使用PyTorch实现Chatbot

> hello?
Bot: hello .
> where am I?
Bot: you re in a hospital .
> who are you?
Bot: i m a lawyer .
> how are you doing?
Bot: i m fine .
> are you my friend?
Bot: no .
> you're under arrest
Bot: i m trying to help you !
> i'm just kidding
Bot: i m sorry .
> where are you from?
Bot: san francisco .
> it's time for me to leave
Bot: i know .
> goodbye
Bot: goodbye .

b'L1045 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ They do not!\n'
b'L1044 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ They do to!\n'
b'L985 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ I hope so.\n'
b'L984 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ She okay?\n'
b"L925 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ Let's go.\n"
b'L924 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ Wow\n'
b"L872 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ Okay -- you're gonna need to learn how to lie.\n"
b'L871 +++$+++ u2 +++$+++ m0 +++$+++ CAMERON +++$+++ No\n'
b'L870 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ I\'m kidding.  You know how sometimes you just become this "persona"?  And you don\'t know how to quit?\n'
b'L869 +++$+++ u0 +++$+++ m0 +++$+++ BIANCA +++$+++ Like my fear of wearing pastels?\n'

data/cornell movie-dialogs corpus$ head movie_conversations.txt 
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L194', 'L195', 'L196', 'L197']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L198', 'L199']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L200', 'L201', 'L202', 'L203']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L204', 'L205', 'L206']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L207', 'L208']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L271', 'L272', 'L273', 'L274', 'L275']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L276', 'L277']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L280', 'L281']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L363', 'L364']
u0 +++$+++ u2 +++$+++ m0 +++$+++ ['L365', 'L366']

# 把每一行都parse成一个dict，key是lineID、characterID、movieID、character和text
# 分别代表这一行的ID、人物ID、电影ID，人物名称和文本。
# 最终输出一个dict，key是lineID，value是一个dict。
# value这个dict的key是lineID、characterID、movieID、character和text
def loadLines(fileName, fields):
    lines = {}
    with open(fileName, 'r', encoding='iso-8859-1') as f:
        for line in f:
            values = line.split(" +++$+++ ")
            # 抽取fields
            lineObj = {}
            for i, field in enumerate(fields):
                lineObj[field] = values[i]
            lines[lineObj['lineID']] = lineObj
    return lines


# 根据movie_conversations.txt文件和上输出的lines，把utterance组成对话。
# 最终输出一个list，这个list的每一个元素都是一个dict，
# key分别是character1ID、character2ID、movieID和utteranceIDs。
# 分别表示这对话的第一个人物的ID，第二个的ID，电影的ID以及它包含的utteranceIDs
# 最后根据lines，还给每一行的dict增加一个key为lines，其value是个list，
# 包含所有utterance(上面得到的lines的value)
def loadConversations(fileName, lines, fields):
    conversations = []
    with open(fileName, 'r', encoding='iso-8859-1') as f:
        for line in f:
            values = line.split(" +++$+++ ")
            # 抽取fields
            convObj = {}
            for i, field in enumerate(fields):
                convObj[field] = values[i]
            # convObj["utteranceIDs"]是一个字符串，形如['L198', 'L199']
            # 我们用eval把这个字符串变成一个字符串的list。
            lineIds = eval(convObj["utteranceIDs"])
            # 根据lineIds构造一个数组，根据lineId去lines里检索出存储utterance对象。
            convObj["lines"] = []
            for lineId in lineIds:
                convObj["lines"].append(lines[lineId])
            conversations.append(convObj)
    return conversations


# 从对话中抽取句对 
# 假设一段对话包含s1,s2,s3,s4这4个utterance
# 那么会返回3个句对：s1-s2,s2-s3和s3-s4。
def extractSentencePairs(conversations):
    qa_pairs = []
    for conversation in conversations:
        # 遍历对话中的每一个句子，忽略最后一个句子，因为没有答案。
        for i in range(len(conversation["lines"]) - 1): 
            inputLine = conversation["lines"][i]["text"].strip()
            targetLine = conversation["lines"][i+1]["text"].strip()
            # 如果有空的句子就去掉 
            if inputLine and targetLine:
                qa_pairs.append([inputLine, targetLine])
    return qa_pairs

# 定义新的文件 
datafile = os.path.join(corpus, "formatted_movie_lines.txt")

delimiter = '\t'
# 对分隔符delimiter进行decode，这里对tab进行decode结果并没有变
delimiter = str(codecs.decode(delimiter, "unicode_escape"))

# 初始化dict lines，list conversations以及前面我们介绍过的field的id数组。
lines = {}
conversations = []
MOVIE_LINES_FIELDS = ["lineID", "characterID", "movieID", "character", "text"]
MOVIE_CONVERSATIONS_FIELDS = ["character1ID", "character2ID", "movieID", "utteranceIDs"]

