Terry Chen
Terry Chen
Can you try this commit 8c158f54278975a074f9cadf12982bcba259436d, it works well on my side.
IMG2IMG demo does not support v2.0 model, feel free to open a PR when you fix it.
v2.0 and v1.5 are different architectures, so we have to recompile and load the new params
Yes, BN params can be always fused into conv: https://github.com/facebookincubator/AITemplate/blob/main/examples/01_resnet-50/weight_utils.py#L59
@cypw how did you do batch normalization refine after training? Do you have plan to release this code?
Great idea! top1 accuracy is improved after this procedure? about 1%?
Ok, thanks! I will try this step on resnext.