def train():
device = Config.device
# 准备数据
train_data, dev_data = build_dataset(Config)
train_iter = DatasetIterater(train_data, Config)
dev_iter = DatasetIterater(dev_data, Config)
model = Model().to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
# optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
# 这里我们用bertAdam优化器
optimizer = BertAdam(optimizer_grouped_parameters,
lr=Config.learning_rate,
warmup=0.05,
t_total=len(train_iter) * Config.num_epochs)
model.to(device)
model.train()
best_loss = 100000.0
for epoch in range(Config.num_epochs):
print('Epoch [{}/{}]'.format(epoch + 1, Config.num_epochs))
for step, batch in enumerate(train_iter):
input_ids, input_mask, start_positions, end_positions = \
batch[0], batch[1], batch[2], batch[3]
input_ids, input_mask, start_positions, end_positions = \
input_ids.to(device), input_mask.to(device), start_positions.to(device), end_positions.to(device)
loss, _, _ = model(input_ids, attention_mask=input_mask,
start_positions=start_positions, end_positions=end_positions)
loss.backward()
optimizer.step()
print('epoch:{}, step:{}, loss:{}')
train_loss.append(loss)
if step % 100 == 0:
eval_loss = evaluate(model, dev_iter)
if eval_loss < best_loss:
best_loss = eval_loss
torch.save(model.state_dict(), './save_model/'+'best_model')
model.train()
def train():
device = Config.device
# 准备数据
train_data, dev_data = build_dataset(Config)
train_iter = DatasetIterater(train_data, Config)
dev_iter = DatasetIterater(dev_data, Config)
model = Model().to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
# optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
# 这里我们用bertAdam优化器
optimizer = AdamW(
optimizer_grouped_parameters,
lr=Config.learning_rate,
correct_bias=False) # 要重现BertAdam特定的行为，请设置correct_bias = False
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=0.05,
num_training_steps=len(train_iter) *
Config.num_epochs) # PyTorch调度程序用法如下：
model.to(device)
model.train()
best_loss = 100000.0
for epoch in range(Config.num_epochs):
print('Epoch [{}/{}]'.format(epoch + 1, Config.num_epochs))
for step, batch in enumerate(train_iter):
start_time = time.time()
ids, input_ids, input_mask, start_positions, end_positions = \
batch[0], batch[1], batch[2], batch[3], batch[4]
input_ids, input_mask, start_positions, end_positions = \
input_ids.to(device), input_mask.to(device), start_positions.to(device), end_positions.to(device)
# print(input_ids.size())
# print(input_mask.size())
# print(start_positions.size())
# print(end_positions.size())
loss, _, _ = model(input_ids,
attention_mask=input_mask,
start_positions=start_positions,
end_positions=end_positions)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
max_grad_norm=20)
optimizer.step()
scheduler.step()
time_str = datetime.datetime.now().isoformat()
log_str = 'time:{}, epoch:{}, step:{}, loss:{:8f}, spend_time:{:6f}'.format(
time_str, epoch, step, loss,
time.time() - start_time)
rainbow(log_str)
train_loss.append(loss)
if epoch % 1 == 0:
eval_loss = valid(model, dev_iter)
if eval_loss < best_loss:
best_loss = eval_loss
torch.save(model.state_dict(),
'./save_model/' + 'best_model.bin')
model.train()