Yonglong Tian
Yonglong Tian
currently `syncBN` is not working, and actually you can remove all the code related to `apex`, sry about this.
@amsword, excellent question. But I don't think this is a fairness issue. IIRC, the baseline you proposed here is sub-optimal compared to end-to-end cross-entropy training. In theory it also should...
`mask.sum(1)` should be at least 1, if you are feeding features of two views of the same image following [here](https://github.com/HobbitLong/SupContrast/blob/master/losses.py#L27).
@ml-illustrated , Thank you for your nice input, I like your modification a lot! Just wonder would it be possible to make it compatible with both torchvision 0.4.2 and previous...
Good catch! You are right, @haozhaopop . Please either follow `test.py` to render shapes or calling for `render_block` function.
Hi, I might not be able to clean and release that part, but it's exactly the same implementation (loss-wise).
Maybe this [snippets](https://github.com/HobbitLong/shape2prog/blob/master/test.py#L162-L173) help?
Hey, sry for the late response. For the rand init. experiments, I forgot all the details since it's long time ago. But giving that rand int. has nothing to do...
For most of the methods, I used the hyper-parameter that was used by original authors. While I suppose authors of each paper have optimized for those parameters, it also might...
Check this line [here](https://github.com/HobbitLong/SupContrast/blob/master/losses.py#L89).