self.rnn = nn.GRU(
    input_size = 20,
    hidden_size = 50,
    num_layers = 1,
    batch_first = True,
    bidirectional = True
)

>>> import torch
>>> from torch import nn
>>> x = torch.zeros((3, 3, 1))
>>> lengths = torch.tensor([1, 2, 3])
>>> rnn = torch.nn.GRU(input_size=1, hidden_size=1, batch_first=True, bidirectional=True)
>>> packed = nn.utils.rnn.pack_padded_sequence(x, lengths, batch_first=True, enforce_sorted=False)
>>> output, hidden = rnn(packed)
>>> unpacked, _ = torch.nn.utils.rnn.pad_packed_sequence(output, batch_first=True)
>>> unpacked
tensor([[[0.2916, 0.2377],
         [0.0000, 0.0000],
         [0.0000, 0.0000]],

        [[0.2916, 0.2239],
         [0.3949, 0.2377],
         [0.0000, 0.0000]],

        [[0.2916, 0.2243],
         [0.3949, 0.2239],
         [0.4263, 0.2377]]], grad_fn=<IndexSelectBackward>)

>>> hidden_size = unpacked.shape[2]//2
>>> forward_last = unpacked[:,:,:hidden_size].gather(1, (lengths - 1).reshape(-1, 1, 1).repeat(1, 1, hidden_size))
>>> forward_last
tensor([[[0.2916]],

        [[0.3949]],

        [[0.4263]]], grad_fn=<GatherBackward>)
>>> backward_first = unpacked[:,:1,hidden_size:]
>>> backward_first
tensor([[[0.2377]],

        [[0.2239]],

        [[0.2243]]], grad_fn=<SliceBackward>)
>>> combined = torch.cat((forward_last, backward_first), dim=2)
>>> combined
tensor([[[0.2916, 0.2377]],

        [[0.3949, 0.2239]],

        [[0.4263, 0.2243]]], grad_fn=<CatBackward>)
>>> combined.shape
torch.Size([3, 1, 2])

import jieba
f = open('chinese.text8', 'w')
for line in open('goods_zh.txt', 'r'):
    line = "".join(line.split(',')[:-2])
    words = list(jieba.cut(line))
    words = [w for w in words if not (w.isascii() or w in ("，", "。", "！"))]
    words.insert(0, "<BEG>")
    words.append("<EOF>")
    f.write(" ".join(words))
    f.write(" ")

import torch
from gensim.models import word2vec
sentences = word2vec.Text8Corpus('chinese.text8')
model = word2vec.Word2Vec(sentences, size=100)

v = torch.tensor(model.wv.vectors)
v1 = (v > 0).float()
model.wv.vectors = v1.numpy()

print("wv shape:", v1.shape)
print("wv unique shape:", v1.unique(dim=0).shape)

model.save("chinese.model")

import os
import sys
import torch
import gzip
import itertools
import jieba
import json
import random
from gensim.models import word2vec
from torch import nn
from matplotlib import pyplot

class MyModel(nn.Module):
    """根据上下文预测句子中的单词"""
    def __init__(self, w2v):
        super().__init__()
        self.hidden_size = 500
        self.embedded_in_size = 100
        self.embedded_out_size = 100
        self.linear_l1_size = 600
        self.linear_l2_size = 300
        self.embedding_in = nn.Embedding(
            num_embeddings=len(w2v.wv.vocab),
            embedding_dim=self.embedded_in_size,
            padding_idx=0
        )
        self.rnn = nn.LSTM(
            input_size = self.embedded_in_size,
            hidden_size = self.hidden_size,
            num_layers = 1,
            batch_first = True,
            bidirectional = True
        )
        self.linear = nn.Sequential(
            nn.Linear(in_features=self.hidden_size*2, out_features=self.linear_l1_size),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(in_features=self.linear_l1_size, out_features=self.linear_l2_size),
            nn.ReLU(),
            nn.Dropout(0.05),
            nn.Linear(in_features=self.linear_l2_size, out_features=self.embedded_out_size),
            nn.Sigmoid())

