\n",
    "        # Normalize energies to weights in range 0 to 1, resize to 1 x B x S\n",
    "        return F.softmax(attn_energies).unsqueeze(1)\n",
    "    \n",
    "    def score(self, hidden, encoder_output):\n",
    "        # hidden : Decoder 上一步的输出，shape = (1, hidden_size)\n",
    "        # encoder_output : encoder 的输出，\n",
    "        # 返回值： 一个float 值, 表示encoder_output 对应位置的energy \n",
    "        if self.method == 'dot':\n",
    "            energy = hidden.dot(encoder_output)\n",
    "            return energy\n",
    "        \n",
    "        elif self.method == 'general':\n",
    "            energy = self.attn(encoder_output)\n",
    "            #print(\"shape of energy:%s  shape of hidden :%s\" %(energy.shape,hidden.shape))\n",
    "            # 【此处修改代码】\n",
    "            # energy = hidden.dot(energy) \n",
    "            energy = hidden.squeeze().dot(energy.squeeze())\n",
    "            #print(\"energy shape: \" , energy.shape)\n",
    "            return energy\n",
    "        \n",
    "        elif self.method == 'concat':\n",
    "            energy = self.attn(torch.cat((hidden, encoder_output), 1))\n",
    "            energy = self.v.dot(energy)\n",
    "            return energy"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Decoder 部分 (两篇论文的方法都实现了,这是第一篇xxx to align and translate)\n",
    "### Implementing the Bahdanau et al. model\n",
    "\n",
    "In summary our decoder should consist of four main parts - (1)an embedding layer turning an input word into a vector; (2)a layer to calculate the attention energy per encoder output; (3) a RNN layer; and (4) an output layer.\n",
    "\n",
    "The decoder's inputs are the last RNN hidden state $s_{i-1}$, last output $y_{i-1}$, and all encoder outputs $h_*$.\n",
    "\n",
    "* 【单词做embedding操作,主要是对$y_{i-1}$ 即前一个Decoder输出的单词做embedding】embedding layer with inputs $y_{i-1}$\n",
    "    * `embedded = embedding(last_rnn_output)`\n",
    "* 【Attention权值生成】 attention layer $a$ with inputs $(s_{i-1}, h_j)$ and outputs $e_{ij}$, normalized to create $a_{ij}$\n",
    "    * `attn_energies[j] = attn_layer(last_hidden, encoder_outputs[j])`\n",
    "    * `attn_weights = normalize(attn_energies)`\n",
    "* 【Context 加权】context vector $c_i$ as an attention-weighted average of encoder outputs\n",
    "    * `context = sum(attn_weights * encoder_outputs)`\n",
    "* 【生成下一个hidden-state】RNN layer(s) $f$ with inputs $(s_{i-1}, y_{i-1}, c_i)$ and internal hidden state, outputting $s_i$\n",
    "    * `rnn_input = concat(embedded, context)`   【embeded是$y_{i-1}$做embedding后的vector】\n",
    "    * `rnn_output, rnn_hidden = rnn(rnn_input, last_hidden)` \n",
    "* an output layer $g$ with inputs $(y_{i-1}, s_i, c_i)$, outputting $y_i$\n",
    "    * `output = out(embedded, rnn_output, context)`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "class BahdanauAttnDecoderRNN(nn.Module):\n",
    "    def __init__(self, hidden_size, output_size, n_layers=1, dropout_p=0.1):\n",
    "        super(BahdanauAttnDecoderRNN, self).__init__()\n",
    "        # output_size : vocab_size\n",
    "        # hidden_size : \n",
    "        \n",
    "        # Define parameters\n",
    "        self.hidden_size = hidden_size\n",
    "        self.output_size = output_size\n",
    "        self.n_layers = n_layers\n",
    "        self.dropout_p = dropout_p\n",
    "        self.max_length = max_length\n",
    "        \n",
    "        # Define layers\n",
    "        self.embedding = nn.Embedding(output_size, hidden_size) # embedding 两个语言不同，需要不同的embedding层\n",
    "        self.dropout = nn.Dropout(dropout_p)\n",
    "        self.attn = Attn('concat', hidden_size)\n",
    "        self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=dropout_p)\n",
    "        self.out = nn.Linear(hidden_size, output_size)\n",
    "    \n",
    "    def forward(self, word_input, last_hidden, encoder_outputs):\n",
    "        # Note: we run this one step at a time\n",
    "        # TODO: FIX BATCHING\n",
    "        \n",
    "        # Get the embedding of the current input word (last output word)\n",
    "        word_embedded = self.embedding(word_input).view(1, 1, -1) # S=1 x B x N   # Tensor.View 即reshape操作，-1表示自动调整,batch为1？【有问题】\n",
    "        word_embedded = self.dropout(word_embedded)\n",
    "        \n",
    "        # Calculate attention weights and apply to encoder outputs\n",
    "        attn_weights = self.attn(last_hidden[-1], encoder_outputs)\n",
    "        # bmm ：batch Matrix Mutiple ，三维Tensor，第一维为batch，其他两维作为matrix，进行batch_size次矩阵乘法\n",
    "        context = attn_weights.bmm(encoder_outputs.transpose(0, 1)) # B x 1 x N\n",
    "        context = context.transpose(0, 1) # 1 x B x N\n",
    "        \n",
    "        # Combine embedded input word and attended context, run through RNN\n",
    "        rnn_input = torch.cat((word_embedded, context), 2)\n",
    "        output, hidden = self.gru(rnn_input, last_hidden)\n",
    "        \n",
    "        # Final output layer\n",
    "        output = output.squeeze(0) # B x N\n",
    "        output = F.log_softmax(self.out(torch.cat((output, context), 1)))\n",
    "        \n",
    "        # Return final output, hidden state, and attention weights (for visualization)\n",
