Ning Wang

@Jiaoma I met the same problem, have you solved the problem?

@haibochina Thanks for your comments. Actually NAS is sensitive to datasets but it can also have some transfer ability since we search on VOC datasets and apply the architecture to...

We are still experimenting and improving on the searching part of the code. After that, the complete code will be open sourced.

@memoiry I am sorry , we don't have a specific timeline yet. If there are new updates, I will promptly notify you under this issue. Hope you can understand it....

@mcarilli I have one question about training process. In one epoch, I measured four part time consuming. Part 1: for _, img, tgt in enumerate(dataloader) which costs 4.18 sec Part...

@mcarilli Thank you very much ! I have used Nvidia Visual Profiler to analyse my code under your guidance. I think my problem in dataloader may be related with initialization....

@mcarilli Sorry for my late reply. I have used pin_memory=True in Pytorch dataloader. The code did not report an obvious error. However, after record_stream(), the features have become NoneType ,...

I try to split the features list and targets dictionary into single tensor. The code can run normally. However, it seems very slow when getting batch data compared with traditional...

@youngfly11 Not yet. Maybe https://zhuanlan.zhihu.com/p/80695364 can help you.