def main(args):
init_logger()
tokenizer = BertTokenizer.from_pretrained(
args.model_name_or_path) # 选择模型BertTokenizer执行from_pretrained
train_dataset = load_and_cache_examples(args, tokenizer, mode="train")
dev_dataset = load_and_cache_examples(args, tokenizer, mode="dev")
trainer = Trainer(args, train_dataset, dev_dataset)
if args.do_train:
trainer.train()
if args.do_eval:
trainer.load_model()
trainer.evaluate()
batch_size=config.batch_size,
lr=config.lr,
num_layers=config.num_layers)
best_valid_loss = float('inf')
target_pad_idx = trainer.trg.vocab.stoi[trainer.trg.pad_token]
criterion = nn.CrossEntropyLoss(ignore_index=target_pad_idx)
if config.mode == 'train':
for epoch in range(trainer.number_of_epochs):
start_time = time.time()
train_loss = trainer.train(criterion)
valid_loss = trainer.evaluate(trainer.valid_iterator, criterion)
end_time = time.time()
epoch_mins, epoch_secs = trainer.epoch_time(start_time, end_time)
if valid_loss < best_valid_loss:
best_valid_loss = valid_loss
torch.save(trainer.model.state_dict(), 'convseq.pt')
print('Epoch: {} | Time: {}m {}s'.format(epoch, epoch_mins,
epoch_secs))
print('Train Loss: {:7.3f} | Train PPL: {:7.3f}'.format(
train_loss, math.exp(train_loss)))
print('Val. Loss: {:7.3f} | Val. PPL: {:7.3f}'.format(
valid_loss, math.exp(valid_loss)))
download=True)
train_loader = DataLoader(train_dataset, batch_size=128, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=128)
# 加载模型
net = models.resnet18()
net.fc = nn.Linear(512, 10)
trainer = Trainer(net)
# 定义损失函数
criterion = nn.CrossEntropyLoss()
# 定义优化器
optimizer = torch.optim.SGD(trainer.get_parameters(), lr=1e-1)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
viz = Visdom(port=17000)
viz.line([[0., 0.]], [0],
win='train&eval loss',
opts={
'title': 'train&eval loss',
'legend': ['train', 'eval']
})
viz.line([0.], [0], win='accuracy', opts={'title': 'accuracy'})
for i in range(1000):
print('*' * 5 + str(i) + '*' * 5)
train_loss = trainer.train(train_loader, optimizer, criterion, device)
eval_loss, acc = trainer.evaluate(test_loader, criterion, device)
print('epoch: {} train_loss: {:.6f} eval_loss: {:.6f} acc: {:.2%}'.format(
i, train_loss, eval_loss, acc))
viz.line([[train_loss, eval_loss]], [i],
win='train&eval loss',
update='append')
viz.line([acc], [i], win='accuracy', update='append')
print('*' * 10)
})
for epoch in range(args.epochs):
# model.train()
# losses = list()
# for i, data in enumerate(train_loader):
# data = data.to(args.device)
# out = model(data)
# loss = F.nll_loss(out, data.y)
# print("Training loss:{}".format(loss.item()))
# losses.append(loss.item())
# loss.backward()
# optimizer.step()
# optimizer.zero_grad()
# val_acc,val_loss = eval(model,val_loader)
train_loss = trainer.train(train_loader, optimizer, criterion, DEVICE)
val_loss, acc = trainer.evaluate(val_loader, criterion, DEVICE)
viz.line([[train_loss, val_loss]], [epoch],
win='train&eval loss',
update='append')
viz.line([acc], [epoch], win='accuracy', update='append')
print("Validation loss:{}\taccuracy:{}".format(val_loss, acc))
if val_loss < min_loss:
torch.save(model.state_dict(), 'latest.pth')
print("Model saved at epoch{}".format(epoch))
min_loss = val_loss
patience = 0
else:
patience += 1
if patience > args.patience:
break