facebookresearch / octconv Goto Github PK

If the network is downsampled n times, the input size must be 2^a (a >= n), other sizes are not supported, such as 600x1000.

I read your code and found that only not depthwise do the # High -> Low and # Low -> High.
So how to get the first data_l in this setting.

Just for reference, there is a PyTorch implementation of OctConv
https://github.com/d-li14/octconv.pytorch, including detailed training logs and pre-trained models of ResNet-50 on the ImageNet benchmark.

First of all, feel the author's open source spirit, great idea!

I know that low and high frequencies are divided according to a given proportion, but how are they divided along the channel dimension? Are they divided randomly according to proportion? Or are there other pre-processing in it?

Thank you very much.

Hi,
Thanks for sharing the code. I am just wondering if there is any implementation for the Octave Transposed Convolution (octave de-convolution)? I have not found it in your code. Is there a plan to implement it?
Thanks.

Incompatible attr in node L4_B01_conv1_conv-h at 1-th input: expected [1,96,7,7], got [1,96,6,6]

Hello @cypw. First of all, thanks to your amazing research. I had a question regarding Table 2 in your paper. This is not an issue but rather a follow up question. You reported CPU inference times of Resnet-50 for various values of alpha. I'm wondering if you had observed a similar trend of decreasing run times on GPU ?
I've replicated your implementation in Tensorflow and here are my inference times on 2080Ti for Resnet-50 imagenet model. Image size 224x224x3

From the above table, octave convolution performance that I get is worse on GPU.
Did you see a similar performance on GPU from your side? Does this suggest that Octave convolution improvement is more suitable for CPU rather than GPU?

Thank you !!

I have demonstrated the effectiveness of octave Conv on detection tasks.

Hi,
Thanks for your work.
Could you also release the code or the pre-trained model of Oct-I3D for action recognition?

Why OctConv processes low-frequency information with corresponding (low-frequency) convolutions can effectively enlarges the receptive field in the original pixel space.

Thank you very much for your work. Will the pre-trained model of the OctResNet-26 or 18 be released? Thanks.

or we need to train from scratch every time?