    "        return output, hidden, attn_weights"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Now we can build a decoder that plugs this Attn module in after the RNN to calculate attention weights, and apply those weights to the encoder outputs to get a context vector."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "class LuongAttnDecoderRNN(nn.Module):\n",
    "    # 计算路径：Ht -> At -> Ct -> H~t\n",
    "    def __init__(self, attn_model, hidden_size, output_size, n_layers=1, dropout=0.1):\n",
    "        # attn_model : dot ; general ; concat\n",
    "        # hidden-size : hidden-state-size\n",
    "        # output-size : 输出单词的个数\n",
    "        super(LuongAttnDecoderRNN, self).__init__()\n",
    "\n",
    "        # Keep for reference\n",
    "        self.attn_model = attn_model\n",
    "        self.hidden_size = hidden_size\n",
    "        self.output_size = output_size\n",
    "        self.n_layers = n_layers\n",
    "        self.dropout = dropout\n",
    "\n",
    "        # Define layers\n",
    "        self.embedding = nn.Embedding(output_size, hidden_size)         # 输出单词个数* 每个单词的Embedding 维度(即Hidden_size)\n",
    "        self.embedding_dropout = nn.Dropout(dropout)\n",
    "        self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=dropout)\n",
    "        self.concat = nn.Linear(hidden_size * 2, hidden_size)           # 逻辑回归; 输入的特征hidden_size*2 ，输出维度：hidden_size\n",
    "        self.out = nn.Linear(hidden_size, output_size)                  # 将RNN 输出映射到one-hot格式; self.out.shape= (batch_size,vocab_size)\n",
    "        \n",
    "        # Choose attention model\n",
    "        if attn_model != 'none':\n",
    "            self.attn = Attn(attn_model, hidden_size)          # attn_model : dot;general; concat;\n",
    "\n",
    "    def forward(self, input_seq, last_hidden, encoder_outputs):  \n",
    "        # input_seq : RNN前一个时刻，输出的单词Index，shape为 [batch_size]。\n",
    "        # last_hidden ： 前一个时刻的Hidden-state，时刻0时，Hidden-state 为Encoder的最后一个输出\n",
    "        # encoder_outputs : (句子长度, batch_size, hidden_size)\n",
    "        # Note: we run this one step at a time\n",
    "        \n",
    "        # Step (1) ： 对前一个预测的单词做 embedding\n",
    "        # Get the embedding of the current input word (last output word)\n",
    "        batch_size = input_seq.size(0)\n",
    "        embedded = self.embedding(input_seq)                      # 对前一个时刻的输出单词做embedding.\n",
    "        embedded = self.embedding_dropout(embedded)\n",
    "        embedded = embedded.view(1, batch_size, self.hidden_size) # S=1 x B x N  ，第一维1表示：只有一个time-step要执行\n",
    "\n",
    "        # Step (2) : 执行GRU ，获取 Ht\n",
    "        # Get current hidden state from input word and last hidden state \n",
    "        rnn_output, hidden = self.gru(embedded, last_hidden)      # 只执行一个time-step。【此处gru input要加入context】\n",
    "\n",
    "        # Step (3) : 获取 attention 权值，加权给context\n",
    "        \n",
    "        # Calculate attention from current RNN state and all encoder outputs;\n",
    "        # apply to encoder outputs to get weighted average \n",
    "        print(\"rnn_output shape:\",rnn_output.shape)               # rnn_output.shape = (1, batch_size, hidden_size)\n",
    "        print(\"encoder_outputs.shape:\", encoder_outputs.shape)\n",
    "        attn_weights = self.attn(rnn_output, encoder_outputs)\n",
    "        context = attn_weights.bmm(encoder_outputs.transpose(0, 1)) # B x S=1 x N\n",
    "\n",
    "        # Attentional vector using the RNN hidden state and context vector\n",
    "        # concatenated together (Luong eq. 5)\n",
    "        rnn_output = rnn_output.squeeze(0) # S=1 x B x N -> B x N\n",
    "        context = context.squeeze(1)       # B x S=1 x N -> B x N\n",
    "        concat_input = torch.cat((rnn_output, context), 1)  # concat_input，即论文中的H~t , shape= (Batch_size , hidden_size*2)\n",
    "        # 修改代码\n",
    "        # concat_output = F.tanh(self.concat(concat_input))\n",
    "        concat_output = torch.tanh(self.concat(concat_input))  # 映射到word embedding 维度，比如300\n",
    "\n",
    "        # Finally predict next token (Luong eq. 6, without softmax)\n",
    "        output = self.out(concat_output)                       # 映射到 one-hot 编码\n",
    "\n",
    "        # Return final output, hidden state, and attention weights (for visualization)\n",
    "        return output, hidden, attn_weights"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Testing the models\n",
    "\n",
    "To make sure the encoder and decoder modules are working (and working together) we'll do a full test with a small batch."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "input_batches torch.Size([8, 3])\n",
      "target_batches torch.Size([7, 3])\n"
     ]
    }
   ],
   "source": [
    "# 随机生成一个batch\n",
    "small_batch_size = 3\n",
    "input_batches, input_lengths, target_batches, target_lengths = random_batch(small_batch_size)\n",
    "\n",
    "print('input_batches', input_batches.size()) # (max_len x batch_size)\n",
    "print('target_batches', target_batches.size()) # (max_len x batch_size)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Create models with a small size (a good idea for eyeball inspection):"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "small_hidden_size = 19           # gru的hidden-state 和 word-embedding size\n",
