shape '[-1, 2, 4, 28]' is invalid for input of size 768

Question

I'm trying to train IAM Dataset on TPSSpatialTransformerNetwork but finally I got an error: shape '[-1, 2, 4, 28]' is invalid for input of size 768

Each image in the dataset has the size of (32,128). I can not figure out the shape it got in the error step. And here is the code:

class TPS_SpatialTransformerNetwork(nn.Module):
    def __init__(self):
        super(TPS_SpatialTransformerNetwork, self).__init__()
        self.conv1 = nn.Conv2d(1, 79, kernel_size=5)
        self.conv2 = nn.Conv2d(79, 256, kernel_size=5)
        self.conv2_drop = nn.Dropout2d()
        self.fc1 = nn.Linear(256, 512)
        self.fc2 = nn.Linear(512, 79)

        # Spatial transformer localization-network
        self.localization = nn.Sequential(
            nn.Conv2d(1, 8, kernel_size=7),
            nn.MaxPool2d(2, stride=2),
            nn.ReLU(True),
            nn.Conv2d(8, 79, kernel_size=5),
            nn.MaxPool2d(2, stride=2),
            nn.ReLU(True)
        )

        # Regressor for the 3 * 2 affine matrix
        self.fc_loc = nn.Sequential(
            nn.Linear(79 * 4 * 28, 32),
            nn.ReLU(True),
            nn.Linear(32, 3 * 2)
        )

        # Initialize the weights/bias with identity transformation
        self.fc_loc[2].weight.data.zero_()
        self.fc_loc[2].bias.data.copy_(torch.tensor([1, 0, 0, 0, 1, 0], dtype=torch.float))

    # Spatial transformer network forward function
    def stn(self, x):
        xs = self.localization(x)
        xs = xs.view(-1, 79 * 4 * 28)
        theta = self.fc_loc(xs)
        theta = theta.view(-1, 2, 4,28)


        grid = F.affine_grid(theta, x.size())
        x = F.grid_sample(x, grid)

        return x

    def forward(self, x):
        # transform the input
        x = self.stn(x)

        # Perform the usual forward pass
        x = F.relu(F.max_pool2d(self.conv1(x), 2))
        x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
        x = x.view(-1, 320)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        x = self.fc2(x)
        return F.log_softmax(x, dim=1)

4 frames
/content/drive/My Drive/OCR/transformation.py in stn(self, x)
     41         xs = xs.view(-1, 79 * 4 * 28)
     42         theta = self.fc_loc(xs)
---> 43         theta = theta.view(-1, 2, 4,28)
     44 
     45 

RuntimeError: shape '[-1, 2, 4, 28]' is invalid for input of size 768

Answer 1

I did a quick search after my comment and found this link which details how to do STN in pytorch (on the official pytorch site). I don't know how you've gotten your resize command but what I suggested in my initial comment above seems correct, you're trying to resize (view) 6 features in a matrix of size [2,4,28] which will never work. As you can see below, the way that it's done on the pytorch site is:

def stn(self, x):
    xs = self.localization(x)
    xs = xs.view(-1, 10 * 3 * 3)
    theta = self.fc_loc(xs)
    theta = theta.view(-1, 2, 3) #<-----------KEY LINE HERE

    grid = F.affine_grid(theta, x.size())
    x = F.grid_sample(x, grid)
    return x

Where the theta tensor is reshaped using dimensions which correspond to the depth of 6.

The reason that the theta tensor is of size 6 is because it is provided the the self.fc_loc method, which itself is given as:

    self.fc_loc = nn.Sequential(
        nn.Linear(79 * 4 * 28, 32),
        nn.ReLU(True),
        nn.Linear(32, 3 * 2)
    )

If we look at the last line, we can see that the output of this sequential block (where each line is constructed in the graph in order, ie. SEQUENTIAL!) is a linear block with 32 inputs and 6 outputs (3*2). Therefore your theta will be of shape [-1, 6], where -1 is the a placeholder for the batch size in this bit of code.