[Feature Request] tf.keras expose align_corners

[bhack]

---

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
  No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
  N/A
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
  N/A
• TensorFlow installed from (source or binary):
  N/A
• TensorFlow version (use command below):
  N/A
• Python version:
  N/A
• Bazel version (if compiling from source):
  N/A
• GCC/Compiler version (if compiling from source):
  N/A
• CUDA/cuDNN version:
  N/A
• GPU model and memory:
  N/A
• Exact command to reproduce:
  N/A

Describe the problem

We need to expose align_corners in this Keras upsample
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/layers/convolutional.py#L1935

See also keras-team/keras#10994 (comment)

[omalleyt12]

I can take a look at this, in the meantime you could always use a Lambda layer:

Lambda(
  lambda x: tf.image.resize_bilinear(x, (H, W), align_corners=True), output_shape=(H, W, C))(data))

[bhack]

/cc @Shathe

[bhack]

@omalleyt12 Do you want to pass trough upstream keras or directly in tf.keras?

[bhack]

@ewilderj A classical opensource upstream downstream issue.
We are at risk that users and devs start to duplicate bugs against tf.keras interfaces and on the upstream Keras github repo.

[ewilderj]

cc @martinwicke @fchollet for visibility. This is a reasonable point, have we seen it happen much?

[bhack]

@ewilderj I've saw this behavior but I've not annotated the tickets.
Also in theory general Keras API could not be modified troughs direct PR against tf.keras right?
So I suppose that it is at high risk to become a chicken or eggs problem especially for FR where newcomers start to be exposed to tf.keras directly instead of coming from a Keras experience and I supposed this hypothesis it is not so much remote with the landing of TF 2.0 right?.

[bhack]

In the meantime probably could you create a specific tf.keras template?

[martinwicke]

@dksb do you think a separate keras template makes sense to simplify triage?

[martinwicke]

@bhack It is definitely possible to modify Keras via PR to tf.keras. We may ask for a mirror PR to keras-team/keras, but we are doing regular syncs, and contributions can move both ways.

[bhack]

@martinwicke Probably this topic needs to be better investigated cause we could have a tf.keras api with only 1 backend that breaks compatibility with Keras API with its own tensorflow backend. More and more users will be exposed to Keras trough tf.keras directly.

[bhack]

@martinwicke @ewilderj This is just another bidirectional case example: #24009 VS keras-team/keras#9898 (comment).
Could we have divergent opinions?