    "small_n_layers = 2\n",
    "\n",
    "encoder_test = EncoderRNN(input_lang.n_words, small_hidden_size, small_n_layers)\n",
    "decoder_test = LuongAttnDecoderRNN('general', small_hidden_size, output_lang.n_words, small_n_layers)\n",
    "# decoder_test = BahdanauAttnDecoderRNN( output_lang.n_words,small_hidden_size, small_n_layers)\n",
    "\n",
    "\n",
    "if USE_CUDA:\n",
    "    encoder_test.cuda()\n",
    "    decoder_test.cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[8, 6, 5]"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "input_lengths"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To test the encoder, run the input batch through to get per-batch encoder outputs:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Encoder forword\n",
      "encoder_outputs torch.Size([8, 3, 19])\n",
      "encoder_hidden torch.Size([4, 3, 19])\n"
     ]
    }
   ],
   "source": [
    "encoder_outputs, encoder_hidden = encoder_test(input_batches, input_lengths, None)\n",
    "\n",
    "print('encoder_outputs', encoder_outputs.size()) # max_len x batch_size x hidden_size\n",
    "print('encoder_hidden', encoder_hidden.size())   # n_layers * 2 x batch_size x hidden_size\n",
    "# hidden.size : 叠加几层，就会有几层的hidden-state\n",
    "# 但为什么是n_layers* 2？ 因为是双向的！！"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'# GRU output和hidden-state 的关系\\n# 正向的最后一个output = 正向的hidden_state\\nprint(encoder_outputs.view(7,3,2,8)[6,0,0,:])  #seq_len, batch, num_directions, hidden_size)  同 print(encoder_outputs[6,0,:])  \\nprint(encoder_hidden.view(1,2,3,8)[0,0,0,:])#num_layers, num_directions, batch, hidden_size)\\n\\n# 反向的第一个output = 反向的hidden_state\\n\\nprint(encoder_outputs.view(7,3,2,8)[0,0,1,:])  #seq_len, batch, num_directions, hidden_size)\\nprint(encoder_hidden.view(1,2,3,8)[0,1,0,:])#num_layers, num_directions, batch, hidden_size) \\n'"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "\"\"\"# GRU output和hidden-state 的关系\n",
    "# 正向的最后一个output = 正向的hidden_state\n",
    "print(encoder_outputs.view(7,3,2,8)[6,0,0,:])  #seq_len, batch, num_directions, hidden_size)  同 print(encoder_outputs[6,0,:])  \n",
    "print(encoder_hidden.view(1,2,3,8)[0,0,0,:])#num_layers, num_directions, batch, hidden_size)\n",
    "\n",
    "# 反向的第一个output = 反向的hidden_state\n",
    "\n",
    "print(encoder_outputs.view(7,3,2,8)[0,0,1,:])  #seq_len, batch, num_directions, hidden_size)\n",
    "print(encoder_hidden.view(1,2,3,8)[0,1,0,:])#num_layers, num_directions, batch, hidden_size) \n",
    "\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'max_target_length' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-30-b7be3c239bbb>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmax_target_length\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0msmall_batch_size\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mdecoder_test\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0moutput_size\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m: name 'max_target_length' is not defined"
     ]
    }
   ],
   "source": [
    "#print((max_target_length, small_batch_size, decoder_test.output_size))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then starting with a SOS token, run word tokens through the decoder to get each next word token. Instead of doing this with the whole sequence, it is done one at a time, to support using it's own predictions to make the next prediction. This will be one time step at a time, but batched per time step. In order to get this to work for short padded sequences, the batch size is going to get smaller each time."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([2, 3, 8])\n",
      "torch.Size([3])\n",
      "shape:torch.Size([3, 6, 1711]) torch.Size([3, 6])\n",
      "loss tensor(7.3703, grad_fn=<DivBackward0>)\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:45: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.\n"
     ]
    }
   ],
   "source": [
    "max_target_length = max(target_lengths)\n",
    "\n",
    "# SOS(Start Of Sentence 作为Decoder RNN 的第一个输出。句子的第一个单词就是SOS\n",
    "# Prepare decoder input and outputs\n",
    "decoder_input = Variable(torch.LongTensor([SOS_token] * small_batch_size))\n",
    "\n",
    "# decoder_hidden = encoder_hidden_state (正向的)\n",
    "decoder_hidden = encoder_hidden[:decoder_test.n_layers] # Use last (forward) hidden state from encoder\n",
    "\n",
    "# 保存Decoder的输出，Tensor Size : 句子长度* batch_size * vocab_size\n",
    "all_decoder_outputs = Variable(torch.zeros(max_target_length, small_batch_size, decoder_test.output_size)) # \n",
    "\n",
    "if USE_CUDA:\n",
    "    all_decoder_outputs = all_decoder_outputs.cuda()\n",
    "    decoder_input = decoder_input.cuda()\n",
    "\n",
    "# Decoder 由于无法\n",
    "print(decoder_hidden.shape)\n",
    "print(decoder_input.shape)\n",
    "# Run through decoder one time step at a time\n",
    "for t in range(max_target_length): \n",
    "    decoder_output, decoder_hidden, decoder_attn = decoder_test(\n",
    "        decoder_input, decoder_hidden, encoder_outputs\n",
    "    )\n",
