Will give more detail about the network?

[Windaway]

I have reimplement the paper. How ever my implement cost a lot of memory with a just resnet50 backbone, which is very strange. The paper said one P100 can hold a batchsize of 4. Can you give more detail? Such as output stride of the backbone, and the channel number of the ASPP.

By analyzing the Diagram, did resnext101 remove the maxpool layer? And use the strides [1,2,2,1] and dilations[1, 1, 1, 2] like os=16 (delete maxpool to still make output stride = 8)?

[Twice22]

Hey!

I have myself reimplemented the whole paper with some assumptions regarding some parts of the Neural Network. The code is running and actually the network is learning but then entered into failure mode with a discriminator loss of 0.

If you still work on it, maybe we could help each other ^^.

Also, for your information, I can train with batch-size of 2 on a 11GB single gpu :)

[aricsong1995]

@Twice22 Hey! i'm new and i'm also working on reimplementing the network and got stucked. Would you mind share your network modle (just the network) with me please? Thanks a lot!

[Twice22]

Well, I can help you ofc but you need to show me first that you worked on it ^^

[aricsong1995]

Hi Victor, Thanks a lot for replying to me! As I may have mentioned, I'm still a newbee to python and pytorch.This is my first time trying to re-implement a complex network. And it's really challenging for me to be honest, So, this is what I have done for the last two weeks(almost):

Reading:

ResNext Basebone: Deep Residual Learning for Image Recognition Aggregated Residual Transformations for Deep Neural Networks ASAP: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. Spatial and Channel wise Attention: SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning PatchGan: Image-to-Image Translation with Conditional Adversarial Nets Cycle-Gan

Programming: Since I spent lots of time reading, I felt overwhelmed by all those new concepts and information. So my friend suggested that I start to write and test some code for each module and then try to put them together. So from this week, I start to explore each module, I also have attached a part of the code just to show that I'm actually learning each module of HAttmatting.I'm using tensorboard to visualize each module(that i found on github) for me to get better understanding. I also found that the true understanding must be at the code, not the paper! Anyway!The biggest problem I encounter is that I still can't find any code reference for the HattMatting network, which I found frustrating to get going.Not to even mention that the paper did not cover all the network details. Since I'm new to all module concepts, it's really hard for me to put everything together. I guess it's always hard to do it the first time. But if I can successfully implement the network, I will definitely share it back with you as well or on github!

Victor, Thanks again if you could help! And If you could, it would be great if you can share the code of the network model(pytorch) with me. I;m going to read your code carefully first and then try to build mine.

Best Regards, Aric

On Thu, 3 Dec 2020 at 06:07, Victor BUSA notifications@github.com wrote:

Well, I can help you ofc but you need to show me first that you worked on it ^^

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/wukaoliu/CVPR2020-HAttMatting/issues/9#issuecomment-737871900, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKRLYK54MBFXDO2RWN3G3HDSS5WNXANCNFSM4OYTLKMQ .