PyTorch backend doesn't work in Kaggle notebook

[romanvelichkin]

I'm trying to run model on Kaggle notebook using PyTorch backend. I guess notebook by defaul has both Tensorflow and PyTorch. And because of that package create tensorflow model no matter if environ set to "torch".

Code:

os.environ["KECAM_BACKEND"] = "torch"
import torch
from keras_cv_attention_models import efficientnet

model = efficientnet.EfficientNetV2B0(is_torch_mode=True,
                                      pretrained='imagenet21k', # weights pretrained on imagenet21k
                                      num_classes=num_classes,
                                      classifier_activation=activation_fn,
                                      dropout=0.25,
                                     )
print(f"{isinstance(model, torch.nn.Module) = }")

Output:

torch
>>>> Load pretrained from: /root/.keras/models/efficientnetv2-b0-21k.h5
isinstance(model, torch.nn.Module) = False

[romanvelichkin]

I deleted tensorflow and keras, restarted kernel, and then kecam started to use pytorch backend.

[leondgarse]

Sorry for the late reply. Just back from a vacation.
Deleting tensorflow and keras shouldn't be necessary, as most time I'm also using all them installed. Also this colab basic test kecam_test.ipynb. I'm not sure, maybe keras_cv_attention_models is already imported before setting KECAM_BACKEND env?

[romanvelichkin]

Sorry for the late reply. Just back from a vacation. Deleting tensorflow and keras shouldn't be necessary, as most time I'm also using all them installed. Also this colab basic test kecam_test.ipynb. I'm not sure, maybe keras_cv_attention_models is already imported before setting KECAM_BACKEND env?

The problem is not about os variable. Kecam creates tensorflow (not pytorch) model even if variable was set to 'torch' in situation when environment has both pytorch and tensorflow. I'm not sure about google collab or desktop jupyter, I met problem in Kaggle notebook. You can check again my first post.

[leondgarse]

Ya, It should work not matter colab or kaggle notebook. Here's my basic kaggle notebook test kaggle kecam_test.