    "    all_decoder_outputs[t] = decoder_output # Store this step's outputs .   decoder_output 维度为 :vocab_size\n",
    "    \n",
    "    # 正常来说，需要把decoder_output转化为单词。但是这里只是为了测试，所以直接从数据集去除已知对应的输出。第一个词的前一个单词是SOS。\n",
    "    decoder_input = target_batches[t] # Next input is current target\n",
    "\n",
    "print(\"shape:%s %s\" %(all_decoder_outputs.transpose(0, 1).contiguous().shape,target_batches.transpose(0, 1).contiguous().shape))\n",
    "# Test masked cross entropy loss\n",
    "loss = masked_cross_entropy(\n",
    "    all_decoder_outputs.transpose(0, 1).contiguous(),\n",
    "    target_batches.transpose(0, 1).contiguous(),\n",
    "    target_lengths\n",
    ")\n",
    "# 修改代码\n",
    "# print('loss', loss.data[0])\n",
    "print(\"loss\" , loss)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training\n",
    "\n",
    "## Defining a training iteration\n",
    "\n",
    "To train we first run the input sentence through the encoder word by word, and keep track of every output and the latest hidden state. Next the decoder is given the last hidden state of the decoder as its first hidden state, and the `<SOS>` token as its first input. From there we iterate to predict a next token from the decoder.\n",
    "\n",
    "### Teacher Forcing vs. Scheduled Sampling\n",
    "\n",
    "\"Teacher Forcing\", or maximum likelihood sampling, means using the real target outputs as each next input when training. The alternative is using the decoder's own guess as the next input. Using teacher forcing may cause the network to converge faster, but [when the trained network is exploited, it may exhibit instability](http://minds.jacobs-university.de/sites/default/files/uploads/papers/ESNTutorialRev.pdf).\n",
    "\n",
    "You can observe outputs of teacher-forced networks that read with coherent grammar but wander far from the correct translation - you could think of it as having learned how to listen to the teacher's instructions, without learning how to venture out on its own.\n",
    "\n",
    "The solution to the teacher-forcing \"problem\" is known as [Scheduled Sampling](https://arxiv.org/abs/1506.03099), which simply alternates between using the target values and predicted values when training. We will randomly choose to use teacher forcing with an if statement while training - sometimes we'll feed use real target as the input (ignoring the decoder's output), sometimes we'll use the decoder's output."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train(input_batches, input_lengths, target_batches, target_lengths, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length=MAX_LENGTH):\n",
    "    \n",
    "    # Zero gradients of both optimizers\n",
    "    encoder_optimizer.zero_grad()\n",
    "    decoder_optimizer.zero_grad()\n",
    "    loss = 0 # Added onto for each word\n",
    "\n",
    "    # Run words through encoder\n",
    "    encoder_outputs, encoder_hidden = encoder(input_batches, input_lengths, None)\n",
    "    \n",
    "    # Prepare input and output variables\n",
    "    decoder_input = Variable(torch.LongTensor([SOS_token] * batch_size))\n",
    "    decoder_hidden = encoder_hidden[:decoder.n_layers] # Use last (forward) hidden state from encoder\n",
    "\n",
    "    max_target_length = max(target_lengths)\n",
    "    all_decoder_outputs = Variable(torch.zeros(max_target_length, batch_size, decoder.output_size))\n",
    "\n",
    "    # Move new Variables to CUDA\n",
    "    if USE_CUDA:\n",
    "        decoder_input = decoder_input.cuda()\n",
    "        all_decoder_outputs = all_decoder_outputs.cuda()\n",
    "\n",
    "    # Run through decoder one time step at a time\n",
    "    for t in range(max_target_length):\n",
    "        decoder_output, decoder_hidden, decoder_attn = decoder(\n",
    "            decoder_input, decoder_hidden, encoder_outputs\n",
    "        )\n",
    "\n",
    "        all_decoder_outputs[t] = decoder_output\n",
    "        decoder_input = target_batches[t] # Next input is current target\n",
    "\n",
    "    # Loss calculation and backpropagation\n",
    "    loss = masked_cross_entropy(\n",
    "        all_decoder_outputs.transpose(0, 1).contiguous(), # -> batch x seq\n",
    "        target_batches.transpose(0, 1).contiguous(), # -> batch x seq\n",
    "        target_lengths\n",
    "    )\n",
    "    loss.backward()\n",
    "    \n",
    "    # Clip gradient norms\n",
    "    ec = torch.nn.utils.clip_grad_norm(encoder.parameters(), clip)\n",
    "    dc = torch.nn.utils.clip_grad_norm(decoder.parameters(), clip)\n",
    "\n",
    "    # Update parameters with optimizers\n",
    "    encoder_optimizer.step()\n",
    "    decoder_optimizer.step()\n",
    "    \n",
    "    return loss.data[0], ec, dc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Running training\n",
    "\n",
    "With everything in place we can actually initialize a network and start training.\n",
    "\n",
    "To start, we initialize models, optimizers, a loss function (criterion), and set up variables for plotting and tracking progress:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "ename": "JSONDecodeError",
