First of all, thank you for your code, which has benefited me a lot, but the same error always occurs when loading the pre-training model. The error code is as follows Error(s) in loading state_dict for OctResNet:Missing key(s) in state_dict: Can you help me with it, or can you provide me with a full version of the code? I am grateful! My email is [email protected]

Why is original resnet50 faster than octave-resnet50? In inference on gtx-1080, original resnet50 is about 70 fps and octave-resnet50 is around 42 fps.

Hi, when apply depthwise seperate conv, you directly abandon low-to-high and high-to-low branches. But in the paper, author compress high-to-high branch and preserve communication between high and low branch.

Thanks for your implementation of OctConv!

I see that the OctConv paper reports reduced GFLOPs usage as well as reduced computation time per image. However, running your implementation on ImageNet dataset, it seems the computation time per image actually increases by a multiple of 3! This happens both during training and during validation. I am comparing PyTorch's ResNet50 vs your OctResNet-50 with default alpha parameters.

cifar100 original_resnet26 77% oct_resnet26 65% my training script is like this: https://github.com/uoguelph-mlrg/Cutout/blob/master/train.py could you shared your training script with us?? thank you.

Are you planning on releasing the training code?

Also, did you try to implement the ResNet BasicBlock with OctConv?

I'm trying to do it in my own implementation, but it is tricky due to the lack of down sampling on the first layer.

def forward(self, x):
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)

    x_h, x_l = self.layer1(x)
    x_h, x_l = self.layer2((x_h,x_l))
    x_h, x_l = self.layer3((x_h,x_l))
    x_h, x_l = self.layer4((x_h,x_l))
    x = self.avgpool(x_h)
    x = x.view(x.size(0), -1)
    x = self.fc(x)

    return x

The alpha_in value should not be set as zero for the output layer. If alpha_in=0 for the output layer, the low frequency feature (X_L) will be abandoned and cannot be merged into the final output (Y_H). The alpha_in can only be set as zero for the input layer. The paper author has claimed the aforementioned opinion in Figure 2.

I ran through the code and i couldnt find any snippet for initialisation of layers. do the layers need initialisation ?

and when "pretrain" parameter is set to TRUE, then nothing happens. Do we need to download the weights from somewhere ?

I don't see a main run script to run the training and evaluation for benchamarking. Seems you have create a module for octave convolution to plug into a custom written main script for benchmarking.

Thanks for your implementation of this work. I‘m using your code to reproduce the result you reported and I'm wondering if I could refer to your training settings. Could you please kindly provide your training config files if possible?

Thanks for your implementation of the Octave Conv paper.

I have a remark/question about the Conv_BN_ACT module. As BN after ACT sometimes makes more sense, following the PyTorch example I made a small OctaveCNN (each using 6 convs total) for the CIFAR10 dataset. Using PReLU activations, CrossEntropy loss, AmsGrad optimizer and alpha=0.25. After some experimentation with using BatchNorm before or after the activation I found the following results:

| Network description | # epochs | accuracy (%) | training loss | test loss | | --- | --- | --- | --- | --- | | Conv_BN_ACT | 15 | 78.46 | 0.7093 | 0.6362 | | | 30 | 82.20 | 0.4613 | 0.5456 | | Conv_ACT_BN | 15 | 82.84 | 0.3917 | 0.5260 | | | 30 | 84.18 | 0.1614 | 0.6036 |

I observe that Conv_ACT_BN has a tendency to overfit more as its training loss is noticeably lower than testing loss when compared to those of Conv_BN_ACT. However, Conv_ACT_BN does have a much higher accuracy.

Have you looked at this before? Is this the reason why you choose to include Conv_BN_ACT and not Conv_ACT_BN?