training procedure

train NN.py

'''
num_epochs: 运行epoch的轮数
train: 训练数据
dev: 验证集数据
evalp: 几轮进行一次验证
model: 模型
metric_best: 记录最佳的metric
metric_stop:  训练停止的metric
cnt_stop: 训练到dev几次不能超过metric_best就停止
'''
from tqdm import tqdm
pbar_epochs = tqdm(range(num_epochs))
for epoch in pbar_epochs:
    pbar_epochs.set_description("Epoch:{:d}.".format(epoch))
    pbar = tqdm(train)
    for data in pbar: 
        model.train_batch(data)    
        pbar.set_description("Loss:{:.4f}.".format(model.loss))
        pbar.refresh() 
    if (epoch+1) % evalp == 0:    
        metric = model.evaluate(dev)    
        if metric >= metric_best:
            metric_best = metric
            cnt = 0
        else:
            cnt += 1
        if metric > metric_stop: 
            break 
        if cnt > cnt_stop:
            break
    pbar_epochs.refresh()