Incomplete logic in handling input shape about efficientnet

[bfialkoff]

This arose because I tried to build a Unet with b5 backbone, and got a shape error. Is this why the b4 and b5 models haven't been updated on segmentation-models master?

[qubvel]

It should be ok:

1. (None None 3) shape is ok for segmentation models as they are fylly-conv and dont require input size - so this is expected behavior
2. It is not updated cause I did not prepare release for efficient net (I was waiting for b6/7 weights)
3. If you install efficientnet from master and use segmentation-models from b4-b5 brach should work ok, but did not test it

[bfialkoff]

1. I understand that its a valid input for segmentation models, but it causes the efficientnet backbone to not fall back onto its default input. For example the B5's default input is (456, 456, 3) and if you call Unet with the b5 backbone without specifying the shape, it falls back to (224, 224, 3) instead of the expected (456, 456, 3).

2. I have done so, in fact it does not work. See my comment.

[qubvel]

Sorry if I dont catch your logic.
From my point of veiw could be next cases:

• input_shape = None -> go to default shape
• input_shape = (None, None, 3) and include_top = False -> ok
• input_shape = (h, w, c) -> ok
  else:
  raise error

If I replace with such statement if input_shape is None or None in input_shape it would be not possible to make model with input shape (None, None, 3)

[bfialkoff]

Could you clarify what is include_top?

It was my understanding that
input_shape = (None, None, 3) -> default_shape?

Am I missing something?
What do you say about the B5?

[qubvel]

If you specify include_top=False parameter the model will not build pooling and dense layers on top.
So there will be no error for input_shape=(None, None, 3) and it will be possible to pass images of different size (useful case for semantic segmentation).