Different DropConnect parameters with EfficientNet? about pycls

[ir413]

Thank you for raising this @D-X-Y. It would be great to know what the effect of using a constant DC_RATIO vs. a linearly increased one is. I would be curious to know what you found if you experimented with this.

In our experiments we found that using a constant DC_RATIO was enough to achieve EN-B0 error very close to the reported one (it may be that it matters more for larger models). However, in most of our experiments we were primarily interested in evaluating network architectures under a fixed training setup without training enhancements (e.g., DropConnect, Dropout, AutoAugment, etc.). Please see https://arxiv.org/abs/2003.13678 for more information.

[D-X-Y]

I haven't tried this experiment yet, and would post it here if I have done the experiments.
Thanks for your reply!