Why is there no bias in collapsed wt_tensor?

[s1ddok]

Hello! Thanks for your incredible work.

I'm closely following your code and I see one weird thing. So you initialize your Conv2D using tf.keras.layers.Conv2D which according to docs has trainable bias by default. But then when you collapse the block you call the conv2d func without bias and you don't even calculate it:

out = tf.nn.conv2d(inputs, wt_tensor, strides=[1, 1, 1, 1], padding="SAME")

can you explain is it a bug or bias gets cancelled somehow during collapsing?

[s1ddok]

@kbhardwa, could you weigh in on this?

[mayujie]

Hello ! I have a similar question as above? Could you explain it? :)

[prathapsoma]

Can you suggest me after training how to test the model.

[zhihongp]

Have the same question. How is the bias handled in collapsing?