     "evalue": "Expecting value: line 1 column 1 (char 0)",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mJSONDecodeError\u001b[0m                           Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-35-3947ff0c62da>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m     40\u001b[0m     \u001b[0;34m'clip'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mclip\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     41\u001b[0m     \u001b[0;34m'teacher_forcing_ratio'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mteacher_forcing_ratio\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 42\u001b[0;31m     \u001b[0;34m'decoder_learning_ratio'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mdecoder_learning_ratio\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     43\u001b[0m })\n\u001b[1;32m     44\u001b[0m \u001b[0mjob\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mplot_every\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mplot_every\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python3.5/dist-packages/sconce/__init__.py\u001b[0m in \u001b[0;36m__init__\u001b[0;34m(self, name, params, hostname)\u001b[0m\n\u001b[1;32m     31\u001b[0m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mhostname\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mhostname\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0msocket\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mgethostname\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     32\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 33\u001b[0;31m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstart\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     34\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     35\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mstart\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python3.5/dist-packages/sconce/__init__.py\u001b[0m in \u001b[0;36mstart\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m     35\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mstart\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     36\u001b[0m         \u001b[0mr\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mreq\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'post'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'jobs'\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m{\u001b[0m\u001b[0;34m'name'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mname\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'params'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mparams\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'hostname'\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mhostname\u001b[0m\u001b[0;34m}\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 37\u001b[0;31m         \u001b[0mbody\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mr\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjson\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     38\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     39\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mr\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstatus_code\u001b[0m \u001b[0;34m!=\u001b[0m \u001b[0;36m200\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/lib/python3/dist-packages/requests/models.py\u001b[0m in \u001b[0;36mjson\u001b[0;34m(self, **kwargs)\u001b[0m\n\u001b[1;32m    806\u001b[0m                     \u001b[0;31m# used.\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    807\u001b[0m                     \u001b[0;32mpass\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 808\u001b[0;31m         \u001b[0;32mreturn\u001b[0m \u001b[0mcomplexjson\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mloads\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtext\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    809\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    810\u001b[0m     \u001b[0;34m@\u001b[0m\u001b[0mproperty\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/lib/python3.5/json/__init__.py\u001b[0m in \u001b[0;36mloads\u001b[0;34m(s, encoding, cls, object_hook, parse_float, parse_int, parse_constant, object_pairs_hook, **kw)\u001b[0m\n\u001b[1;32m    317\u001b[0m             \u001b[0mparse_int\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mNone\u001b[0m \u001b[0;32mand\u001b[0m \u001b[0mparse_float\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mNone\u001b[0m \u001b[0;32mand\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    318\u001b[0m             parse_constant is None and object_pairs_hook is None and not kw):\n\u001b[0;32m--> 319\u001b[0;31m         \u001b[0;32mreturn\u001b[0m \u001b[0m_default_decoder\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdecode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    320\u001b[0m     \u001b[0;32mif\u001b[0m \u001b[0mcls\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    321\u001b[0m         \u001b[0mcls\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mJSONDecoder\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/lib/python3.5/json/decoder.py\u001b[0m in \u001b[0;36mdecode\u001b[0;34m(self, s, _w)\u001b[0m\n\u001b[1;32m    337\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    338\u001b[0m         \"\"\"\n\u001b[0;32m--> 339\u001b[0;31m         \u001b[0mobj\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mend\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mraw_decode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0midx\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0m_w\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    340\u001b[0m         \u001b[0mend\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0m_w\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mend\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mend\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    341\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mend\u001b[0m \u001b[0;34m!