convLSTM 大小不匹配

Question

我正在尝试只实现一个卷积 LSTM 单元并在其中传递一个张量 (1,3,128,128)。我收到大小不匹配错误。


    class ConvLSTMCell(nn.Module):
        def __init__(self, input_size, input_dim, hidden_dim, kernel_size, bias):
            """
            Parameters
            ----------
            input_size: (int, int)
                Height and width of input tensor as (height, width).
            input_dim: int
                Number of channels of input tensor.
            hidden_dim: int
                Number of channels of hidden state.
            kernel_size: (int, int)
                Size of the convolutional kernel.
            bias: bool
                Whether or not to add the bias.
            """
            super(ConvLSTMCell, self).__init__()

            self.height, self.width = input_size
            self.input_dim  = input_dim
            self.hidden_dim = hidden_dim

            self.kernel_size = kernel_size
            # self.padding     = kernel_size[0] // 2, kernel_size[1] // 2
            self.bias        = bias

            self.conv = nn.Conv2d(in_channels=self.input_dim + self.hidden_dim,
                              out_channels=4 * self.hidden_dim,
                              kernel_size=self.kernel_size,
                              #padding=self.padding,
                              bias=self.bias)

        def forward(self, input, prev_state):
            h_prev, c_prev = prev_state
            print('x: {}\nh_prev: {}\nc_prev: {}'.format(x.size(), h_prev.size(), c_prev.size()))
            combined = torch.cat((input, h_prev), dim=1) # concatenate along channel axis
            print('combined: {}'.format(combined.size()))

            combined_conv = self.conv(combined)
            print('combined_conv: {}'.format(combined_conv.size()))
            cc_i, cc_f, cc_o, cc_g = torch.split(combined_conv, self.hidden_dim, dim=1)
            print('cc_i: {}\ncc_f: {}\ncc_o: {}\ncc_g: {}'.format(cc_i.size(), cc_f.size(),   cc_o.size(), cc_g.size()))

            i = torch.sigmoid(cc_i)
            f = torch.sigmoid(cc_f)
            o = torch.sigmoid(cc_o)
            g = torch.tanh(cc_g)
            print('i: {}\nf: {}\no: {}\ng: {}'.format(i.size(), f.size(), o.size(), g.size()))

            c_cur = f * c_prev + i * g
            h_cur = o * F.tanh(c_cur)
            print('c_cur: {}\nh_cur: {}'.format(c_cur.size(), h_cur.size()))

            return h_cur, c_cur

        def init_hidden(self, batch_size):
            return (Variable(torch.zeros(batch_size, self.hidden_dim, self.height, self.width)),
                    Variable(torch.zeros(batch_size, self.hidden_dim, self.height, self.width)))



    x = torch.randn(1,3,128,128)
    model = ConvLSTMCell(input_size=(128,128), input_dim=3, hidden_dim=3, kernel_size=(5,5),
                         bias=True)
    hc = model.init_hidden(batch_size=1)
    if gpu:
        x.cuda()
        model.cuda()
        hc.cuda()

    out = model(x, hc)
    print(out.size())

我收到以下错误：

x: torch.Size([1, 3, 128, 128])

h_prev: torch.Size([1, 3, 128, 128])

c_prev: torch.Size([1, 3, 128, 128])

combined: torch.Size([1, 6, 128, 128])

combined_conv: torch.Size([1, 12, 124, 124])

cc_i: torch.Size([1, 3, 124, 124])

cc_f: torch.Size([1, 3, 124, 124])

cc_o: torch.Size([1, 3, 124, 124])

cc_g: torch.Size([1, 3, 124, 124])

i: torch.Size([1, 3, 124, 124])

f: torch.Size([1, 3, 124, 124])

o: torch.Size([1, 3, 124, 124])

g: torch.Size([1, 3, 124, 124])

Traceback (most recent call last):

File "trial.py", line 87, in

out = model(x, hc)

File "/Users/abcde/opt/anaconda3/envs/matrix/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in call

result = self.forward(*input, **kwargs)

File "trial.py", line 66, in forward

c_cur = f * c_prev + i * g

RuntimeError: The size of tensor a (124) must match the size of tensor b (128) at non-singleton >dimension 3

我希望使用它构建一个包含 17 个单元格的网络，并且我想使用每个单元格的输出来计算相对于真实情况的损失。地面实况是 18 (3,128,128) 张图像。

如何让我的网络输出相同大小的隐藏状态？

Answer 1

由于边界效应，您的输出较小 - 卷积运算仅计算内核可以完全适合输入形状的坐标处的值。最简单的解决方案是将填充应用于卷积层（您似乎已经尝试过，这有什么问题吗？）。如果您的内核大小为 5，则应填充 2，然后卷积输出的形状将与输入的形状相同。

convLSTM 大小不匹配

convLSTM size mismatch

convolution

conv-neural-network

pytorch