神经网络无法为单个输入融合

如何解决神经网络无法为单个输入融合

我正在尝试复制neural network for shape and appearance disentangling。网络是用Tensorflow编写的，我想用PyTorch编写。该模型如下所示：

class Model(nn.Module):
    def __init__(self,parts=16,n_features=32):
        super(Model,self).__init__()
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.E_sigma = E(3,parts,residual_dim=64,sigma=True)
        self.E_alpha = E(1,n_features,sigma=False)
        self.decoder = Decoder(parts,n_features)

    def forward(self,x):
        sig,stack = self.E_sigma(x)
        f_xs = self.E_alpha(stack)
        alpha = get_local_part_appearances(f_xs,sig)
        mu,L_inv = get_mu_and_prec(sig,self.device)
        encoding = feat_mu_to_enc(alpha,mu,L_inv,self.device)
        reconstruction = self.decoder(encoding)
        return reconstruction

该模型由三个nn.modules组成，分别为E_sigma，E_alpha和Decoder。例如，它看起来如下：

class E(nn.Module):
    def __init__(self,depth,n_out,residual_dim,sigma=True):
        super(E,self).__init__()
        self.sigma = sigma
        self.hg = Hourglass(depth,residual_dim)  # depth 4 has bottleneck of 4x4
        self.n_out = Conv(residual_dim,kernel_size=1,stride=1,bn=True,relu=True)
        if self.sigma:
            self.preprocess_1 = Conv(3,64,kernel_size=6,stride=2,relu=True)  # transform to 64 x 64 for sigma
            self.preprocess_2 = Residual(64,residual_dim)
            self.map_transform = Conv(n_out,1,1)    # channels for addition must be increased

    def forward(self,x):
        if self.sigma:
            x = self.preprocess_1(x)
            x = self.preprocess_2(x)
        out = self.hg(x)
        map = self.n_out(out)
        if self.sigma:
            map_normalized = F.softmax(map.reshape(map.size(0),map.size(1),-1),dim=2).view_as(map)
            map_transform = self.map_transform(map_normalized)
            stack = map_transform + x   # Why not stack? x is much larger than map_transform,so it is almost no impact
            return map_normalized,stack
        else:
            return map

还有三个函数可以转换数据。我正在尝试测试，如果仅输入一个输入并让模型训练一段时间，模型是否会收敛：

def train():
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    net = Model().to(device)
    net.train()
    optimizer = torch.optim.Adam(net.parameters(),lr=1e-4)
    criterion = nn.MSELoss().to(device)
    img = torch.randn(1,3,128,128).to(device)
    for epoch in range(1000):
        optimizer.zero_grad()
        prediction = net(img)
        loss = criterion(prediction,img)
        loss.backward()
        optimizer.step()
        if epoch % 10 == 0:
            print(loss)

该模型具有大约1200万个参数，不幸的是，对于该单个输入，损耗始终收敛在大约0.5。这是否表明我的体系结构存在问题？可能是什么原因？