Has anyone seen EfficientNet train faster than resnet50 to the same accuracy? I know the paper says it can be faster for inference, and it has less params, but experiments I've seen for training so far show it train much slower. e.g. RN50 is 67% for this:https://github.com/sdoria/EfficientNet/blob/master/Imagewoof%20from%20scratch.ipynb …
I’m still early in my efforts here, but I’ve been trying to use b3 and b0 as the encoder in a slightly modified DynamicUnet. So far, it’s very sensitive to the specific resolution you use at inference compared to using resnet. Not sure if at all related but it’s interesting...