    def forward(self, x, lengths):
        # 转换单词对应的数值到输入使用的向量
        embedded_in = self.embedding_in(x)
        # 附加长度信息，避免 RNN 计算填充的数据
        packed = nn.utils.rnn.pack_padded_sequence(
            embedded_in, lengths, batch_first=True, enforce_sorted=False)
        # 使用递归模型计算，接下来的步骤需要所有输出，所以忽略最新的隐藏状态
        output, _ = self.rnn(packed)
        # output 内部会连接所有隐藏状态，shape = 实际输入数量合计, hidden_size
        # 为了接下来的处理，需要先整理 shape = batch_size, 每组的最大输入数量, hidden_size
        # 第二个返回值是各个 tensor 的实际长度，内容和 lengths 相同，所以可以省略掉
        unpacked, _ = nn.utils.rnn.pad_packed_sequence(output, batch_first=True)
        # 整理正向输出和反向输出，例如有 8 个单词，2 个填充
        # B 1 2 3 4 5 6 7 8 E 0 0
        # 0 B 1 2 3 4 5 6 7 8 E 0 (对应正向)
        # 1 2 3 4 5 6 7 8 E 0 0 0 (对应反向)
        h = self.hidden_size
        hidden_forward = torch.cat((torch.zeros(unpacked.shape[0], 1, h), unpacked[:,:,:h]), dim=1)[:,:-1,:]
        hidden_backward = torch.cat((unpacked[:,:,h:], torch.zeros(unpacked.shape[0], 1, h)), dim=1)[:,1:,:]
        hidden = torch.cat((hidden_forward, hidden_backward), dim=2)
        # 使用多层线性模型推测各个单词以接近原有句子
        y = self.linear(hidden)
        return y

    def calc_loss(self, loss_function, batch_y, predicted, batch_x_lengths):
        # 剪切 batch_y 使得维度与 predicted 相同，因为子批次的最大长度可能与批次的最大长度不一致
        batch_y = batch_y[:,:predicted.shape[1],:]
        # 根据实际长度清零头尾和填充的部分
        # 不能就地修改否则会导致 gradient computation has been modified by an inplace operation 错误
        mask = torch.ones(predicted.shape)
        for index, length in enumerate(batch_x_lengths):
            mask[index,0,:] = 0
            mask[index,length-1:,:] = 0
        predicted = predicted * mask
        batch_y = batch_y * mask
        return loss_function(predicted, batch_y)

def save_tensor(tensor, path):
    """保存 tensor 对象到文件"""
    torch.save(tensor, gzip.GzipFile(path, "wb"))

def load_tensor(path):
    """从文件读取 tensor 对象"""
    return torch.load(gzip.GzipFile(path, "rb"))

def load_word2vec_model():
    """读取 word2vec 编码库"""
    return word2vec.Word2Vec.load("chinese.model")

def prepare_save_batch(batch, pending_tensors):
    """准备训练 - 保存单个批次的数据"""
    # 打乱单个批次的数据
    random.shuffle(pending_tensors)

    # 划分输入和输出 tensor，另外保存各个输入 tensor 的长度
    in_tensor_unpadded = [p[0] for p in pending_tensors]
    in_tensor_lengths = torch.tensor([t.shape[0] for t in in_tensor_unpadded])
    out_tensor_unpadded = [p[1] for p in pending_tensors]

    # 整合长度不等的 tensor 到单个 tensor，不足的长度会填充 0
    in_tensor = nn.utils.rnn.pad_sequence(in_tensor_unpadded, batch_first=True)
    out_tensor = nn.utils.rnn.pad_sequence(out_tensor_unpadded, batch_first=True)

    # 切分训练集 (60%)，验证集 (20%) 和测试集 (20%)
    random_indices = torch.randperm(in_tensor.shape[0])
    training_indices = random_indices[:int(len(random_indices)*0.6)]
    validating_indices = random_indices[int(len(random_indices)*0.6):int(len(random_indices)*0.8):]
    testing_indices = random_indices[int(len(random_indices)*0.8):]
    training_set = (in_tensor[training_indices], in_tensor_lengths[training_indices], out_tensor[training_indices])
    validating_set = (in_tensor[validating_indices], in_tensor_lengths[validating_indices], out_tensor[validating_indices])
    testing_set = (in_tensor[testing_indices], in_tensor_lengths[testing_indices], out_tensor[testing_indices])

    # 保存到硬盘
    save_tensor(training_set, f"data/training_set.{batch}.pt")
    save_tensor(validating_set, f"data/validating_set.{batch}.pt")
    save_tensor(testing_set, f"data/testing_set.{batch}.pt")
    print(f"batch {batch} saved")

def prepare():
    """准备训练"""
    # 数据集转换到 tensor 以后会保存在 data 文件夹下
    if not os.path.isdir("data"):
        os.makedirs("data")

    # 准备词语到数值的索引
    w2v = load_word2vec_model()
    beg_index = w2v.wv.vocab["<BEG>"].index
    eof_index = w2v.wv.vocab["<EOF>"].index

