Weight Sharing between Custom Convs

1

Hello,

For a Project I am using the Sparsity-Invariant Convolution which is implemented in a class. The dilation rate is an input into this class, so would have to be set when initialized.
Within a Model-Block, I want to use this convolution 3 times in parallel, with differing dilation rates. Is it possible to lock the weights between these 3 convolutions so they share the same weights.

This weight sharing would be like DetectoRS https://arxiv.org/abs/2006.02334

Thank you in advance for your help.

2

Yes, weight sharing is possible and a simple approach would be to use the functional API by defining the filters and bias as nn.Parameters and apply them via F.conv2d using the different dilation rates.

3

Thank you for your answer. Would this be a correct implementation of shared weights with a common gradient?

Convolution In init:

self.conv = nn.Conv2d(in_channels, out_channels, kernel_size,
                              stride=stride, padding=padding,
                              groups=groups, dilation=dilation, bias=False)

Share convolution in forward:

shared_weight = nn.Parameter((self.conv.weight))
shared_bias = self.conv.bias
x_d = F.conv2d(x_d, shared_weight, bias=shared_bias, stride = self.conv.stride, 
                          padding=self.switch_dilation, dilation=self.switch_dilation)
4

Don’t re-wrap the parameters in a new nn.Parameter, but use them directly via F.conv2d(x_d, self.conv.weight, bias=self.conv.bias, ...).

5

Thanks a lot :slight_smile: