Released ResNet models are not trained with the default settings

[wjfwzzc]

I download models from https://github.com/facebookresearch/pycls/blob/master/MODEL_ZOO.md#resnet-models, then use the following code:

import torch
x = torch.load("R-50-1x64d_dds_8gpu.pyth")
print(x["optimizer_state"]["param_groups"][0]["weight_decay"])

It shows 1e-4, which is different to

pycls/configs/dds_baselines/resnet/R-50-1x64d_dds_8gpu.yaml

Line 18 in 2c361ee

WEIGHT_DECAY: 5e-5

So does lr:

print(x["optimizer_state"]["param_groups"][0]["lr"])

It shows 2.467e-05, however it should be 4.934e-05 if

pycls/configs/dds_baselines/resnet/R-50-1x64d_dds_8gpu.yaml

Line 15 in 2c361ee

BASE_LR: 0.2