LambdaResNet50 in comparison to a normal ResNet50 from my testing is 7x slower. I am training on ImageNet, and I am comparing your LambdaResNet50 to a torch.hub ResNet50.
I have run the code as is, and for the default LambdaReset18 model
I am seeing accuracy of only 71.39% on the test dataset (which 96% on the training dataset)
Could you share the model for LambdaReset18
in line 44 lambda_p = F.conv3d(values, self.embedding, padding=(0, self.padding, self.padding)) in model.py
Is there any reason using conv3d?
Can it be implemented using conv2d?