sru源码--language model

Posted 2020-10-17 深度学习1

import sys
import os
import argparse
import time
import random
import math

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable

import cuda_functional as MF


def read_corpus(path, eos="</s>"):
    data = [ ]
    with open(path) as fin:
        for line in fin:
            data += line.split() + [ eos ]
    return data

def create_batches(data_text, map_to_ids, batch_size, cuda=True):
    data_ids = map_to_ids(data_text)
    N = len(data_ids)
    L = ((N-1) // batch_size) * batch_size
    x = np.copy(data_ids[:L].reshape(batch_size,-1).T)
    y = np.copy(data_ids[1:L+1].reshape(batch_size,-1).T)#x和y的结果基本相同
    x, y = torch.from_numpy(x), torch.from_numpy(y)
    x, y = x.contiguous(), y.contiguous()
    if cuda:
        x, y = x.cuda(), y.cuda()
    return x, y


class EmbeddingLayer(nn.Module):#为语料中每一个单词对应的其相应的词向量
    def __init__(self, n_d, words, fix_emb=False):
        super(EmbeddingLayer, self).__init__()
        word2id = {}
        for w in words:
            if w not in word2id:
                word2id[w] = len(word2id)#把文本映射到数字上。

        self.word2id = word2id
        self.n_V, self.n_d = len(word2id), n_d#n_V应该是指词库大小，n_d指hidden state size    
        self.embedding = nn.Embedding(self.n_V, n_d)#赋予每个单词相应的词向量

    def forward(self, x):
        return self.embedding(x)

    def map_to_ids(self, text):#映射
        return np.asarray([self.word2id[x] for x in text],
                 dtype=‘int64‘
        )

class Model(nn.Module):
    def __init__(self, words, args):
        super(Model, self).__init__()
        self.args = args
        self.n_d = args.d
        self.depth = args.depth
        self.drop = nn.Dropout(args.dropout)#防止过拟合的层，变分dropout
        self.embedding_layer = EmbeddingLayer(self.n_d, words)
        self.n_V = self.embedding_layer.n_V
        if args.lstm:
            self.rnn = nn.LSTM(self.n_d, self.n_d,#self.rnn = nn.LSTM(         # if use nn.RNN(), it hardly learns
            input_size=INPUT_SIZE,
            hidden_size=64,         # rnn hidden unit
            num_layers=1,           # number of rnn layer
            batch_first=True,       # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
        )

                self.depth,
                dropout = args.rnn_dropout
            )
        else:
            self.rnn = MF.SRU(self.n_d, self.n_d, self.depth,
                dropout = args.rnn_dropout,
                rnn_dropout = args.rnn_dropout,
                use_tanh = 0
            )
        self.output_layer = nn.Linear(self.n_d, self.n_V)
        # tie weights
        self.output_layer.weight = self.embedding_layer.embedding.weight
        self.init_weights()
        if not args.lstm:
            self.rnn.set_bias(args.bias)

    def init_weights(self):#initial c
        val_range = (3.0/self.n_d)**0.5
        for p in self.parameters():
            if p.dim() > 1:  # matrix
                p.data.uniform_(-val_range, val_range)
            else:
                p.data.zero_()

    def forward(self, x, hidden):
        emb = self.drop(self.embedding_layer(x))
        output, hidden = self.rnn(emb, hidden)#rnn的输入和输出都有两个，即输入和上一层的隐层的值
        output = self.drop(output)
        output = output.view(-1, output.size(2))#改变tensor的size，size（2）表示计算第三维的大小，如size 4x6x7，则.size(3)就等于7
        output = self.output_layer(output)
        return output, hidden

    def init_hidden(self, batch_size):
        weight = next(self.parameters()).data
        zeros = Variable(weight.new(self.depth, batch_size, self.n_d).zero_())
        if self.args.lstm:
            return (zeros, zeros)
        else:
            return zeros

    def print_pnorm(self):#输出范数，范数常常被用来度量某个向量空间（或矩阵）中的每个向量的长度或大小。正则化中就是用范数
        norms = [ "{:.0f}".format(x.norm().data[0]) for x in self.parameters() ]
        sys.stdout.write("\tp_norm: {}\n".format(
            norms
        ))

def train_model(epoch, model, train):
    model.train()
    args = model.args

    unroll_size = args.unroll_size
    batch_size = args.batch_size
    N = (len(train[0])-1)//unroll_size + 1
    lr = args.lr

    total_loss = 0.0
    criterion = nn.CrossEntropyLoss(size_average=False)#每个小批次的损失将被相加。
    hidden = model.init_hidden(batch_size)
    for i in range(N):
        x = train[0][i*unroll_size:(i+1)*unroll_size]
        y = train[1][i*unroll_size:(i+1)*unroll_size].view(-1)#view(-1)是指按列展开
        x, y =  Variable(x), Variable(y)
        hidden = (Variable(hidden[0].data), Variable(hidden[1].data)) if args.lstm             else Variable(hidden.data)

        model.zero_grad()
        output, hidden = model(x, hidden)
        assert x.size(1) == batch_size
        loss = criterion(output, y) / x.size(1)#.size(1)计算列数.size(0)计算行数，must be (1. nn output, 2. target), the target label is NOT one-hotted
        loss.backward()