    # 提前转换输出的编码
    embedding_out = nn.Embedding.from_pretrained(torch.FloatTensor(w2v.wv.vectors))

    # 从 txt 读取原始数据集，分批每次处理 2000 行
    # 这里使用原始方法读取，最后一个标注为 1 代表好评，为 0 代表差评
    batch = 0
    pending_tensors = []
    for line in open("goods_zh.txt", "r"):
        parts = line.split(',')
        phase = ",".join(parts[:-2])
        positive = int(parts[-1])
        # 使用 jieba 分词，然后转换单词到索引
        words = jieba.cut(phase)
        word_indices = [beg_index] # 代表语句开始
        for word in words:
            vocab = w2v.wv.vocab.get(word)
            if vocab:
                word_indices.append(vocab.index)
        word_indices.append(eof_index) # 代表语句结束
        if len(word_indices) <= 2:
            continue # 没有单词在编码库中
        # 输入是各个句子对应的索引值列表，输出是各个各个句子对应的向量列表
        tensor_in = torch.tensor(word_indices)
        tensor_out = embedding_out(tensor_in)
        pending_tensors.append((tensor_in, tensor_out))
        if len(pending_tensors) >= 2000:
            prepare_save_batch(batch, pending_tensors)
            batch += 1
            pending_tensors.clear()
    if pending_tensors:
        prepare_save_batch(batch, pending_tensors)
        batch += 1
        pending_tensors.clear()

def train():
    """开始训练"""
    # 创建模型实例
    w2v = load_word2vec_model()
    model = MyModel(w2v)

    # 创建损失计算器
    loss_function = torch.nn.BCELoss()

    # 创建参数调整器
    optimizer = torch.optim.Adam(model.parameters())

    # 记录训练集和验证集的正确率变化
    training_accuracy_history = []
    validating_accuracy_history = []

    # 记录最高的验证集正确率
    validating_accuracy_highest = -1
    validating_accuracy_highest_epoch = 0

    # 读取批次的工具函数
    def read_batches(base_path):
        for batch in itertools.count():
            path = f"{base_path}.{batch}.pt"
            if not os.path.isfile(path):
                break
            yield load_tensor(path)

    # 计算正确率的工具函数，除去头尾和填充值
    def calc_accuracy(actual, predicted, lengths):
        acc = 0
        for x in range(len(lengths)):
            l = lengths[x]
            predicted_record = (predicted[x][1:l-1] > 0.5).int()
            actual_record = actual[x][1:l-1].int()
            acc += (predicted_record == actual_record).sum().item() / predicted_record.numel()
        acc /= len(lengths)
        return acc
 
    # 划分输入和长度的工具函数
    def split_batch_xy(batch, begin=None, end=None):
        # shape = batch_size, input_size
        batch_x = batch[0][begin:end]
        # shape = batch_size, 1
        batch_x_lengths = batch[1][begin:end]
        # shape = batch_size. input_size, embedded_size
        batch_y = batch[2][begin:end]
        return batch_x, batch_x_lengths, batch_y

    # 开始训练过程
    for epoch in range(1, 10000):
        print(f"epoch: {epoch}")

        # 根据训练集训练并修改参数
        # 切换模型到训练模式，将会启用自动微分，批次正规化 (BatchNorm) 与 Dropout
        model.train()
        training_accuracy_list = []
        for batch_index, batch in enumerate(read_batches("data/training_set")):
            # 切分小批次，有助于泛化模型
            training_batch_accuracy_list = []
            for index in range(0, batch[0].shape[0], 100):
                # 划分输入和长度
                batch_x, batch_x_lengths, batch_y = split_batch_xy(batch, index, index+100)
                # 计算预测值
                predicted = model(batch_x, batch_x_lengths)
                # 计算损失
                loss = model.calc_loss(loss_function, batch_y, predicted, batch_x_lengths)
                # 从损失自动微分求导函数值
                loss.backward()
                # 使用参数调整器调整参数
                optimizer.step()
                # 清空导函数值
                optimizer.zero_grad()
                # 记录这一个批次的正确率，torch.no_grad 代表临时禁用自动微分功能
                with torch.no_grad():
                    training_batch_accuracy_list.append(calc_accuracy(batch_y, predicted, batch_x_lengths))
            # 输出批次正确率
            training_batch_accuracy = sum(training_batch_accuracy_list) / len(training_batch_accuracy_list)
            training_accuracy_list.append(training_batch_accuracy)
            print(f"epoch: {epoch}, batch: {batch_index}: batch accuracy: {training_batch_accuracy}")
        training_accuracy = sum(training_accuracy_list) / len(training_accuracy_list)
        training_accuracy_history.append(training_accuracy)
        print(f"training accuracy: {training_accuracy}")

        # 检查验证集
        # 切换模型到验证模式，将会禁用自动微分，批次正规化 (BatchNorm) 与 Dropout
        model.eval()
        validating_accuracy_list = []
        for batch in read_batches("data/validating_set"):
            batch_x, batch_x_lengths, batch_y = split_batch_xy(batch)
            predicted = model(batch_x, batch_x_lengths)
            validating_accuracy_list.append(calc_accuracy(batch_y, predicted, batch_x_lengths))
        validating_accuracy = sum(validating_accuracy_list) / len(validating_accuracy_list)
        validating_accuracy_history.append(validating_accuracy)
        print(f"validating accuracy: {validating_accuracy}")

