【论文复现】使用高层api进行训练时acc达不到底层api实现的效果 - Python Paddle

已有作者用使paddle复现成功论文，但该作者是使用底层api实现的。 我在使用高层api对该项目进行复现时，遇到了loss很难收敛，并且acc达不到底层api实现的效果的情况 数据处理方式、优化器与超参等均相同 这是原作者实现的过程，5次epoch后acc达到79.97%

# 训练
def training():
    global save_every
    model.train()
    if not os.path.exists(save_dir):
        os.makedirs(save_dir)
    best_acc=0
    losses = []
    for epoch in range(epochs):
        if epoch > 180:
            save_every = 2
        # train for one epoch
        start = time.time()
        # model.train()
        for _iter,(x,y) in enumerate(train_loader):
            y = paddle.reshape(y, (-1, 1))
            loss = loss_fn(model(x), y)
            loss.backward()
            opt.step()
            opt.clear_grad()
            losses.append(np.mean(loss.numpy()))
            if _iter % 10 == 0:
                print('iter:%d  loss:%.4f'%(_iter,np.mean(losses)))
        print('Time per epoch:%.2f,loss:%.4f'%(time.time()-start,np.mean(losses)))

        if (epoch+1)%save_every==0 or epoch+1==epochs:
            # evaluate on validation set
            eval_acc,eval_loss = test()
            print("Validation accuracy/loss: %.2f%%,%.4f"%(eval_acc, eval_loss))
            model.train()
            paddle.save(model.state_dict(),os.path.join(save_dir, 'checkpoint_{}.pdparams'.format(epoch)))
            paddle.save(opt.state_dict(),os.path.join(save_dir, 'checkpoint_{}.pdopt'.format(epoch)))
            if eval_acc > best_acc:
                paddle.save(model.state_dict(),os.path.join(save_dir, 'checkpoint.pdparams'))
                paddle.save(opt.state_dict(),os.path.join(save_dir, 'checkpoint.pdopt')) 
            best_acc = max(eval_acc, best_acc)      
        scheduler.step()
    paddle.save(model.state_dict(),os.path.join(save_dir, 'model.pdparams'))
    paddle.save(opt.state_dict(),os.path.join(save_dir, 'model.pdopt'))         
    print('Best accuracy on validation dataset: %.2f%%'%(best_acc))

def test():
    model.eval()
    accuracies = []
    losses = []
    for (x,y) in val_loader:
        with paddle.no_grad():
            logits = model(x)
            y = paddle.reshape(y, (-1, 1))
            loss = loss_fn(logits, y)
            acc = acc_fn(logits, y)
            accuracies.append(np.mean(acc.numpy()))
            losses.append(np.mean(loss.numpy()))
    return np.mean(accuracies)*100, np.mean(losses) 

import warnings

warnings.filterwarnings("ignore", category=Warning)
model = WideResNet(28,10,20,0.3)

epochs = 400
save_every = 1
loss_fn = paddle.nn.CrossEntropyLoss()
acc_fn = paddle.metric.accuracy
scheduler=paddle.optimizer.lr.PiecewiseDecay(boundaries=[60,120,160,200,240,260,280],values=[0.05,0.01,0.002,0.0004,0.0002,0.0001,0.00005],verbose=True)
opt = paddle.optimizer.Momentum(parameters=model.parameters(), learning_rate=scheduler, momentum=0.9,weight_decay=0.0005)
save_dir = '/home/aistudio/models/cifar10/ResNet_wide'

training()

这是我使用高层api实现的过程，5次epoch后acc仅为48.01%

import math
from wide_resnet import WideResNet
model = paddle.Model(WideResNet(28,10,20,0.3))

loss_fn = paddle.nn.CrossEntropyLoss()
acc_fn =  paddle.metric.Accuracy()
scheduler=paddle.optimizer.lr.PiecewiseDecay(boundaries=[60,120,160,200,240,260,280],values=[0.05,0.01,0.002,0.0004,0.0002,0.0001,0.00005],verbose=False)

opt = paddle.optimizer.Momentum(parameters=model.parameters(), learning_rate=scheduler, momentum=0.9,weight_decay=0.0005)
model.prepare(opt,loss_fn,acc_fn)
model.fit(train_loader,            # 训练数据集
          val_loader,            # 评估数据集
          epochs=400,            # 总的训练轮次
          save_freq=5,
          verbose=1,                # 日志展示格式
          save_dir='./chk_points/',)

以下为aistudio项目 https://aistudio.baidu.com/aistudio/projectdetail/2304027?shared=1

Asked Oct 10 '21 15:10

valium123