import numpy as np import torch # from torch.utils.tensorboard import SummaryWriter import torch.nn as nn import argparse from tqdm import tqdm from config import device, print_freq, vocab_size, sos_id, eos_id from data_gen import AiShellDataset, pad_collate from transformer.decoder import Decoder from transformer.encoder import Encoder from transformer.loss import cal_performance from transformer.optimizer import TransformerOptimizer from transformer.transformer import Transformer from utils import parse_args, save_checkpoint, AverageMeter, get_logger def train_net(args): torch.manual_seed(7) #定义随机种子 np.random.seed(7) checkpoint = args.checkpoint start_epoch = 0 best_loss = float('inf') # writer = SummaryWriter() epochs_since_improvement = 0 # Initialize / load checkpoint if checkpoint is None: #判断模型是否被中断过 # model encoder = Encoder(args.d_input * args.LFR_m, args.n_layers_enc, args.n_head, args.d_k, args.d_v, args.d_model, args.d_inner, dropout=args.dropout, pe_maxlen=args.pe_maxlen) decoder = Decoder(sos_id, eos_id, vocab_size, args.d_word_vec, args.n_layers_dec, args.n_head, args.d_k, args.d_v, args.d_model, args.d_inner, dropout=args.dropout, tgt_emb_prj_weight_sharing=args.tgt_emb_prj_weight_sharing, pe_maxlen=args.pe_maxlen) model = Transformer(encoder, decoder) # print(model) # model = nn.DataParallel(model) # optimizer optimizer = TransformerOptimizer( torch.optim.Adam(model.parameters(), lr=args.lr, betas=(0.9, 0.98), eps=1e-09)) #model.parameters()：可用于迭代优化的参数或者定义参数组的dicts。 #lr (float, optional) ：学习率(默认: 1e-3) #betas (Tuple[float, float], optional)：用于计算梯度的平均和平方的系数(默认: (0.9, 0.98)) #eps (float, optional)：为了提高数值稳定性而添加到分母的一个项(默认: 1e-09) #weight_decay (float, optional)：权重衰减(如L2惩罚) else: checkpoint = torch.load(checkpoint) start_epoch = checkpoint['epoch'] + 1 epochs_since_improvement = checkpoint['epochs_since_improvement'] model = checkpoint['model'] optimizer = checkpoint['optimizer'] logger = get_logger() #日志 # Move to GPU, if available model = model.to(device) # Custom dataloaders train_dataset = AiShellDataset(args, 'train') #从train路径下获取train数据，并对wav类型数据进行预处理 #对数据设置批量和填充，pin_memory=True：锁页内存（不与硬盘进行交换）；shuffle=True：打乱顺序；num_workers：工作进程数，越大批量处理越快，但加重CPU负担 train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, collate_fn=pad_collate, pin_memory=True, shuffle=True, num_workers=args.num_workers) valid_dataset = AiShellDataset(args, 'dev') valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=args.batch_size, collate_fn=pad_collate, pin_memory=True, shuffle=False, num_workers=args.num_workers) # Epochs for epoch in range(start_epoch, args.epochs): # One epoch's training train_loss = train(train_loader=train_loader, model=model, optimizer=optimizer, epoch=epoch, logger=logger) # writer.add_scalar('model/train_loss', train_loss, epoch) lr = optimizer.lr #获取学习率值 print(' Learning rate: {}'.format(lr)) # writer.add_scalar('model/learning_rate', lr, epoch) step_num = optimizer.step_num #优化器更新学习率的次数 print('Step num: {} '.format(step_num)) # One epoch's validation valid_loss = valid(valid_loader=valid_loader, model=model, logger=logger) #测试不需要优化器 # writer.add_scalar('model/valid_loss', valid_loss, epoch) # Check if there was an improvement is_best = valid_loss < best_loss #判断等式右边是否成立，成立is_best=1，否则is_best=0 best_loss = min(valid_loss, best_loss) #获得最小的loss值 if not is_best: #比较当前测试损失和以前最好的损失谁更小，并做标记 epochs_since_improvement += 1 print(" Epochs since last improvement: %d " % (epochs_since_improvement,)) else: epochs_since_improvement = 0 # Save checkpoint # 保存最小损失的数据信息 save_checkpoint(epoch, epochs_since_improvement, model, optimizer, best_loss, is_best) def train(train_loader, model, optimizer, epoch, logger): #数据训练 model.train() # train mode (dropout and batchnorm is used) 训练模式，有梯度，参数更新等 losses = AverageMeter() #损失平均值 # Batches for i, (data) in enumerate(train_loader): #train_loader中有数据，标签和length # Move to GPU, if available padded_input, padded_target, input_lengths = data padded_input = padded_input.to(device) #将输入数据放入设备中 padded_target = padded_target.to(device) input_lengths = input_lengths.to(device) # Forward prop. pred, gold = model(padded_input, input_lengths, padded_target) #将数据放入模型中训练得到预测值和目标值 loss, n_correct = cal_performance(pred, gold, smoothing=args.label_smoothing) #将目标值和预测值放入损失函数中得到损失和准确个数，smoothing（平滑正则化）：防止过拟合 # Back prop. optimizer.zero_grad() #将优化器梯度归零 loss.backward() #反向传播 # Update weights optimizer.step() #更新参数 # Keep track of metrics losses.update(loss.item()) #获得loss平均值 # Print status if i % print_freq == 0: #默认print_freq = 100,每一百个数据训练完成后日志中记录一次平均损失等信息 logger.info('Epoch: [{0}][{1}/{2}] ' 'Loss {loss.val:.5f} ({loss.avg:.5f})'.format(epoch, i, len(train_loader), loss=losses)) return losses.avg #返回损失平均值 def valid(valid_loader, model, logger): #模型预测 model.eval() #预测模式 losses = AverageMeter() # Batches for data in tqdm(valid_loader): # Move to GPU, if available padded_input, padded_target, input_lengths = data padded_input = padded_input.to(device) padded_target = padded_target.to(device) input_lengths = input_lengths.to(device) with torch.no_grad(): # Forward prop. pred, gold = model(padded_input, input_lengths, padded_target) loss, n_correct = cal_performance(pred, gold, smoothing=args.label_smoothing) # Keep track of metrics losses.update(loss.item()) # Print status logger.info(' Validation Loss {loss.val:.5f} ({loss.avg:.5f}) '.format(loss=losses)) return losses.avg def main(): global args args = parse_args() train_net(args) if __name__ == '__main__': main()