Closed NivekNey closed 5 years ago
@NivekNey, the resnet50.ResNet50
is a legacy and resnet.ResNet50
(=resnet_common.ResNet50
) is factorized over the ResNet variations and new one. But the reason why we are keeping the weight file is not to break the current usages. For example, someone, who are thresholding of softmax values of ResNet50
, may get the different results if we are replacing the resnet50.ResNet50
with resnet.ResNet50
.
Ok got it, I guess this is more a model zoo than exposed user endpoint, so it makes sense to keep both for people to choose/use. Thanks for the official reply!
These are for
include_top=False
, the weight files are not pointing to the same weight file. They also differ in terms of layer naming. Is there a reason to keep them both? Also noticedtf.keras.applications
is pointing toresnet50.ResNet50
.For
resnet50.ResNet50
: https://github.com/keras-team/keras-applications/blob/d506dc82d0/keras_applications/resnet50.py#L27-L29For
resnet.ResNet50
/resnet_common.ResNet50
: https://github.com/keras-team/keras-applications/blob/d506dc82d0/keras_applications/resnet_common.py#L40-L42