# 首先使用loadLines函数处理movie_lines.txt 
print("\nProcessing corpus...")
lines = loadLines(os.path.join(corpus, "movie_lines.txt"), MOVIE_LINES_FIELDS)
# 接着使用loadConversations处理上一步的结果，得到conversations
print("\nLoading conversations...")
conversations = loadConversations(os.path.join(corpus, "movie_conversations.txt"),
                                  lines, MOVIE_CONVERSATIONS_FIELDS)

# 输出到一个新的csv文件
print("\nWriting newly formatted file...")
with open(datafile, 'w', encoding='utf-8') as outputfile:
    writer = csv.writer(outputfile, delimiter=delimiter, lineterminator='\n')
    # 使用extractSentencePairs从conversations里抽取句对。
    for pair in extractSentencePairs(conversations):
        writer.writerow(pair)

# 输出一些行用于检查 
print("\nSample lines from file:")
printLines(datafile)

b"Can we make this quick?  Roxanne Korrine and Andrew Barrett are having an incredibly horrendous public break- up on the quad.  Again.\tWell, I thought we'd start with pronunciation, if that's okay with you.\n"
b"Well, I thought we'd start with pronunciation, if that's okay with you.\tNot the hacking and gagging and spitting part.  Please.\n"
b"Not the hacking and gagging and spitting part.  Please.\tOkay... then how 'bout we try out some French cuisine.  Saturday?  Night?\n"
b"You're asking me out.  That's so cute. What's your name again?\tForget it.\n"
b"No, no, it's my fault -- we didn't have a proper introduction ---\tCameron.\n"
b"Cameron.\tThe thing is, Cameron -- I'm at the mercy of a particularly hideous breed of loser.  My sister.  I can't date until she does.\n"
b"The thing is, Cameron -- I'm at the mercy of a particularly hideous breed of loser.  My sister.  I can't date until she does.\tSeems like she could get a date easy enough...\n"
b'Why?\tUnsolved mystery.  She used to be really popular when she started high school, then it was just like she got sick of it or something.\n'
b"Unsolved mystery.  She used to be really popular when she started high school, then it was just like she got sick of it or something.\tThat's a shame.\n"
b'Gosh, if only we could find Kat a boyfriend...\tLet me see what I can do.\n'

# 预定义的token
PAD_token = 0  # 表示padding 
SOS_token = 1  # 句子的开始 
EOS_token = 2  # 句子的结束 

class Voc:
    def __init__(self, name):
        self.name = name
        self.trimmed = False
        self.word2index = {}
        self.word2count = {}
        self.index2word = {PAD_token: "PAD", SOS_token: "SOS", EOS_token: "EOS"}
        self.num_words = 3  # 目前有SOS, EOS, PAD这3个token。

    def addSentence(self, sentence):
        for word in sentence.split(' '):
            self.addWord(word)

    def addWord(self, word):
        if word not in self.word2index:
            self.word2index[word] = self.num_words
            self.word2count[word] = 1
            self.index2word[self.num_words] = word
            self.num_words += 1
        else:
            self.word2count[word] += 1

    # 删除频次小于min_count的token 
    def trim(self, min_count):
        if self.trimmed:
            return
        self.trimmed = True

        keep_words = []

        for k, v in self.word2count.items():
            if v >= min_count:
                keep_words.append(k)

        print('keep_words {} / {} = {:.4f}'.format(
            len(keep_words), len(self.word2index), len(keep_words) / len(self.word2index)
        ))

        # 重新构造词典 
        self.word2index = {}
        self.word2count = {}
        self.index2word = {PAD_token: "PAD", SOS_token: "SOS", EOS_token: "EOS"}
        self.num_words = 3 # Count default tokens
        
        # 重新构造后词频就没有意义了(都是1)
        for word in keep_words:
            self.addWord(word)

MAX_LENGTH = 10  # 句子最大长度是10个词(包括EOS等特殊词)

# 把Unicode字符串变成ASCII
# 参考https://stackoverflow.com/a/518232/2809427
def unicodeToAscii(s):
    return ''.join(
        c for c in unicodedata.normalize('NFD', s)
        if unicodedata.category(c) != 'Mn'
    )
 
def normalizeString(s):
    # 变成小写、去掉前后空格，然后unicode变成ascii
    s = unicodeToAscii(s.lower().strip())
    # 在标点前增加空格，这样把标点当成一个词
    s = re.sub(r"([.!?])", r" \1", s)
    # 字母和标点之外的字符都变成空格
    s = re.sub(r"[^a-zA-Z.!?]+", r" ", s)
    # 因为把不用的字符都变成空格，所以可能存在多个连续空格
    # 下面的正则替换把多个空格变成一个空格，最后去掉前后空格
    s = re.sub(r"\s+", r" ", s).strip()
    return s

