Colabtools: Tensorflow eager mode dose not work when GPU is enabled in google colab

[alinamadchian]

recently I have faced this problem.
when I want to use GPU as the accelerator in google collab, the eager mode won't work and when I try to enable the eager mode even at the beginning of the program it gives me the error:
ValueError: tf.enable_eager_execution must be called at program startup.
while when the GPU is not used the eager mode can be activated and be used successfully in google collab

[colaboratory-team]

Sorry for the trouble. Please try the following and let us know if the issue recurs:

1. Reset your environment using the 'Reset all runtimes' item from the Runtime menu.
   

2. Execute the following in a code cell:

import tensorflow as tf
tf.enable_eager_execution()

[alinamadchian]

Unfortunately, I tried again and the issue is not resolved. I did not have such a problem two days ago, it seems something goes wrong recently

[craigcitro]

Yep, this is a bug (which will only happen with GPUs) -- to work around it for now, reset all runtimes, and then run

import os
os.environ['DISABLE_COLAB_TF_IMPORT_HOOK'] = '1'

before any TF-using code, and you should be unblocked.

[alinamadchian]

Yes, It works.
thank you very much.

[craigcitro]

FTR fix is now live, this should be working without the custom env var (after possibly restarting any runtimes).

[ParasharaRamesh]

I tried the @craigcitro 's solution of importing the os module. But it still gives me an error saying module 'tensorflow' has no attribute 'enable_eager_execution'.
The colab tutorial I am trying is the following
"https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/r2/tutorials/generative/image_captioning.ipynb#scrollTo=Dx_fvbVgRPGQ"
and in the cell "Caching the features extracted from InceptionV3" the last line with "np.save(path_of_feature, bf.numpy())" gives me an error saying "Tensor has no attribute called numpy()" , so i am guessing it is to do with enabling eager execution but i am not sure.

[craigcitro]

@ParasharaRamesh Can you file a separate issue for what you're hitting?

[MarkDaoust]

@ParasharaRamesh Can you file a separate issue for what you're hitting?

Yes please.

Everything in that r2/ directory is updates in progress for tf2.0.
Not everything works yet. But this looks worse than an error in the notebook.

That batch_features = image_features_extract_model(img) line is returning a graph tensor, instead of an eager tensor. It's not supposed to do that.

Can you raise an issue on tensorflow/tensorflow, and cc me on it?

[ParasharaRamesh]

@MarkDaoust I was able to get past that issue by doing the following:

• Not pip installing tf-nightly
• using the default tensorflow 1.13.0rc version and then executing tf.enable_eager_execution() which happened to work where i was previously getting an error