Defining conv2 Cin size dynamically

1

Hi. I want to have one layer of Con2d with which the C_in size is defined after some processing in the forward method and is not mentioned in advanced. How can I do this since I need to declare in my __init__ method the fix size for C_in. is there any way like tensorflow which we can mention Noun which means the size can be mentioned dynamically during run time?

2

You could use the functional API to define your parameters in the forward method.
Here is a small example using a random number of kernels for the conv layer:

class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.conv_weight = None
        self.conv_bias = None
        
    def forward(self, x):
        if self.conv_weight is None:
            nb_kernels = torch.randint(1, 10, (1,))
            self.conv_weight = nn.Parameter(torch.randn(nb_kernels, x.size(1), 3, 3))
            self.conv_bias = nn.Parameter(torch.randn(nb_kernels))

        x = F.conv2d(x, self.conv_weight, self.conv_bias, stride=1, padding=1)
        return x


model = MyModel()
x = torch.randn(1, 3, 24, 24)
output = model(x)
output.mean().backward()

print(model.conv_weight.grad)
print(list(model.parameters()))
1 Like