imankgoyal / nondeepnetworks Goto Github PK

I am very interested in your work and would like to further study. I hope you can release the code as soon as possible in your busy schedule. Thank you！

Can you provide some cfg files and annotation information? The annotation information and images always don't match!

Which three layers' outputs for neck fusion when the three stream is parallel.
Those three ones?
Thanks a lot！

I want to finetune the model for other classification task,could you provide a finetune model?

1. what is your shuffle code in your fusion module?
2. what is your model architecture in cifar-100? I just changed front two downsample modules based on the ParNet for Imagenet in the paper. But the accuracy is lower. And How do you set the LR, MILESTONES and NUM_EPOCH to meet high accuracy?

Could you introduce more details in parallelizing across GPUs, like how to implement through PyTorch.

您好，非常有幸阅读了贵团队的论文，感谢作者们作出的贡献，这篇文章对我有很大的启发，就是不知道会不会发布pytorch版本的？

Will I be able to see the code at the end of 2021？
I am very interested in your work, I have waited from October until now, hoping to open source，^_^

Hello. Thank you for your great study. I wonder the meaning of 'Shuffle' of fusion block in Fig. A1.
Is it pixel shuffle layer?
Please let me know the meaning of that.

Thank you.

Hello, my friend, appreciate for your great work! I have tested the code on https://github.com/Pritam-N/ParNet by Pritam-N and change the ResNet code in my model by using your ParNet , but the actual time is quite slow than the paper said. My block size is [64, 128, 256, 512, 2048], and the time of "forward()" is more than 5s average while the Resnet is 0.02s in my device. I have use the time function for every line in the forward(), find that the encode stuff is the main reason. I continue write time.perf_counter() in the encode stuff, find that the "self.stream2_fusion" and "self.stream3_fusion" is the most time user. Do you know why ?

I am very interested in your research, when will the code of the model be released? I saw on October 23rd that you said it would be released in 4 weeks

How's speed comparasion on batch size 64 or input resolution up to 800w input?

Hi. Figure 2b shows that there's one 1x1conv in a branch of SSE, how to match the channel of output by 1x1conv with the channel of input after shortcut? If I set the output channel of 1x1conv the same as input, the channels of the outputs by RepVGG block and SSE will not match.