fix #1824

[fsx950223]

Sometimes miss Tensor shape in Graph mode.
Set shape explicitly.

[bhack]

Does def cutout( require @tf.function decoration?

[bhack]

@fsx950223 Do you have a minimal code to reproduce #1824?

Edit: I am a little bit worried about "sometimes miss.."

[fsx950223]

Since Dataset runs map function in Graph mode. I won't add a test code.
You could try with reporter's minimal code.
Batch Size is None(Tensor shape is (None, 224, 224, 3)) after call cutout and after call tf.squeeze the Tensor shape misses(Tensor shape is None or unknown).
I have to explicitly set shape(Tensor shape is (1, 224, 224, 3)).

[bhack]

Yes I saw that his squeeze example the Tensor shape misses.
Have you seen: #1824 (comment)

[fsx950223]

Yes I saw that his squeeze example the Tensor shape misses.
Have you seen: #1824 (comment)

It won't change anything.
Maybe it's useful to users but not ops.
For details:
#1338 (comment)

[bhack]

It won't change anything.
Maybe it's useful to users but not ops.

This why I asked to you before what do you mean with "sometimes".
Do you mean that Is the case when an user need to use tf.squeeze over the Op result? Or that tf.squeeze Is not working correctly?

[fsx950223]

It won't change anything.
Maybe it's useful to users but not ops.

This why I asked to you before what do you mean with "sometimes".
Do you mean that Is the case when an user need to use tf.squeeze over the Op result? Or that tf.squeeze Is not working correctly?

We don't support 3D and 2D images after we discussed at #1338 (comment) and data_format too.
We have spent lots of time on it.
But now, we have to change it back?

[fsx950223]

How do you process images with batch_size>1 without set shape again?

[bhack]

Ok for 3D and 2D images I've not followed that review but just the new thread at #1825

[WindQAQ]

Sorry for the late reply. I believe this is a right fix so I am going to merge this first. Let's open other issues for further discussion. Thank you all!