Custom Conv2d Layer

vision
1

I am trying to define a custom CNN layer with some manipulations to weights before forward pass. Below is my implementation:

class Myconv(Linear):
    def __init__(self, in_channels, out_channels, kernel_size, iter, stride=1, padding=0, dilation=1, groups=1,  bias=True):
    super(Myconv, self).__init__(in_channels, out_channels, kernel_size)
    
    self.iter=iter
    self.params = nn.ParameterList([torch.nn.Parameter(torch.randn(out_channels), requires_grad=True) for i in range(self.iter)])
           
    
          
    def modify(self, param):
        ..........          
        return W
    
    @property
    def modifyW(self):
        W=self.modify(self.params)       
        return W   
  
    def forward(self, input):
        return F.conv2d(input, self.modifyW, self.bias)

However, during forward pass I get the error: ‘Myconv’ object has no attribute ‘modifyW’.
I implemented the same for linear layer using F.linear() and it works fine.

Any help would be appreciated!

thanks!

2

Hi,
Could it be something weird because of the missing capital in the super that use some other variable of your code? Or is that just a copy paste error?

3

@albanD sorry for the typo while posting here, but the error still occurs.

4

No problem. Weird indeed.
And if you replace the property with a function call it works fine?

5

@albanD If I remove @property then I get this error: argument ‘weight’ (position 2) must be Tensor, not method

6

Sorry I meant remove the @property and replace self.modifyW by self.modifyW().
But still it means it found the method, only the property was not found…

7

@albanD Thanks. Yes it works. But can you explain why is this the case for conv layers and not with linear layer

8

Hi,

I have no idea, I wanted you to check that to make sure it’s not a problem with code indentation or something like that that would make your definition appear as if it was outside the class.
Glad it works with a function call. But I don’t know why the property does not…

Also your conv inherit from Linear is that expected?

9

@albanD I have checked the code for indentation errors. In-fact I have just used the same class written for linear layers by replacing
F.linear in forward() to F.conv2d() and
making sure modifyW() returns the weights in correct shape

I will update here in case I could make it work with @property.

Best