# 读取问答句对并且返回Voc词典对象 
def readVocs(datafile, corpus_name):
    print("Reading lines...")
    # 文件每行读取到list lines中。 
    lines = open(datafile, encoding='utf-8').\
        read().strip().split('\n')
    # 每行用tab切分成问答两个句子，然后调用normalizeString函数进行处理。
    pairs = [[normalizeString(s) for s in l.split('\t')] for l in lines]
    voc = Voc(corpus_name)
    return voc, pairs

def filterPair(p): 
    return len(p[0].split(' ')) < MAX_LENGTH and len(p[1].split(' ')) < MAX_LENGTH

# 过滤太长的句对 
def filterPairs(pairs):
    return [pair for pair in pairs if filterPair(pair)]

# 使用上面的函数进行处理，返回Voc对象和句对的list 
def loadPrepareData(corpus, corpus_name, datafile):
    print("Start preparing training data ...")
    voc, pairs = readVocs(datafile, corpus_name)
    print("Read {!s} sentence pairs".format(len(pairs)))
    pairs = filterPairs(pairs)
    print("Trimmed to {!s} sentence pairs".format(len(pairs)))
    print("Counting words...")
    for pair in pairs:
        voc.addSentence(pair[0])
        voc.addSentence(pair[1])
    print("Counted words:", voc.num_words)
    return voc, pairs


# Load/Assemble voc and pairs
# save_dir = os.path.join("data", "save")
voc, pairs = loadPrepareData(corpus, corpus_name, datafile)
# 输出一些句对
print("\npairs:")
for pair in pairs[:10]:
    print(pair)

Start preparing training data ...
Reading lines...
Read 221282 sentence pairs
Trimmed to 64271 sentence pairs
Counting words...
Counted words: 18008

MIN_COUNT = 3    # 阈值为3


def trimRareWords(voc, pairs, MIN_COUNT):
    # 去掉voc中频次小于3的词 
    voc.trim(MIN_COUNT)
    # 保留的句对 
    keep_pairs = []
    for pair in pairs:
        input_sentence = pair[0]
        output_sentence = pair[1]
        keep_input = True
        keep_output = True
        # 检查问题
        for word in input_sentence.split(' '):
            if word not in voc.word2index:
                keep_input = False
                break
        # 检查答案
        for word in output_sentence.split(' '):
            if word not in voc.word2index:
                keep_output = False
                break

        # 如果问题和答案都只包含高频词，我们才保留这个句对
        if keep_input and keep_output:
            keep_pairs.append(pair)

    print("Trimmed from {} pairs to {}, {:.4f} of total".format(len(pairs), 
		len(keep_pairs), len(keep_pairs) / len(pairs)))
    return keep_pairs


# 实际进行处理
pairs = trimRareWords(voc, pairs, MIN_COUNT)

keep_words 7823 / 18005 = 0.4345
Trimmed from 64271 pairs to 53165, 0.8272 of total

# 把句子的词变成ID
def indexesFromSentence(voc, sentence):
    return [voc.word2index[word] for word in sentence.split(' ')] + [EOS_token]

# l是多个长度不同句子(list)，使用zip_longest padding成定长，长度为最长句子的长度。
def zeroPadding(l, fillvalue=PAD_token):
    return list(itertools.zip_longest(*l, fillvalue=fillvalue))

# l是二维的padding后的list
# 返回m和l的大小一样，如果某个位置是padding，那么值为0，否则为1
def binaryMatrix(l, value=PAD_token):
    m = []
    for i, seq in enumerate(l):
        m.append([])
        for token in seq:
            if token == PAD_token:
                m[i].append(0)
            else:
                m[i].append(1)
    return m

# 把输入句子变成ID，然后再padding，同时返回lengths这个list，标识实际长度。
# 返回的padVar是一个LongTensor，shape是(batch, max_length)，
# lengths是一个list，长度为(batch,)，表示每个句子的实际长度。
def inputVar(l, voc):
    indexes_batch = [indexesFromSentence(voc, sentence) for sentence in l]
    lengths = torch.tensor([len(indexes) for indexes in indexes_batch])
    padList = zeroPadding(indexes_batch)
    padVar = torch.LongTensor(padList)
    return padVar, lengths

# 对输出句子进行padding，然后用binaryMatrix得到每个位置是padding(0)还是非padding，
# 同时返回最大最长句子的长度(也就是padding后的长度)
# 返回值padVar是LongTensor，shape是(batch, max_target_length)
# mask是ByteTensor，shape也是(batch, max_target_length)
def outputVar(l, voc):
    indexes_batch = [indexesFromSentence(voc, sentence) for sentence in l]
    max_target_len = max([len(indexes) for indexes in indexes_batch])
    padList = zeroPadding(indexes_batch)
    mask = binaryMatrix(padList)
    mask = torch.ByteTensor(mask)
    padVar = torch.LongTensor(padList)
    return padVar, mask, max_target_len

