fix build in Travis

[Borda]

drop testing models because of the timeout failer
start adding CircleCI which should be able to run full test

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (master@a5d6d01). Click here to learn what that means.
The diff coverage is 67%.

@@           Coverage Diff            @@
##             master   #1137   +/-   ##
========================================
  Coverage          ?     66%           
========================================
  Files             ?      41           
  Lines             ?    2335           
  Branches          ?       0           
========================================
  Hits              ?    1532           
  Misses            ?     803           
  Partials          ?       0

[hgaiser]

@Borda this might be of interest to you: #1123 (comment)

[Borda]

thx, I will try to use it in testing if it helps... but it means that we restrict to TF usage only, right?

import tensorflow as tf
tf.compat.v1.disable_v2_behavior()

[Borda]

[Borda]

Still failing on CircleCI...

[hgaiser]

A general question, why should we use CircleCI instead of Travis?

@de-vri-es do you have any experience with CircleCI vs Travis?

[Borda]

I think that both Travis and CircleCI can be used in parallel... But I agree that it is not necessary so I will remove CircleCI from this PR
From my experience CircleCI has better machines and the tests runs faster, about 20%

[de-vri-es]

@de-vri-es do you have any experience with CircleCI vs Travis?

I think that both Travis and CircleCI can be used in parallel... But I agree that it is not necessary so I will remove CircleCI from this PR
From my experience CircleCI has better machines and the tests runs faster, about 20%

I have very limited experience with CircleCI. They have some cool features (you can get a shell in the VM/container of a failed run, for example), but I think we should stick with one CI system in total.

I'd be fine with either CircleCI or Travis. If it really runs faster that could be a good reason to switch. But whichever we go with, they must be able to complete the test suite =]

[Borda]

@de-vri-es I have stick this PR just with Travis and I may create a new PR with CircleCI and clearly state if and what is the performance boost compare to Travis...

[de-vri-es]

@de-vri-es I have stick this PR just with Travis and I may create a new PR with CircleCI and clearly state if and what is the performance boost compare to Travis...

Makes sense 👍

[Borda]

@de-vri-es is there anything else I should do for this PR? :)

[de-vri-es]

I think we want to merge most of this, but I'm not so sure about disabling the tests, even if it is temporary. We're close to a fix in #1146, and I'm a bit afraid that if we disable the tests now, we won't enable them again soon.

@hgaiser, @vcarpani: What do you think?

[hgaiser]

I agree with @de-vri-es, we're hopefully close to having working unit tests in keras 2.3 so we should aim for that.

[vcarpani]

Let's wait for working unit tests in Keras 2.3, it should not take too much time

[Borda]

I would consider adding to CI testing for Tensorflow 1.x and 2.x

[Borda]

@vcarpani @de-vri-es @hgaiser I have implemented the behaviour change and run multi TF testing...
in my opinion, this implementation of v1 behaviour is much smoother :]
BUT it seems that it needs also enable eager (as mentioned #1123 (comment)) which makes densenet freeze - https://travis-ci.org/Borda/keras-retinanet/jobs/604432952#L924

[Borda]

@de-vri-es @hgaiser @vcarpani any update so far? :]

[hgaiser]

Nope, seems keras is dead.

[Borda]

@hgaiser does it mean that this repo is also dead?

[hsahin]

@hgaiser ?

[hgaiser]

It will be deprecated in time, in favor of the implementation using tf.keras. For now it's not dead.