模型权重未更新,但损失正在减少

时间:2019-09-19 13:29:32

标签: python machine-learning neural-network pytorch

以下代码是使用损失|| output-input || ^ 2来训练图像尺寸为64 * 64的MLP。

由于某种原因,我的每个时期的权重并未如最后所示进行更新。

class MLP(nn.Module):
    def __init__(self, size_list):
        super(MLP, self).__init__()
        layers = []
        self.size_list = size_list
        for i in range(len(size_list) - 2):
            layers.append(nn.Linear(size_list[i],size_list[i+1]))
            layers.append(nn.ReLU())
        layers.append(nn.Linear(size_list[-2], size_list[-1]))
        self.net = nn.Sequential(*layers)

    def forward(self, x):
        return self.net(x)

model_1 = MLP([4096, 64, 4096])

对于训练每个时代:

def train_epoch(model, train_loader, criterion, optimizer):
    model.train()
    model.to(device)

running_loss = 0.0

    start_time = time.time()
    # train batch
    for batch_idx, (data) in enumerate(train_loader):   
        optimizer.zero_grad() 

        data = data.to(device)

        outputs = model(data)
        loss = criterion(outputs, data)
        running_loss += loss.item()

        loss.backward()
        optimizer.step()

    end_time = time.time()

    weight_ll = model.net[0].weight
    running_loss /= len(train_loader)

    print('Training Loss: ', running_loss, 'Time: ',end_time - start_time, 's')
    return running_loss, outputs, weight_ll

用于训练数据:

n_epochs = 20
Train_loss = []
weights=[]

criterion = nn.MSELoss()

optimizer = optim.SGD(model_1.parameters(), lr = 0.1)


for i in range(n_epochs):
    train_loss, output, weights_ll = train_epoch(model_1, trainloader, criterion, optimizer)
    Train_loss.append(train_loss)
    weights.append(weights_ll)
    print('='*20)

现在,当我按每个时代打印第一个完全连接的图层的权重时,它们的权重不会被更新。

print(weights[0][0])
print(weights[19][0])

上面的输出是(显示第0阶段和第19阶段的权重):

tensor([ 0.0086,  0.0069, -0.0048,  ..., -0.0082, -0.0115, -0.0133],
       grad_fn=<SelectBackward>)
tensor([ 0.0086,  0.0069, -0.0048,  ..., -0.0082, -0.0115, -0.0133],
       grad_fn=<SelectBackward>)

可能出了什么问题?看着我的损失,它以稳定的速度下降,但是权重没有变化。

1 个答案:

答案 0 :(得分:0)

尝试weight_ll = model.net[0].weight.clone().detach()或仅在weight_ll = model.net[0].weight.clone()函数中更改train_epoch()。您会发现权重有所不同。

说明:weights_ll始终是最后一个纪元值(如果不克隆它)。在图中它将被视为相同的张量。这就是为什么您的weights[0][0]等于weights[19][0]的原因,它们实际上是相同的张量。