# 处理一个batch的pair句对 
def batch2TrainData(voc, pair_batch):
    # 按照句子的长度(词数)排序
    pair_batch.sort(key=lambda x: len(x[0].split(" ")), reverse=True)
    input_batch, output_batch = [], []
    for pair in pair_batch:
        input_batch.append(pair[0])
        output_batch.append(pair[1])
    inp, lengths = inputVar(input_batch, voc)
    output, mask, max_target_len = outputVar(output_batch, voc)
    return inp, lengths, output, mask, max_target_len


# 示例
small_batch_size = 5
batches = batch2TrainData(voc, [random.choice(pairs) for _ in range(small_batch_size)])
input_variable, lengths, target_variable, mask, max_target_len = batches

print("input_variable:", input_variable)
print("lengths:", lengths)
print("target_variable:", target_variable)
print("mask:", mask)
print("max_target_len:", max_target_len)

input_variable: tensor(
       [[   92,   101,    76,    50,    34],
        [    7,   250,    37,     6,     4],
        [  123,   279,   628,     2,     2],
        [   40,    75,     4,     0,     0],
        [  359,    53,  7216,     0,     0],
        [ 2763,   217,     4,     0,     0],
        [  637,     4,     2,     0,     0],
        [    6,     2,     0,     0,     0],
        [    2,     0,     0,     0,     0]])
lengths: tensor([ 9,  8,  7,  3,  3])
target_variable: tensor(
       [[   25,    34,   404,     7,    25],
        [  283,     4,    76,    24,  1464],
        [   25,     2,    37,     4,    70],
        [   72,     0,  7217,     2,  1465],
        [  829,     0,     4,     0,     6],
        [  234,     0,     2,     0,     2],
        [    4,     0,     0,     0,     0],
        [    2,     0,     0,     0,     0]])
mask: tensor(
       [[ 1,  1,  1,  1,  1],
        [ 1,  1,  1,  1,  1],
        [ 1,  1,  1,  1,  1],
        [ 1,  0,  1,  1,  1],
        [ 1,  0,  1,  0,  1],
        [ 1,  0,  1,  0,  1],
        [ 1,  0,  0,  0,  0],
        [ 1,  0,  0,  0,  0]], dtype=torch.uint8)
max_target_len: 8

class EncoderRNN(nn.Module):
    def __init__(self, hidden_size, embedding, n_layers=1, dropout=0):
        super(EncoderRNN, self).__init__()
        self.n_layers = n_layers
        self.hidden_size = hidden_size
        self.embedding = embedding

        # 初始化GRU，这里输入和hidden大小都是hidden_size，这里假设embedding层的输出大小是hidden_size
        # 如果只有一层，那么不进行Dropout，否则使用传入的参数dropout进行GRU的Dropout。
        self.gru = nn.GRU(hidden_size, hidden_size, n_layers,
                          dropout=(0 if n_layers == 1 else dropout), bidirectional=True)

    def forward(self, input_seq, input_lengths, hidden=None):
        # 输入是(max_length, batch)，Embedding之后变成(max_length, batch, hidden_size)
        embedded = self.embedding(input_seq)
        # Pack padded batch of sequences for RNN module
        # 因为RNN(GRU)要知道实际长度，所以PyTorch提供了函数pack_padded_sequence把输入向量和长度
        # pack到一个对象PackedSequence里，这样便于使用。
        packed = torch.nn.utils.rnn.pack_padded_sequence(embedded, input_lengths)
        # 通过GRU进行forward计算，需要传入输入和隐变量
        # 如果传入的输入是一个Tensor (max_length, batch, hidden_size)
        # 那么输出outputs是(max_length, batch, hidden_size*num_directions)。
        # 第三维是hidden_size和num_directions的混合，它们实际排列顺序是num_directions在前面，
        # 因此我们可以使用outputs.view(seq_len, batch, num_directions, hidden_size)得到4维的向量。
        # 其中第三维是方向，第四位是隐状态。
        
        # 而如果输入是PackedSequence对象，那么输出outputs也是一个PackedSequence对象，我们需要用
        # 函数pad_packed_sequence把它变成shape为(max_length, batch, hidden*num_directions)的向量以及
        # 一个list，表示输出的长度，当然这个list和输入的input_lengths完全一样，因此通常我们不需要它。
        outputs, hidden = self.gru(packed, hidden)
        # 参考前面的注释，我们得到outputs为(max_length, batch, hidden*num_directions)
        outputs, _ = torch.nn.utils.rnn.pad_packed_sequence(outputs)
        # 我们需要把输出的num_directions双向的向量加起来
        # 因为outputs的第三维是先放前向的hidden_size个结果，然后再放后向的hidden_size个结果
        # 所以outputs[:, :, :self.hidden_size]得到前向的结果
        # outputs[:, :, self.hidden_size:]是后向的结果
        # 注意，如果bidirectional是False，则outputs第三维的大小就是hidden_size，
        # 这时outputs[:, : ,self.hidden_size:]是不存在的，因此也不会加上去。
        # 对Python slicing不熟的读者可以看看下面的例子：
        
