float number in randint Python 3.12

[dk222dk]

I noticed in the source code there are some calls of generator.randint(1, 1e9) 1e9 is converted to float but in python 3.12 calling float in randint causes an error

System information.

• Python 3.12
• tf_keras 2.17.0
• tensorflow 2.17.0

Running with env variable: TF_USE_LEGACY_KERAS=1

Simple steps to reproduce:

import tensorflow as tf
tf.keras.utils.set_random_seed(10)
from tensorflow.keras.initializers import VarianceScaling
vs = VarianceScaling()
vs((3, 3))

Error log

Traceback (most recent call last):
  File "/mnt/g/PycharmProjects/testing/first.py", line 8, in <module>
    vs((3, 3))
  File "/home/user/miniconda3/envs/t2/lib/python3.12/site-packages/tf_keras/src/initializers/initializers.py", line 648, in __call__
    return self._generate_init_val(shape=shape, dtype=dtype, nonce=nonce)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/miniconda3/envs/t2/lib/python3.12/site-packages/tf_keras/src/initializers/initializers.py", line 663, in _generate_init_val
    return self._random_generator.truncated_normal(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/miniconda3/envs/t2/lib/python3.12/site-packages/tf_keras/src/backend.py", line 2134, in truncated_normal
    self._maybe_init()
  File "/home/user/miniconda3/envs/t2/lib/python3.12/site-packages/tf_keras/src/backend.py", line 1976, in _maybe_init
    self._seed = self._create_seed(self._seed)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/miniconda3/envs/t2/lib/python3.12/site-packages/tf_keras/src/backend.py", line 2032, in _create_seed
    return _SEED_GENERATOR.generator.randint(1, 1e9)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/miniconda3/envs/t2/lib/python3.12/random.py", line 336, in randint
    return self.randrange(a, b+1)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "/home/user/miniconda3/envs/t2/lib/python3.12/random.py", line 312, in randrange
    istop = _index(stop)
            ^^^^^^^^^^^^
TypeError: 'float' object cannot be interpreted as an integer 

[tilakrayal]

@dk222dk,
I tried to execute the mentioned code on both Keras2.0 & keras3.0 and observed that the code was executed without any issues/fail. Kindly find the gist of it here. Thank you!

[dk222dk]

@tilakrayal You didn't use python 3.12 this error occurs only with python >=3.12 and in previous versions you get warnings that calling randint with float is deprecated and support will be removed in next python versions

Please try in fresh docker environment with python 3.12:

docker pull python:3.12
docker run --rm --entrypoint /bin/bash -it python:3.12

inside docker:

pip install tf_keras==2.17.0 tensorflow==2.17.0
apt update
apt install nano
nano script.py

copy this:

import tensorflow as tf
tf.keras.utils.set_random_seed(10)
from tensorflow.keras.initializers import VarianceScaling
vs = VarianceScaling()
vs((3, 3))


export TF_USE_LEGACY_KERAS=1
python script.py

The solution would be to find all calls of .randint(...) with floats and wrap them with int(), I saw there was one place in source code where it was done but not everywhere

for example: generator.randint(1, 1e9) -> generator.randint(1, int(1e9))

[dk222dk]

@tilakrayal are you able to create a MR with that change and release new version of tf_keras? it should be a simple change currently it's something that blocks users from using tf_keras with python 3.12 since it doesn't work with that version.

[github-actions]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[dk222dk]

I noticed there are other test fails with python 3.12

[github-actions]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[github-actions]

This issue was closed because it has been inactive for 28 days. Please reopen if you'd like to work on this further.

[MartinNowak]

Mostly fixed by 6f44991
Though it seems like these remained → #815.