Edit

Now I am not so sure, since a careful reading of the papers seems to show the authors believe that strided convolutions cause misaligned feature maps, which is an erroneous conclusion. I go into more detail in the linked issue, below.

Original issue text

I don't think strided convolutions receive any special treatment in the paper. I think maybe there was some confusion about the discussion of how downsampling is implemented. That refers to the H->L downsampling operation, it does not refer to the way a strided convolution layer is ported to OctConv.

Since the paper is not explicit about strided convolutions, I assume they are ported to OctConv just as any other convolutional layer. So I would just use the given stride argument in each of the convolutions, L->L, L->H, H->L, H-H.

Unless somebody finds information to the contrary, this appears to be the only sensible reading of the paper.

  X_h = X_l2h + X_h2h
RuntimeError: The size of tensor a (6) must match the size of tensor b (7) at non-singleton dimension 3

because X_l comes from X_h throgh AvgPooling , when H×W=7×7, the output is 3×3.Through upsample, the result is 6×6, not 7×7

Thank you very much. I would welcome if you can share the 1d version of OctConv as well.

In this line shouldn't we pass x to gap, otherwise if our input and output number of channels is different this will throw error, because x1 will have different number of channels than x and fc1 layers maps from number of input channels to number of output channels.

gap = self.gap(x1)

Please let me know if I am missing something.

I tried to train oct_res50 on 4 TITAN X (Pascal). It took around 0.9 s/iter, while original res50 took around 0.4~0.5 s/iter. Is it reasonable?

the test speed of the oct101 is 17fps is much slower than res101(55fps) ,is that right?

Hi, lxtGH.

Thanks for the implement, however, there is some details that confused me.

As defined under this line

X_h2h = F.conv2d(X_h, self.weights[0:end_h_y, 0:end_h_x, :,:], self.bias[0:end_h_y], 1,
                        self.padding, self.dilation, self.groups)

X_l2l = F.conv2d(X_l, self.weights[end_h_y:, end_h_x:, :,:], self.bias[end_h_y:], 1,
                        self.padding, self.dilation, self.groups)

X_h2l = F.conv2d(X_h2l, self.weights[end_h_y:, 0: end_h_x, :,:], self.bias[end_h_y:], 1,
                        self.padding, self.dilation, self.groups)

X_l2h = F.conv2d(X_l, self.weights[0:end_h_y, end_h_x:, :,:], self.bias[0:end_h_y], 1,
                        self.padding, self.dilation, self.groups)

why the calculation of X_h2h and X_l2h sharing convolution bias? Which same as X_l2l and X_h2l.

I didn't find any details about bias sharing written in the paper, is this sharing reasonable?

Thanks in advance.

https://github.com/lxtGH/OctaveConv_pytorch/blob/master/nn/AdaptiveConvResnet.py
158行

forward:3x3后的x1，in->out，通道数为out
而作为fc1的输入，通道数为in

你好，我实际应用于我的一个算法，替代resnet50，其他保持不变，但是发现训练速度明显变慢了，而且模型收敛反而更不好。不知道是和原因。

Recently, I am doing experiments on octave convolution. Should I be getting the same performance as vanilla convolution neural network by keeping the learning strategy same while training my model with octave convolution or I might have to change my learning strategy in order to get the same performance as that of vanilla convolution neural network?

首先非常感谢您的代码，让人受益匪浅，在这里提一点发现的可能的问题：在文件resnet_adaptiveconv.py的class AdaptiveConv中的代码行“weight = self.softmax(self.w)”，看起来直接对shape为torch.Size([3, 1, 64, 64])的self.w按您代码中如下softmax并没起作用，操作后weight仍然是全1的矩阵

When I used the model in libs/nn/OCtaveResnet.py, I found tow things.

One is the stride of last Bottleneck is 2, which cause bug when using this model, because the input size of last Bottleneck is 1414 and 77, and after conv block of stride=2, there will be two tensor, 66 and 77 which can not be added together, bug comes. Setting the stride of last Bottleneck equals to 1 will solve this bug, with a little change of resnet.

Another one is about stride of Octaveconv2.py. I found that when stride=2, you still used conv layer with stride=1, and implement pooing(stride=2) before ocnv layer instead, which is totally different from conv layer with stride=2. I think the latter one is what author do in original paper.

Still, thanks for sharing your work!

In the paper, the upsample operation occurs after the L->H convolution, not before.

Looking forward to your reply.

Hi @lxtGH ,

Thank you for your contribution. It would be nice if you can share the required libraries(i.e. the version of pytorch). Just tiny advice.

exp/train_val_step_se_resnet50.sh : tha last parameter name should be --warmup
libs/nn/res2net.py : line 57: missing paranthesis (generates syntax error)

@lxtGH, thanks for the great repo.

I just have a question regarding your implementation of Res2Net. Your implementation has sub-convolution along with BN & ReLU.

https://github.com/lxtGH/OctaveConv_pytorch/blob/master/libs/nn/res2net.py#L57

However, I couldn't find any mention of non-linearity in the paper so it is bit unclear to me. Have you tested the performance with & without non-linearity? I think adding BN & ReLU to every sub conv might slow down the training process.

Dear Sir, when I use nn.OCtaveResnet in image classification--input image size is [224,224,3], batch_size=72-- the shape of X_l2h doesn't match X_h2h (from OCtaveConv2--OCtaveConv).

it seems that it is something wrong with upsample. I don't know how to fix it

resnet 预训练模型如何得到？就是用以前的就可以吗？

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 ?

你好，你能方便告知一下你的octconv所运行的环境，即pytorch的版本号，torchvision的版本号，cuda的版本号吗？我一调试就提示找不到transform的模块，希望得到你的帮助