        # >>> a=[1,2,3]
        # >>> a[:3]
        # [1, 2, 3]
        # >>> a[3:]
        # []
        # >>> a[:3]+a[3:]
        # [1, 2, 3]
        
        # 这样就不用写下面的代码了：
        # if bidirectional:
        #     outputs = outputs[:, :, :self.hidden_size] + outputs[:, : ,self.hidden_size:]
        outputs = outputs[:, :, :self.hidden_size] + outputs[:, : ,self.hidden_size:]
        # 返回最终的输出和最后时刻的隐状态。 
        return outputs, hidden

# Luong 注意力layer
class Attn(torch.nn.Module):
    def __init__(self, method, hidden_size):
        super(Attn, self).__init__()
        self.method = method
        if self.method not in ['dot', 'general', 'concat']:
            raise ValueError(self.method, "is not an appropriate attention method.")
        self.hidden_size = hidden_size
        if self.method == 'general':
            self.attn = torch.nn.Linear(self.hidden_size, hidden_size)
        elif self.method == 'concat':
            self.attn = torch.nn.Linear(self.hidden_size * 2, hidden_size)
            self.v = torch.nn.Parameter(torch.FloatTensor(hidden_size))

    def dot_score(self, hidden, encoder_output):
        # 输入hidden的shape是(1, batch=64, hidden_size=500)
        # encoder_outputs的shape是(input_lengths=10, batch=64, hidden_size=500)
        # hidden * encoder_output得到的shape是(10, 64, 500)，然后对第3维求和就可以计算出score。
        return torch.sum(hidden * encoder_output, dim=2)

    def general_score(self, hidden, encoder_output):
        energy = self.attn(encoder_output)
        return torch.sum(hidden * energy, dim=2)

    def concat_score(self, hidden, encoder_output):
        energy = self.attn(torch.cat((hidden.expand(encoder_output.size(0), -1, -1), 
				      encoder_output), 2)).tanh()
        return torch.sum(self.v * energy, dim=2)
    
    # 输入是上一个时刻的隐状态hidden和所有时刻的Encoder的输出encoder_outputs
    # 输出是注意力的概率，也就是长度为input_lengths的向量，它的和加起来是1。
    def forward(self, hidden, encoder_outputs):
        # 计算注意力的score，输入hidden的shape是(1, batch=64, hidden_size=500)，
        # 表示t时刻batch数据的隐状态
        # encoder_outputs的shape是(input_lengths=10, batch=64, hidden_size=500) 
        if self.method == 'general':
            attn_energies = self.general_score(hidden, encoder_outputs)
        elif self.method == 'concat':
            attn_energies = self.concat_score(hidden, encoder_outputs)
        elif self.method == 'dot':
            # 计算内积，参考dot_score函数
            attn_energies = self.dot_score(hidden, encoder_outputs)

        # Transpose max_length and batch_size dimensions
        # 把attn_energies从(max_length=10, batch=64)转置成(64, 10)
        attn_energies = attn_energies.t()

        # 使用softmax函数把score变成概率，shape仍然是(64, 10)，然后用unsqueeze(1)变成
        # (64, 1, 10) 
        return F.softmax(attn_energies, dim=1).unsqueeze(1)

class LuongAttnDecoderRNN(nn.Module):
    def __init__(self, attn_model, embedding, hidden_size, output_size, n_layers=1, dropout=0.1):
        super(LuongAttnDecoderRNN, self).__init__()

        # 保存到self里，attn_model就是前面定义的Attn类的对象。
        self.attn_model = attn_model
        self.hidden_size = hidden_size
        self.output_size = output_size
        self.n_layers = n_layers
        self.dropout = dropout

        # 定义Decoder的layers
        self.embedding = embedding
        self.embedding_dropout = nn.Dropout(dropout)
        self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=(0 if n_layers == 1 else dropout))
        self.concat = nn.Linear(hidden_size * 2, hidden_size)
        self.out = nn.Linear(hidden_size, output_size)

        self.attn = Attn(attn_model, hidden_size)

    def forward(self, input_step, last_hidden, encoder_outputs):
        # 注意：decoder每一步只能处理一个时刻的数据，因为t时刻计算完了才能计算t+1时刻。
        # input_step的shape是(1, 64)，64是batch，1是当前输入的词ID(来自上一个时刻的输出)
        # 通过embedding层变成(1, 64, 500)，然后进行dropout，shape不变。
        embedded = self.embedding(input_step)
        embedded = self.embedding_dropout(embedded)
        # 把embedded传入GRU进行forward计算
        # 得到rnn_output的shape是(1, 64, 500)
        # hidden是(2, 64, 500)，因为是两层的GRU，所以第一维是2。
        rnn_output, hidden = self.gru(embedded, last_hidden)
        # 计算注意力权重， 根据前面的分析，attn_weights的shape是(64, 1, 10)
        attn_weights = self.attn(rnn_output, encoder_outputs)
        