=\u001b[0m \u001b[0mlen\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/lib/python3.5/json/decoder.py\u001b[0m in \u001b[0;36mraw_decode\u001b[0;34m(self, s, idx)\u001b[0m\n\u001b[1;32m    355\u001b[0m             \u001b[0mobj\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mend\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mscan_once\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0ms\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0midx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    356\u001b[0m         \u001b[0;32mexcept\u001b[0m \u001b[0mStopIteration\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0merr\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 357\u001b[0;31m             \u001b[0;32mraise\u001b[0m \u001b[0mJSONDecodeError\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"Expecting value\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0ms\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0merr\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mvalue\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;32mfrom\u001b[0m \u001b[0;32mNone\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    358\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mobj\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mend\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mJSONDecodeError\u001b[0m: Expecting value: line 1 column 1 (char 0)"
     ]
    }
   ],
   "source": [
    "# Configure models\n",
    "attn_model = 'general'\n",
    "hidden_size = 300\n",
    "n_layers = 2\n",
    "dropout = 0.1 \n",
    "batch_size = 32\n",
    "\n",
    "# Configure training/optimization\n",
    "clip = 50.0\n",
    "teacher_forcing_ratio = 0.5\n",
    "learning_rate = 0.0001\n",
    "decoder_learning_ratio = 5.0\n",
    "n_epochs = 50000\n",
    "epoch = 0\n",
    "plot_every = 20\n",
    "print_every = 100\n",
    "evaluate_every = 1000\n",
    "\n",
    "# Initialize models\n",
    "encoder = EncoderRNN(input_lang.n_words, hidden_size, n_layers, dropout=dropout)\n",
    "decoder = LuongAttnDecoderRNN(attn_model, hidden_size, output_lang.n_words, n_layers, dropout=dropout)\n",
    "\n",
    "# Initialize optimizers and criterion\n",
    "encoder_optimizer = optim.Adam(encoder.parameters(), lr=learning_rate)\n",
    "decoder_optimizer = optim.Adam(decoder.parameters(), lr=learning_rate * decoder_learning_ratio)\n",
    "criterion = nn.CrossEntropyLoss()\n",
    "\n",
    "# Move models to GPU\n",
    "if USE_CUDA:\n",
    "    encoder.cuda()\n",
    "    decoder.cuda()\n",
    "\n",
    "import sconce\n",
    "job = sconce.Job('seq2seq-translate', {\n",
    "    'attn_model': attn_model,\n",
    "    'n_layers': n_layers,\n",
    "    'dropout': dropout,\n",
    "    'hidden_size': hidden_size,\n",
    "    'learning_rate': learning_rate,\n",
    "    'clip': clip,\n",
    "    'teacher_forcing_ratio': teacher_forcing_ratio,\n",
    "    'decoder_learning_ratio': decoder_learning_ratio\n",
    "})\n",
    "job.plot_every = plot_every\n",
    "job.log_every = print_every\n",
    "\n",
    "# Keep track of time elapsed and running averages\n",
    "start = time.time()\n",
    "plot_losses = []\n",
    "print_loss_total = 0 # Reset every print_every\n",
    "plot_loss_total = 0 # Reset every plot_every"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(plot_every)\n",
    "print(print_every)  \n",
    "job = sconce.Job('seq2seq-translate', {\n",
    "    'attn_model': attn_model,\n",
    "    'n_layers': n_layers,\n",
    "    'dropout': dropout,\n",
    "    'hidden_size': hidden_size,\n",
    "    'learning_rate': learning_rate,\n",
    "    'clip': clip,\n",
    "    'teacher_forcing_ratio': teacher_forcing_ratio,\n",
    "    'decoder_learning_ratio': decoder_learning_ratio\n",
    "})\n",
    "job.plot_every = plot_every\n",
    "job.log_every = print_every "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Plus helper functions to print time elapsed and estimated time remaining, given the current time and progress."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def as_minutes(s):\n",
    "    m = math.floor(s / 60)\n",
    "    s -= m * 60\n",
    "    return '%dm %ds' % (m, s)\n",
    "\n",
    "def time_since(since, percent):\n",
    "    now = time.time()\n",
    "    s = now - since\n",
    "    es = s / (percent)\n",
    "    rs = es - s\n",
    "    return '%s (- %s)' % (as_minutes(s), as_minutes(rs))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluating the network\n",
    "\n",
