Question about Axial-Res50

[lxa9867]

Thank you for sharing this great work. I have a question. Is Axial-Res50 same to the pretrained backbone you used in MaX-deeplab-S? BTW, would you like to let me know the performance of MaX-Deeplab-S & L when you pretrain it on ImageNet? I implemented both of them and got about 77% top-1 acc for S, does it seem normal? I want to sanity check whether or not my implementation is correct. Thanks.

[csrhddlam]

Thanks for your interest in our new work MaX-DeepLab.

1. No. Axial-Res50 is not MaX-deeplab-S backbone. Please see MaX-DeepLab Appendix figure A.9 for more the architecture details.
2. 77% is definitely lower than expected. Please double check if the architecture and the number of parameters match those in the MaX-DeepLab paper. I would expect around 79% and 80% accuracy for MaX-deeplab-S and L (the pretrained part) at a 224x224 resolution. One might get higher numbers if one applies strong augmentation, label smoothing or super long training schedule, but these are not used in our pre-training.
3. We are working on releasing the code for MaX-DeepLab as well. Please stay tuned.

[lxa9867]

Thank you for your reply. I found bugs in my pretraining code while some details are still hard for me to reimplement like drop path in backbone. If you already have a codebase for MaX-Deeplab, I am more than happy to help you train the models.