        # encoder_outputs是(10, 64, 500) 
        # encoder_outputs.transpose(0, 1)后的shape是(64, 10, 500)
        # attn_weights.bmm后是(64, 1, 500)
        
        # bmm是批量的矩阵乘法，第一维是batch，我们可以把attn_weights看成64个(1,10)的矩阵
        # 把encoder_outputs.transpose(0, 1)看成64个(10, 500)的矩阵
        # 那么bmm就是64个(1, 10)矩阵 x (10, 500)矩阵，最终得到(64, 1, 500)
        context = attn_weights.bmm(encoder_outputs.transpose(0, 1))
        # 把context向量和GRU的输出拼接起来
        # rnn_output从(1, 64, 500)变成(64, 500)
        rnn_output = rnn_output.squeeze(0)
        # context从(64, 1, 500)变成(64, 500)
        context = context.squeeze(1)
        # 拼接得到(64, 1000)
        concat_input = torch.cat((rnn_output, context), 1)
        # self.concat是一个矩阵(1000, 500)，
        # self.concat(concat_input)的输出是(64, 500)
        # 然后用tanh把输出返回变成(-1,1)，concat_output的shape是(64, 500)
        concat_output = torch.tanh(self.concat(concat_input))

        # out是(500, 词典大小=7826)    
        output = self.out(concat_output)
        # 用softmax变成概率，表示当前时刻输出每个词的概率。
        output = F.softmax(output, dim=1)
        # 返回 output和新的隐状态 
        return output, hidden

def maskNLLLoss(inp, target, mask):
    # 计算实际的词的个数，因为padding是0，非padding是1，因此sum就可以得到词的个数
    nTotal = mask.sum()
    
    crossEntropy = -torch.log(torch.gather(inp, 1, target.view(-1, 1)).squeeze(1))
    loss = crossEntropy.masked_select(mask).mean()
    loss = loss.to(device)
    return loss, nTotal.item()

>>> x = torch.randn(3, 4)
>>> x
tensor([[ 0.3552, -2.3825, -0.8297,  0.3477],
        [-1.2035,  1.2252,  0.5002,  0.6248],
        [ 0.1307, -2.0608,  0.1244,  2.0139]])
>>> mask = x.ge(0.5)
>>> mask
tensor([[ 0,  0,  0,  0],
        [ 0,  1,  1,  1],
        [ 0,  0,  0,  1]], dtype=torch.uint8)
>>> torch.masked_select(x, mask)
tensor([ 1.2252,  0.5002,  0.6248,  2.0139])

0.3 0.2 0.4 0.1
0.2 0.1 0.4 0.3

inp = torch.tensor([[0.3, 0.2, 0.4, 0.1], [0.2, 0.1, 0.4, 0.3]])
target = torch.tensor([2, 3])
selected = torch.gather(inp, 1, target.view(-1, 1))
print(selected)
输出：
tensor([[ 0.4000],
        [ 0.3000]])

def train(input_variable, lengths, target_variable, mask, max_target_len, encoder, decoder, embedding,
          encoder_optimizer, decoder_optimizer, batch_size, clip, max_length=MAX_LENGTH):

    # 梯度清空
    encoder_optimizer.zero_grad()
    decoder_optimizer.zero_grad()

    # 设置device，从而支持GPU，当然如果没有GPU也能工作。
    input_variable = input_variable.to(device)
    lengths = lengths.to(device)
    target_variable = target_variable.to(device)
    mask = mask.to(device)

    # 初始化变量
    loss = 0
    print_losses = []
    n_totals = 0

    # encoder的Forward计算
    encoder_outputs, encoder_hidden = encoder(input_variable, lengths)

    # Decoder的初始输入是SOS，我们需要构造(1, batch)的输入，表示第一个时刻batch个输入。
    decoder_input = torch.LongTensor([[SOS_token for _ in range(batch_size)]])
    decoder_input = decoder_input.to(device)

    # 注意：Encoder是双向的，而Decoder是单向的，因此从下往上取n_layers个
    decoder_hidden = encoder_hidden[:decoder.n_layers]

    # 确定是否teacher forcing
    use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False

    # 一次处理一个时刻 
    if use_teacher_forcing:
        for t in range(max_target_len):
            decoder_output, decoder_hidden = decoder(
                decoder_input, decoder_hidden, encoder_outputs
            )
            # Teacher forcing: 下一个时刻的输入是当前正确答案
            decoder_input = target_variable[t].view(1, -1)
            # 计算累计的loss
            mask_loss, nTotal = maskNLLLoss(decoder_output, target_variable[t], mask[t])
            loss += mask_loss
            print_losses.append(mask_loss.item() * nTotal)
            n_totals += nTotal
    else:
        for t in range(max_target_len):
            decoder_output, decoder_hidden = decoder(
                decoder_input, decoder_hidden, encoder_outputs
            )
            # 不是teacher forcing: 下一个时刻的输入是当前模型预测概率最高的值
            _, topi = decoder_output.topk(1)
            decoder_input = torch.LongTensor([[topi[i][0] for i in range(batch_size)]])
            decoder_input = decoder_input.to(device)
            # 计算累计的loss
            mask_loss, nTotal = maskNLLLoss(decoder_output, target_variable[t], mask[t])
            loss += mask_loss
            print_losses.append(mask_loss.item() * nTotal)
            n_totals += nTotal