    "Evaluation is mostly the same as training, but there are no targets. Instead we always feed the decoder's predictions back to itself. Every time it predicts a word, we add it to the output string. If it predicts the EOS token we stop there. We also store the decoder's attention outputs for each step to display later."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def evaluate(input_seq, max_length=MAX_LENGTH):\n",
    "    input_lengths = [len(input_seq)]\n",
    "    input_seqs = [indexes_from_sentence(input_lang, input_seq)]\n",
    "    input_batches = Variable(torch.LongTensor(input_seqs), volatile=True).transpose(0, 1)\n",
    "    \n",
    "    if USE_CUDA:\n",
    "        input_batches = input_batches.cuda()\n",
    "        \n",
    "    # Set to not-training mode to disable dropout\n",
    "    encoder.train(False)\n",
    "    decoder.train(False)\n",
    "    \n",
    "    # Run through encoder\n",
    "    encoder_outputs, encoder_hidden = encoder(input_batches, input_lengths, None)\n",
    "\n",
    "    # Create starting vectors for decoder\n",
    "    decoder_input = Variable(torch.LongTensor([SOS_token]), volatile=True) # SOS\n",
    "    decoder_hidden = encoder_hidden[:decoder.n_layers] # Use last (forward) hidden state from encoder\n",
    "    \n",
    "    if USE_CUDA:\n",
    "        decoder_input = decoder_input.cuda()\n",
    "\n",
    "    # Store output words and attention states\n",
    "    decoded_words = []\n",
    "    decoder_attentions = torch.zeros(max_length + 1, max_length + 1)\n",
    "    \n",
    "    # Run through decoder\n",
    "    for di in range(max_length):\n",
    "        decoder_output, decoder_hidden, decoder_attention = decoder(\n",
    "            decoder_input, decoder_hidden, encoder_outputs\n",
    "        )\n",
    "        decoder_attentions[di,:decoder_attention.size(2)] += decoder_attention.squeeze(0).squeeze(0).cpu().data\n",
    "\n",
    "        # Choose top word from output\n",
    "        topv, topi = decoder_output.data.topk(1)\n",
    "        ni = topi[0][0]\n",
    "        if ni == EOS_token:\n",
    "            decoded_words.append('<EOS>')\n",
    "            break\n",
    "        else:\n",
    "            decoded_words.append(output_lang.index2word[ni])\n",
    "            \n",
    "        # Next input is chosen word\n",
    "        decoder_input = Variable(torch.LongTensor([ni]))\n",
    "        if USE_CUDA: decoder_input = decoder_input.cuda()\n",
    "\n",
    "    # Set back to training mode\n",
    "    encoder.train(True)\n",
    "    decoder.train(True)\n",
    "    \n",
    "    return decoded_words, decoder_attentions[:di+1, :len(encoder_outputs)]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can evaluate random sentences from the training set and print out the input, target, and output to make some subjective quality judgements:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def evaluate_randomly():\n",
    "    [input_sentence, target_sentence] = random.choice(pairs)\n",
    "    evaluate_and_show_attention(input_sentence, target_sentence)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Visualizing attention\n",
    "\n",
    "A useful property of the attention mechanism is its highly interpretable outputs. Because it is used to weight specific encoder outputs of the input sequence, we can imagine looking where the network is focused most at each time step.\n",
    "\n",
    "You could simply run `plt.matshow(attentions)` to see attention output displayed as a matrix, with the columns being input steps and rows being output steps:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import io\n",
    "import torchvision\n",
    "from PIL import Image\n",
    "import visdom\n",
    "vis = visdom.Visdom()\n",
    "\n",
    "def show_plot_visdom():\n",
    "    buf = io.BytesIO()\n",
    "    plt.savefig(buf)\n",
    "    buf.seek(0)\n",
    "    attn_win = 'attention (%s)' % hostname\n",
    "    vis.image(torchvision.transforms.ToTensor()(Image.open(buf)), win=attn_win, opts={'title': attn_win})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "For a better viewing experience we will do the extra work of adding axes and labels:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def show_attention(input_sentence, output_words, attentions):\n",
    "    # Set up figure with colorbar\n",
    "    fig = plt.figure()\n",
    "    ax = fig.add_subplot(111)\n",
    "    cax = ax.matshow(attentions.numpy(), cmap='bone')\n",
    "    fig.colorbar(cax)\n",
    "\n",
    "    # Set up axes\n",
    "    ax.set_xticklabels([''] + input_sentence.split(' ') + ['<EOS>'], rotation=90)\n",
    "    ax.set_yticklabels([''] + output_words)\n",
    "\n",
    "    # Show label at every tick\n",
    "    ax.xaxis.set_major_locator(ticker.MultipleLocator(1))\n",
    "    ax.yaxis.set_major_locator(ticker.MultipleLocator(1))\n",
    "\n",
    "    show_plot_visdom()\n",
    "    plt.show()\n",
    "    plt.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def evaluate_and_show_attention(input_sentence, target_sentence=None):\n",
    "    output_words, attentions = evaluate(input_sentence)\n",
    "    output_sentence = ' '.join(output_words)\n",
    "    print('>', input_sentence)\n",
    "    if target_sentence is not None:\n",
    "        print('=', target_sentence)\n",
