Why did you compare RegNet with your EfficientNet results instead of original EfficientNet results from the paper?

[AlexeyAB]

1. Why did you compare RegNet with your EfficientNet results instead of original EfficientNet results from their paper https://arxiv.org/abs/1905.11946 Table 2?

2. Why you didn't use enhancements (DropPath, more epochs, RMSProp, AutoAugment, ...) from Table 7 RegNet-paper for training RegNet?

• EfficientNet-paper https://arxiv.org/abs/1905.11946 Table 2: EfficientNet-B5 83.6% Top1
• RegNet-paper https://arxiv.org/abs/2003.13678 Table 4: EfficientNet-B5 78.5% Top1
• RegNet-model zoo facebookresearch pycls url: EfficientNet-B5 78.3% Top1

[ir413]

Hi @AlexeyAB, thanks for the questions. Please find the answers below.

Re 1: Our focus in the paper was on evaluating network architectures under simple training settings without training enhancements. Please see Section 5 of the RegNet paper for discussion.

Re 2: As mentioned above, our focus was on simple settings and we have not explored training RegNet models with enhancements. See also #112 for a related discussion.

Re bullet points:

• Original EN numbers at the time of publishing are shown grayed out in Table 4
• EN numbers in Table 4 are our reproductions under simple settings (see Figure 22; averaged across 5 runs)
• The models provided in the model zoo are from the runs with errors closest to the average of 5 runs