ModuleNotFoundError: No module named 'scenic.train_lib_deprecated'

[molu-ggg]

I want to use vivit ,
python==3.10, scenic==0.0.1 ,
but train_lib_deprecated not in scenic.
Do you have a solution? Or which version of Scenic should be used?

[ffffffly923]

+1

[pepper725]

+1

[pepper725]

Did you solve the problem?

[dddcyy]

just replace 'scenic.train_lib_depreacated' with 'scenic.train_lib', it works on my settings

[ffffffly923]

can u train the model correctly?

[dddcyy]

no, I can't. There are always many problems if I want to run vivit. I am still fixing this horrible thing[sadly]

[dddcyy]

I really hope people who need vivit model can gather to solve these issues (I have spent like two weeks on this f****** s***, still imcompatible denpendencies)

[tj-messi]

@dddcyy can we have a talk?

[zmasih]

I faced the same issue. Any updates?

[UsmanNiz]

Same problem with me, it has been a week and I stuck with the dependencies issues.

[zmasih]

Same problem with me, it has been a week and I stuck with the dependencies issues.

As far as I remember, that happened since Scenic’s older train_lib_deprecated was removed. I switched to the newer Optax-based code, and it solved the issue

[rickyC3]

Hi, I’ve managed to resolve this issue.

The main change was switching from the customized ViViT trainer to train_lib/classification_trainer.py. This setup now works correctly. There are still some minor issues related to tensorflow_addons, as it has been deprecated, but these can be resolved with small modifications, mainly in video_ops.py (this can be fixed fairly easily with the help of ChatGPT).

In addition, the label mapping function in DMVR needs to be corrected (see dmvr/example/generate_from_file.py). Otherwise, you may encounter a situation where the training accuracy reaches around 90% while the validation accuracy remains near 5%, which is caused by label mismatches between the training and validation datasets.

However, I am currently facing another issue: I am unable to properly load and use the pre-trained checkpoint, and the performance is significantly worse than expected. The detailed discussion can be found here:
#227

I hope we can work together to resolve this problem.

[dddcyy]

@dddcyy can we have a talk?

Well, I gave up