pytorch自编码器训练

Question

batchsize设置问题，设置 batch_size = 1，32，64 , 训练速度很慢 ，而且loss也很大，达到上千，
数据的shape为 torch.Size([5152, 1, 2000]) 设置为 batch_size = len(model.encoder(tain_data)) = 5152 时，
训练速度变快 loss也降低很大，但效果并不好 最好才到 loss = 2 左右.
全连接层出nn.Linear(1998, 256))，这里的256是可以人为随意设置么，还是需要与输入有对应关系？

class autoencoder(nn.Module):
    def __init__(self):
        super(autoencoder, self).__init__()
        self.encoder = nn.Sequential(
            nn.Conv1d(1, 1, 3, 1, padding=0),
            nn.ReLU(),
            nn.Linear(1998, 256))

        self.decoder = nn.Sequential(
            nn.Linear(256, 1998),
            nn.ConvTranspose1d(1, 1, 3, 1, padding=0),
            nn.Sigmoid())
    def forward(self, x):
        x = self.encoder(x)
        x = self.decoder(x)
        return x
num_epochs = 100
dataloader = DataLoader(trian_data, batch_size=len(model.encoder(trian_data)), shuffle=True)
model = autoencoder()
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.01,
                             weight_decay=1e-5)
for epoch in range(num_epochs):
   total_loss = 0
   for data in dataloader:
        x = data
        x = Variable(x)
        # ===================forward=====================
        output = model(x)
        loss = criterion(output, x)
        # ===================backward====================
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        total_loss += loss.data
     # ===================log========================
   print('epoch [{}/{}], loss:{:.4f}'
       .format(epoch+1, num_epochs, total_loss))

》》》epoch [95/100], loss:2.0771
           epoch [96/100], loss:2.0772
           epoch [97/100], loss:2.0769
           epoch [98/100], loss:2.0767
           epoch [99/100], loss:2.0769
           epoch [100/100], loss:2.0766

Answer 1

batch_size一般为32、64就可以了，不需要用到全部的数据，这样子会导致模型收敛慢。
256只是编码器的输出位数，没有特定约束，跟1998没有任何关系，可以是任意数，只需要保证跟解码器的输入是一样的就行了