    # 反向计算 
    loss.backward()

    # 对encoder和decoder进行梯度裁剪
    _ = torch.nn.utils.clip_grad_norm_(encoder.parameters(), clip)
    _ = torch.nn.utils.clip_grad_norm_(decoder.parameters(), clip)

    # 更新参数
    encoder_optimizer.step()
    decoder_optimizer.step()

    return sum(print_losses) / n_totals

def trainIters(model_name, voc, pairs, encoder, decoder, encoder_optimizer, decoder_optimizer, 
              embedding, encoder_n_layers, decoder_n_layers, save_dir, n_iteration, batch_size, 
              print_every, save_every, clip, corpus_name, loadFilename):

    # 随机选择n_iteration个batch的数据(pair)
    training_batches = [batch2TrainData(voc, [random.choice(pairs) for _ in range(batch_size)])
                      for _ in range(n_iteration)]

    # 初始化
    print('Initializing ...')
    start_iteration = 1
    print_loss = 0
    if loadFilename:
        start_iteration = checkpoint['iteration'] + 1

    # 训练
    print("Training...")
    for iteration in range(start_iteration, n_iteration + 1):
        training_batch = training_batches[iteration - 1]
        
        input_variable, lengths, target_variable, mask, max_target_len = training_batch

        # 训练一个batch的数据
        loss = train(input_variable, lengths, target_variable, mask, max_target_len, encoder,
                     decoder, embedding, encoder_optimizer, decoder_optimizer, batch_size, clip)
        print_loss += loss

        # 进度
        if iteration % print_every == 0:
            print_loss_avg = print_loss / print_every
            print("Iteration: {}; Percent complete: {:.1f}%; Average loss: {:.4f}"
			.format(iteration, iteration / n_iteration * 100, print_loss_avg))
            print_loss = 0

        # 保存checkpoint
        if (iteration % save_every == 0):
            directory = os.path.join(save_dir, model_name, corpus_name, '{}-{}_{}'
		.format(encoder_n_layers, decoder_n_layers, hidden_size))
            if not os.path.exists(directory):
                os.makedirs(directory)
            torch.save({
                'iteration': iteration,
                'en': encoder.state_dict(),
                'de': decoder.state_dict(),
                'en_opt': encoder_optimizer.state_dict(),
                'de_opt': decoder_optimizer.state_dict(),
                'loss': loss,
                'voc_dict': voc.__dict__,
                'embedding': embedding.state_dict()
            }, os.path.join(directory, '{}_{}.tar'.format(iteration, 'checkpoint')))

class GreedySearchDecoder(nn.Module):
    def __init__(self, encoder, decoder):
        super(GreedySearchDecoder, self).__init__()
        self.encoder = encoder
        self.decoder = decoder

    def forward(self, input_seq, input_length, max_length):
        # Encoder的Forward计算 
        encoder_outputs, encoder_hidden = self.encoder(input_seq, input_length)
        # 把Encoder最后时刻的隐状态作为Decoder的初始值
        decoder_hidden = encoder_hidden[:decoder.n_layers]
        # 因为我们的函数都是要求(time,batch)，因此即使只有一个数据，也要做出二维的。
        # Decoder的初始输入是SOS
        decoder_input = torch.ones(1, 1, device=device, dtype=torch.long) * SOS_token
        # 用于保存解码结果的tensor
        all_tokens = torch.zeros([0], device=device, dtype=torch.long)
        all_scores = torch.zeros([0], device=device)
        # 循环，这里只使用长度限制，后面处理的时候把EOS去掉了。
        for _ in range(max_length):
            # Decoder forward一步
            decoder_output, decoder_hidden = self.decoder(decoder_input, decoder_hidden, 
								encoder_outputs)
            # decoder_outputs是(batch=1, vob_size)
            # 使用max返回概率最大的词和得分
            decoder_scores, decoder_input = torch.max(decoder_output, dim=1)
            # 把解码结果保存到all_tokens和all_scores里
            all_tokens = torch.cat((all_tokens, decoder_input), dim=0)
            all_scores = torch.cat((all_scores, decoder_scores), dim=0)
            # decoder_input是当前时刻输出的词的ID，这是个一维的向量，因为max会减少一维。
            # 但是decoder要求有一个batch维度，因此用unsqueeze增加batch维度。
            decoder_input = torch.unsqueeze(decoder_input, 0)
        # 返回所有的词和得分。
        return all_tokens, all_scores

