sonnet abstract module in tensorflow2

[creativesh]

Hello friends

I want to use i3d model created by sonnet in tensorflow 2

these are my environment:
ubuntu 20
tensorflow 2.12
python 3.8

this is my i3d code:

`
from future import absolute_import

from future import division
from future import print_function
import sonnet as snt
import tensorflow as tf
class Unit3D(snt.AbstractModule):

`

but when I want to import it in tf2, I face the following error:
AttributeError: module 'sonnet' has no attribute 'AbstractModule'

which is related to this line:

class Unit3D(snt.AbstractModule):
>>> class MultiplyModule(snt.Module):
... def call(self, x):
... if not hasattr(self, 'w'):
... self.w = tf.Variable(2., name='w')
... return x * self.w

  >>> mod = MultiplyModule()
  >>> mod(1.)
  <tf.Tensor: ... numpy=2.0>

how to fix it ?

[tomhennigan]

Hi @creativesh , snt.AbstractModule is part of Sonnet 1, it looks like the code you are wanting to run will only work with TF1 not TF2.

[creativesh]

Hi @tomhennigan

Could I rewrite it to use tf2 ?

[creativesh]

Thanks How should I convert it to tf2 ? It is an i3d model. for action recognition I have some weights in .ckpt format which is in tf1 format and want to load them in a model and retrain the model

…

On Sat, Sep 30, 2023 at 3:37 PM Tom Hennigan ***@***.***> wrote: Hi @creativesh <https://github.com/creativesh> , snt.AbstractModule is part of Sonnet 1, it looks like the code you are wanting to run will only work with TF1 not TF2. — Reply to this email directly, view it on GitHub <#271 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFOLMRPAVS434CME24IOEZDX5ADQNANCNFSM6AAAAAA5NNGEJ4> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[tomhennigan]

Hi, it should be possible to implement this in TF2/Sonnet 2, but this will require learning and understanding Sonnet 2.

If you just want to make use of a pretrained I3D model then perhaps using Keras + TF2 would be simpler, they have an example that includes using I3D and some pretrained weights to make predictions for a video: https://www.tensorflow.org/hub/tutorials/action_recognition_with_tf_hub

[creativesh]

@tomhennigan

Thanks for your response,
exactly, I want to finetune the checkpoints you mentioned in your link :
https://www.tensorflow.org/hub/tutorials/action_recognition_with_tf_hub

I have written the tf1 version of finetune code. now I want to write it in tf2.
the limitation is that Ihad to use .ckpt checkpoints and save the fine-tuned model in .ckpt format. Can I do it with keras? Do you have any link or hint?

[tomhennigan]

I am not very familiar with Keras, I would suggest asking for support on their issue tracker, hopefully someone will have a recipe for you.

There is a generic guide on migrating checkpoints here which you could try to follow https://www.tensorflow.org/guide/migrate/migrating_checkpoints

[creativesh]

@tomhennigan

Do you have any sample code of developing networks with sonnet 2 and tensorflow 2 ?

[tomhennigan]

There are a number of examples here: https://github.com/google-deepmind/sonnet/tree/v2/examples

[creativesh]

@tomhennigan

In this post #128
you mentioned aome changes to convert a class in sonnet 1 to sonnet2

would you please tell me exactly what to do about i3d.py in this link:
https://github.com/google-deepmind/kinetics-i3d

1- converting snt.AbstractModule to snt.Module
2- ?

[tomhennigan]

Hi, the comment you linked to has information on how to port code from sonnet 1 -> 2 and what has changed between the libraries.

Good luck with porting the code for your project.