        torch.nn.utils.clip_grad_norm(model.parameters(), args.clip_grad)#nn.utils.clip_grad_norm()对网络进行梯度裁剪，因为RNN中容易出现梯度爆炸的问题。
        for p in model.parameters():
            if p.requires_grad:
                if args.weight_decay > 0:
                    p.data.mul_(1.0-args.weight_decay)
                p.data.add_(-lr, p.grad.data)
        if math.isnan(loss.data[0]) or math.isinf(loss.data[0]):#如果发生梯度消失或梯度爆炸则退出程序
            sys.exit(0)                                         #math.isinf(x)：如果x = ±inf（inf：infinity ，译为无穷）也就是±∞返回True
            return                                              #math.isnan(x）：如果x = Non (not a number) 返回True；

        total_loss += loss.data[0] / x.size(0)
        if i%10 == 0:
            sys.stdout.write("\r{}".format(i))
            sys.stdout.flush()

    return np.exp(total_loss/N)

def eval_model(model, valid):
    model.eval()
    args = model.args
    total_loss = 0.0
    unroll_size = model.args.unroll_size
    criterion = nn.CrossEntropyLoss(size_average=False)
    hidden = model.init_hidden(1)
    N = (len(valid[0])-1)//unroll_size + 1
    for i in range(N):
        x = valid[0][i*unroll_size:(i+1)*unroll_size]
        y = valid[1][i*unroll_size:(i+1)*unroll_size].view(-1)
        x, y = Variable(x, volatile=True), Variable(y)
        hidden = (Variable(hidden[0].data), Variable(hidden[1].data)) if args.lstm             else Variable(hidden.data)
        output, hidden = model(x, hidden)
        loss = criterion(output, y)
        total_loss += loss.data[0]
    avg_loss = total_loss / valid[1].numel()#numel()返回张量所含元素的个数
    ppl = np.exp(avg_loss)
    return ppl

def main(args):
    train = read_corpus(args.train)
    dev = read_corpus(args.dev)
    test = read_corpus(args.test)

    model = Model(train, args)
    model.cuda()
    sys.stdout.write("vocab size: {}\n".format(
        model.embedding_layer.n_V
    ))
    sys.stdout.write("num of parameters: {}\n".format(
        sum(x.numel() for x in model.parameters() if x.requires_grad)
    ))
    model.print_pnorm()
    sys.stdout.write("\n")

    map_to_ids = model.embedding_layer.map_to_ids
    train = create_batches(train, map_to_ids, args.batch_size)
    dev = create_batches(dev, map_to_ids, 1)
    test = create_batches(test, map_to_ids, 1)

    unchanged = 0
    best_dev = 1e+8
    for epoch in range(args.max_epoch):
        start_time = time.time()#返回当前时间的时间戳（1970纪元后经过的浮点秒数）。
        if args.lr_decay_epoch>0 and epoch>=args.lr_decay_epoch:
            args.lr *= args.lr_decay
        train_ppl = train_model(epoch, model, train)
        dev_ppl = eval_model(model, dev)
        sys.stdout.write("\rEpoch={}  lr={:.4f}  train_ppl={:.2f}  dev_ppl={:.2f}"
                "\t[{:.2f}m]\n".format(
            epoch,
            args.lr,
            train_ppl,
            dev_ppl,
            (time.time()-start_time)/60.0
        ))
        model.print_pnorm()
        sys.stdout.flush()

        if dev_ppl < best_dev:
            unchanged = 0
            best_dev = dev_ppl
            start_time = time.time()
            test_ppl = eval_model(model, test)
            sys.stdout.write("\t[eval]  test_ppl={:.2f}\t[{:.2f}m]\n".format(
                test_ppl,
                (time.time()-start_time)/60.0
            ))
            sys.stdout.flush()
        else:
            unchanged += 1
        if unchanged >= 30: break
        sys.stdout.write("\n")

if __name__ == "__main__":
    argparser = argparse.ArgumentParser(sys.argv[0], conflict_handler=‘resolve‘)
    argparser.add_argument("--lstm", action="store_true")
    argparser.add_argument("--train", type=str, required=True, help="training file")
    argparser.add_argument("--dev", type=str, required=True, help="dev file")
    argparser.add_argument("--test", type=str, required=True, help="test file")
    argparser.add_argument("--batch_size", "--batch", type=int, default=32)
    argparser.add_argument("--unroll_size", type=int, default=35)
    argparser.add_argument("    ", type=int, default=300)
    argparser.add_argument("--d", type=int, default=910)
    argparser.add_argument("--dropout", type=float, default=0.7,
        help="dropout of word embeddings and softmax output"
    )
    argparser.add_argument("--rnn_dropout", type=float, default=0.2,
        help="dropout of RNN layers"
    ) 
    argparser.add_argument("--bias", type=float, default=-3,
        help="intial bias of highway gates",
    )
    argparser.add_argument("--depth", type=int, default=6)
    argparser.add_argument("--lr", type=float, default=1.0)
    argparser.add_argument("--lr_decay", type=float, default=0.98)
    argparser.add_argument("--lr_decay_epoch", type=int, default=175)
    argparser.add_argument("--weight_decay", type=float, default=1e-5)
    argparser.add_argument("--clip_grad", type=float, default=5)

    args = argparser.parse_args()
    print (args)