def evaluate(encoder, decoder, searcher, voc, sentence, max_length=MAX_LENGTH):
    ### 把输入的一个batch句子变成id
    indexes_batch = [indexesFromSentence(voc, sentence)]
    # 创建lengths tensor
    lengths = torch.tensor([len(indexes) for indexes in indexes_batch])
    # 转置 
    input_batch = torch.LongTensor(indexes_batch).transpose(0, 1)
    # 放到合适的设备上(比如GPU)
    input_batch = input_batch.to(device)
    lengths = lengths.to(device)
    # 用searcher解码
    tokens, scores = searcher(input_batch, lengths, max_length)
    # ID变成词。
    decoded_words = [voc.index2word[token.item()] for token in tokens]
    return decoded_words


def evaluateInput(encoder, decoder, searcher, voc):
    input_sentence = ''
    while(1):
        try:
            # 得到用户终端的输入
            input_sentence = input('> ')
            # 是否退出
            if input_sentence == 'q' or input_sentence == 'quit': break
            # 句子归一化
            input_sentence = normalizeString(input_sentence)
            # 生成响应Evaluate sentence
            output_words = evaluate(encoder, decoder, searcher, voc, input_sentence)
            # 去掉EOS后面的内容
            words = []
            for word in output_words:
                if word == 'EOS':
                    break
                elif word != 'PAD':
                    words.append(word)
            print('Bot:', ' '.join(words))

        except KeyError:
            print("Error: Encountered unknown word.")
      

# 配置模型
model_name = 'cb_model'
attn_model = 'dot'
#attn_model = 'general'
#attn_model = 'concat'
hidden_size = 500
encoder_n_layers = 2
decoder_n_layers = 2
dropout = 0.1
batch_size = 64

# 从哪个checkpoint恢复，如果是None，那么从头开始训练。
loadFilename = None
checkpoint_iter = 4000
  

# 如果loadFilename不空，则从中加载模型 
if loadFilename:
    # 如果训练和加载是一条机器，那么直接加载 
    checkpoint = torch.load(loadFilename)
    # 否则比如checkpoint是在GPU上得到的，但是我们现在又用CPU来训练或者测试，那么注释掉下面的代码
    #checkpoint = torch.load(loadFilename, map_location=torch.device('cpu'))
    encoder_sd = checkpoint['en']
    decoder_sd = checkpoint['de']
    encoder_optimizer_sd = checkpoint['en_opt']
    decoder_optimizer_sd = checkpoint['de_opt']
    embedding_sd = checkpoint['embedding']
    voc.__dict__ = checkpoint['voc_dict']


print('Building encoder and decoder ...')
# 初始化word embedding
embedding = nn.Embedding(voc.num_words, hidden_size)
if loadFilename:
    embedding.load_state_dict(embedding_sd)
# 初始化encoder和decoder模型
encoder = EncoderRNN(hidden_size, embedding, encoder_n_layers, dropout)
decoder = LuongAttnDecoderRNN(attn_model, embedding, hidden_size, voc.num_words, 
				decoder_n_layers, dropout)
if loadFilename:
    encoder.load_state_dict(encoder_sd)
    decoder.load_state_dict(decoder_sd)
# 使用合适的设备
encoder = encoder.to(device)
decoder = decoder.to(device)
print('Models built and ready to go!')

# 配置训练的超参数和优化器 
clip = 50.0
teacher_forcing_ratio = 1.0
learning_rate = 0.0001
decoder_learning_ratio = 5.0
n_iteration = 4000
print_every = 1
save_every = 500

# 设置进入训练模式，从而开启dropout 
encoder.train()
decoder.train()

# 初始化优化器 
print('Building optimizers ...')
encoder_optimizer = optim.Adam(encoder.parameters(), lr=learning_rate)
decoder_optimizer = optim.Adam(decoder.parameters(), lr=learning_rate * decoder_learning_ratio)
if loadFilename:
    encoder_optimizer.load_state_dict(encoder_optimizer_sd)
    decoder_optimizer.load_state_dict(decoder_optimizer_sd)

# 开始训练
print("Starting Training!")
trainIters(model_name, voc, pairs, encoder, decoder, encoder_optimizer, decoder_optimizer,
           embedding, encoder_n_layers, decoder_n_layers, save_dir, n_iteration, batch_size,
           print_every, save_every, clip, corpus_name, loadFilename)

# 进入eval模式，从而去掉dropout。 
encoder.eval()
decoder.eval()

# 构造searcher对象 
searcher = GreedySearchDecoder(encoder, decoder)

# 测试
evaluateInput(encoder, decoder, searcher, voc)

> hello
Bot: hello .
> what's your name?
Bot: jacob .
> I am sorry.
Bot: you re not .
> where are you from?
Bot: southern .
> q