Port to tensorflow 2

[renatobellotti]

Running this module using tensorflow 2.0.0 yields an error:
AttributeError: module 'tensorflow' has no attribute 'get_default_graph'

Are there any plans to port the library to tensorflow 2.0.0?

Perhaps all you have to do is change import keras to from tensorflow import keras and perhaps use fully qualified names when using layers...

[renatobellotti]

The following minimal example gives another error message. The same code without ImportanceTraining works fine.

from tensorflow import keras
import numpy as np

from importance_sampling.training import ImportanceTraining

num_x_features = 5
num_y_features = 2

x_train = np.random.uniform(0, 1, (100, num_x_features))
y_train = np.random.uniform(0, 1, (100, num_y_features))

model = keras.models.Sequential()
model.add(keras.layers.Dense(10, activation='relu', input_shape=(5,)))
model.add(keras.layers.Dense(10, activation='relu'))
model.add(keras.layers.Dense(num_y_features))
model.compile("adam", loss="mse")

history = ImportanceTraining(model).fit(
    x_train, y_train,
    batch_size=8,
    epochs=5,
)

Error message:

AttributeError: 'Node' object has no attribute 'output_masks'

[renatobellotti]

Will tensorflow 2 be supported at some point? I'd like to try this for my master's thesis and it would be nice to save the model in tensorflow 2 because keras will soon become obsolete.

[angeloskath]

Hi, sorry for the late reply.

It would be nice to update to tf 2. If instead of from tensorflow import keras you do import keras does it still have issues? I have not had the time to look at it yet.

Feel free to dig in the code and ask question I will help as best I can.

Cheers,
Angelos

[renatobellotti]

It is not enough, unfortunately.

import keras does not work anymore. Starting with tensorflow 2, keras will be completely integrated into tensorflow. Quote from www.keras.io:

At this time, we recommend that Keras users who use multi-backend Keras with the TensorFlow backend switch to tf.keras in TensorFlow 2.0. tf.keras is better maintained and has better integration with TensorFlow features (eager execution, distribution support and other).

Keras 2.2.5 was the last release of Keras implementing the 2.2.* API. It was the last release to only support TensorFlow 1 (as well as Theano and CNTK).

I've tried to upgrade the project automatically, but apparently there are some things that are not caught by the script.

The point where I'm failing is in importance-sampling/training.py. What is CallbackList? I was not able to find any documentation, so I don't know how to replace it.

I can probably invest a bit time in porting this library, but I'd probably need some help.

[MaximilianBoemer]

Try: from tensorflow.python.keras.callbacks import CallbackList

tensorflow/tensorflow#23880 (comment)

[renatobellotti]

I have made some formal changes to the code in my fork (4456623). There are still some things I couldn't solve yet:

• test_model_learning_phase.py: The tests fail.
• test_seq2seq.py: The tests fail.
• test_keras_utils.py: The test fails. Other than that: I think these tests are invalid, because in lines 29 and 47 only the values that already are in memory are compared. This makes the failure even stranger...
• test_finetuning.py: Passes, but the loss value is NaN.

I suspect most of the errors come from lines 262 and 384 model_wrappers.py. I'm pretty sure this should be different, but I couldn't find out how to avoid the error that is thrown in the original code.

I'd appreciate your feedback on the issue.

[angeloskath]

@renatobellotti Sorry for the late reply, if you still need help with anything, let me know...

[renatobellotti]

I have not worked on this again, and my thesis does not allow me to spend more time ob this issue. I hope my contributions are at least a bit helpful.

[maryam2013]

Hi everybody,
I come across the same error,

AttributeError: module 'tensorflow' has no attribute 'get_default_graph',

when applying importance sampling with Keras.
have you found a solution?
I would be pleased if you help me with the issue.
Thank you in advance
Maryam

[angeloskath]

At this point I would advise using TF 1 instead of 2.

[maryam2013]

Dear Angelos, thank you for replying so fast, but have u solved the problem bt using TF1? Best regards Maryam

…

On Mon, Apr 27, 2020 at 1:09 AM Angelos Katharopoulos < ***@***.***> wrote: At this point I would advise using TF 1 instead of 2. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#25 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJCCHE3PYMYHJJXSI63LY2LROSLZRANCNFSM4JMFHO4Q> .

[angeloskath]

In my case it works with keras 2.3.1 and tf 2.1 however when the code was first developed there was tf 1 and keras 2 so if you are having issues, then try to use a tensorflow version < 2.

[renatobellotti]

Perhaps a clarification is needed. With the transition to tf2, the Keras API was merged into TensorFlow and is now a part of it. The standalone version of Keras will be discontinued. That means if this package cannot be run with tf2 only (without the separate Keras package), there will not be much future for it.

I'm sorry to say this so directly, but I think this is a very important issue.

[maryam2013]

Dear all,
I come across the same error when implementing importance sampling with Keras,
from importance_sampling.training import ImportanceTraining

AttributeError: 'Node' object has no attribute 'output_masks'

have you found a solution?
I would be grateful if you help me with the problem.
Thank you in advance
Maryam

[renatobellotti]

Dear all,
I come across the same error when implementing importance sampling with Keras,
from importance_sampling.training import ImportanceTraining

AttributeError: 'Node' object has no attribute 'output_masks'

have you found a solution?
I would be grateful if you help me with the problem.
Thank you in advance
Maryam

@maryam2013 I am not using this package because I'm using TF2, but perhaps the developers need your version of Keras to help you.