Redundant bias paramter in ResNet's conv2d.

[sebastian-sz]

Short description In tf.keras by default Conv2D has bias enabled. We should explicitly set use_bias=False if the convolution layer is followed by BatchNormalization.

This is the done in some existing Keras-CV implementations:

• DarkNet Conv Block - link
• DenseNet - link
• ResNet V2 - link

But not in resnet_v1 - all ResNet's have bias enabled, which causes redundant computation and is not aligned with other implementations.

I think this is rather a quick fix - add use_bias=False parameter to the Conv2d's in the Block function. If you approve I could submit a PR.

[sebastian-sz]

Gently pinging @LukeWood for opinion / green light here.

[LukeWood]

Interesting, could you send a PR so we can verify this? I think you're correct, but I would like to see the updated code.

[sebastian-sz]

@LukeWood yes, submitted the PR here https://github.com/keras-team/keras-cv/pull/697