        # 记录最高的验证集正确率与当时的模型状态，判断是否在 20 次训练后仍然没有刷新记录
        if validating_accuracy > validating_accuracy_highest:
            validating_accuracy_highest = validating_accuracy
            validating_accuracy_highest_epoch = epoch
            save_tensor(model.state_dict(), "model.pt")
            print("highest validating accuracy updated")
        elif epoch - validating_accuracy_highest_epoch > 20:
            # 在 20 次训练后仍然没有刷新记录，结束训练
            print("stop training because highest validating accuracy not updated in 20 epoches")
            break

    # 使用达到最高正确率时的模型状态
    print(f"highest validating accuracy: {validating_accuracy_highest}",
        f"from epoch {validating_accuracy_highest_epoch}")
    model.load_state_dict(load_tensor("model.pt"))

    # 检查测试集
    testing_accuracy_list = []
    for batch in read_batches("data/testing_set"):
        batch_x, batch_x_lengths, batch_y = split_batch_xy(batch)
        predicted = model(batch_x, batch_x_lengths)
        testing_accuracy_list.append(calc_accuracy(batch_y, predicted, batch_x_lengths))
    testing_accuracy = sum(testing_accuracy_list) / len(testing_accuracy_list)
    print(f"testing accuracy: {testing_accuracy}")

    # 显示训练集和验证集的正确率变化
    pyplot.plot(training_accuracy_history, label="training")
    pyplot.plot(validating_accuracy_history, label="validing")
    pyplot.ylim(0, 1)
    pyplot.legend()
    pyplot.show()

def eval_model():
    """使用训练好的模型"""
    # 读取 word2vec 编码库
    w2v = load_word2vec_model()

    # 创建模型实例，加载训练好的状态，然后切换到验证模式
    model = MyModel(w2v)
    model.load_state_dict(load_tensor("model.pt"))
    model.eval()

    # 获取单词索引到向量的 tensor
    embedding_tensor = torch.tensor(w2v.wv.vectors)

    # 查找最接近单词数量的函数，根据欧几里得距离比较
    # 也可以使用 w2v.wv.similar_by_vector
    def find_similar_words(target_tensor):
        top_words = 10
        similar_words = []
        for word, vocab in w2v.wv.vocab.items():
            index = vocab.index
            distance = torch.dist(embedding_tensor[index], target_tensor, 2).item()
            if len(similar_words) < top_words or distance < similar_words[-1][1]:
                similar_words.append((word, distance))
                similar_words.sort(key=lambda v: v[1])
                if len(similar_words) > top_words:
                    similar_words.pop()
        return similar_words

    # 询问输入并预测输出
    # __ 为预测目标，例如下次还来__购买 表示预测 __ 处的单词，只支持一个预测目标
    while True:
        try:
            phase = input("Sentence: ")
            phase = phase.replace("\t", "").replace("__", "\t")
            if "\t" not in phase:
                raise ValueError("Please use __ to represent predict target")
            if phase.count("\t") > 1:
                raise ValueError("Please only use one predict target")
            # 分词
            words = list(jieba.cut(phase))
            # 转换到数值列表
            word_indices = [1] # 代表语句开始
            for word in words:
                if word == '\t':
                    word_indices.append(0) # 预测目标
                    continue
                vocab = w2v.wv.vocab.get(word)
                if vocab:
                    word_indices.append(vocab.index)
            word_indices.append(2) # 代表语句结束
            if len(word_indices) <= 2:
                raise ValueError("No known words")
            # 构建输入
            x = torch.tensor(word_indices).reshape(1, -1)
            lengths = torch.tensor([len(word_indices)])
            # 预测输出
            predicted = model(x, lengths)
            # 找出最接近的单词一览
            target_index = word_indices.index(0)
            target_tensor = (predicted[0, target_index] > 0.5).float()
            similar_words = find_similar_words(target_tensor)
            for word, distance in similar_words:
                print(word, distance)
        except Exception as e:
            print("error:", e)

def main():
    """主函数"""
    if len(sys.argv) < 2:
        print(f"Please run: {sys.argv[0]} prepare|train|eval")
        exit()

    # 给随机数生成器分配一个初始值，使得每次运行都可以生成相同的随机数
    # 这是为了让过程可重现，你也可以选择不这样做
    random.seed(0)
    torch.random.manual_seed(0)

    # 根据命令行参数选择操作
    operation = sys.argv[1]
    if operation == "prepare":
        prepare()
    elif operation == "train":
        train()
    elif operation == "eval":
        eval_model()
    else:
        raise ValueError(f"Unsupported operation: {operation}")

if __name__ == "__main__":
    main()