The resnet50 and 101 models available on TF keras has bias terms after every convolutional layers. It shouldn't
https://github.com/tensorflow/tensorflow/issues/37365
This MR implement both properly