    "    print('<', output_sentence)\n",
    "    \n",
    "    show_attention(input_sentence, output_words, attentions)\n",
    "    \n",
    "    # Show input, target, output text in visdom\n",
    "    win = 'evaluted (%s)' % hostname\n",
    "    text = '<p>&gt; %s</p><p>= %s</p><p>&lt; %s</p>' % (input_sentence, target_sentence, output_sentence)\n",
    "    vis.text(text, win=win, opts={'title': win})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Putting it all together\n",
    "\n",
    "**TODO** Run `train_epochs` for `n_epochs`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To actually train, we call the train function many times, printing a summary as we go.\n",
    "\n",
    "*Note:* If you're running this notebook you can **train, interrupt, evaluate, and come back to continue training**. Simply run the notebook starting from the following cell (running from the previous cell will reset the models)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "# Begin!\n",
    "ecs = []\n",
    "dcs = []\n",
    "eca = 0\n",
    "dca = 0\n",
    "\n",
    "while epoch < n_epochs:\n",
    "    epoch += 1\n",
    "    \n",
    "    # Get training data for this cycle\n",
    "    input_batches, input_lengths, target_batches, target_lengths = random_batch(batch_size)\n",
    "\n",
    "    # Run the train function\n",
    "    loss, ec, dc = train(\n",
    "        input_batches, input_lengths, target_batches, target_lengths,\n",
    "        encoder, decoder,\n",
    "        encoder_optimizer, decoder_optimizer, criterion\n",
    "    )\n",
    "\n",
    "    # Keep track of loss\n",
    "    print_loss_total += loss\n",
    "    plot_loss_total += loss\n",
    "    eca += ec\n",
    "    dca += dc\n",
    "    \n",
    "    job.record(epoch, loss)\n",
    "\n",
    "    if epoch % print_every == 0:\n",
    "        print_loss_avg = print_loss_total / print_every\n",
    "        print_loss_total = 0\n",
    "        print_summary = '%s (%d %d%%) %.4f' % (time_since(start, epoch / n_epochs), epoch, epoch / n_epochs * 100, print_loss_avg)\n",
    "        print(print_summary)\n",
    "        \n",
    "    if epoch % evaluate_every == 0:\n",
    "        evaluate_randomly()\n",
    "\n",
    "    if epoch % plot_every == 0:\n",
    "        plot_loss_avg = plot_loss_total / plot_every\n",
    "        plot_losses.append(plot_loss_avg)\n",
    "        plot_loss_total = 0\n",
    "        \n",
    "        # TODO: Running average helper\n",
    "        ecs.append(eca / plot_every)\n",
    "        dcs.append(dca / plot_every)\n",
    "        ecs_win = 'encoder grad (%s)' % hostname\n",
    "        dcs_win = 'decoder grad (%s)' % hostname\n",
    "        vis.line(np.array(ecs), win=ecs_win, opts={'title': ecs_win})\n",
    "        vis.line(np.array(dcs), win=dcs_win, opts={'title': dcs_win})\n",
    "        eca = 0\n",
    "        dca = 0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Plotting training loss\n",
    "\n",
    "Plotting is done with matplotlib, using the array `plot_losses` that was created while training."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def show_plot(points):\n",
    "    plt.figure()\n",
    "    fig, ax = plt.subplots()\n",
    "    loc = ticker.MultipleLocator(base=0.2) # put ticks at regular intervals\n",
    "    ax.yaxis.set_major_locator(loc)\n",
    "    plt.plot(points)\n",
    "\n",
    "show_plot(plot_losses)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "output_words, attentions = evaluate(\"je suis trop froid .\")\n",
    "plt.matshow(attentions.numpy())\n",
    "show_plot_visdom()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"elle a cinq ans de moins que moi .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"elle est trop petit .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"je ne crains pas de mourir .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"c est un jeune directeur plein de talent .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"est le chien vert aujourd hui ?\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"le chat me parle .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"des centaines de personnes furent arretees ici .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"des centaines de chiens furent arretees ici .\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluate_and_show_attention(\"ce fromage est prepare a partir de lait de chevre .\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "# Exercises\n",
    "\n",
    "* Try with a different dataset\n",
    "    * Another language pair\n",
    "    * Human &rarr; Machine (e.g. IOT commands)\n",
    "    * Chat &rarr; Response\n",
    "    * Question &rarr; Answer\n",
    "* Replace the embedding pre-trained word embeddings such as word2vec or GloVe\n",
    "* Try with more layers, more hidden units, and more sentences. Compare the training time and results.\n",
    "* If you use a translation file where pairs have two of the same phrase (`I am test \\t I am test`), you can use this as an autoencoder. Try this:\n",
    "    * Train as an autoencoder\n",
    "    * Save only the Encoder network\n",
    "    * Train a new Decoder for translation from there"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}