Giters

soumik12345 / wandb-addons

Weights & Biases Addons is a repository consisting of additional unitilities and community contributions for supercharging your Weights & Biases workflows.

Home Page:https://geekyrakshit.dev/wandb-addons/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

[Feature Request]: Channels-first support for `WandBImageClassificationCallback`

soumik12345 opened this issue · comments

commented

Request from Ian Stenbit to make the WandBImageClassificationCallback work with both channels-first and channels-last image formats.

commented

Hi @soumik12345, can I work on this issue?

I figured out that the class WandBImageClassificationCallback exists in the wandb_addons\keras\image_classification.py,
and to work with both channels-first and channels-last image formats, we need to ensure that the image tensor is correctly formatted before logging it to WandB.

commented

@mratanusarkar Thanks for your interest. Please feel free to raise a PR.

commented

@mratanusarkar Here are 2 examples to aid your testing:

commented

@soumik12345 regarding the channels-first notebook you shared,

  • The import from keras_core.utils import TorchModuleWrapper throws the following error: 
    ImportError: cannot import name 'TorchModuleWrapper' from 'keras_core.utils' (/usr/local/lib/python3.10/dist-packages/keras_core/utils/__init__.py)
    Which was resolved with from keras_core.layers import TorchModuleWrapper as per your suggestion.
  • The notebook doesn't use WandBImageClassificationCallback hence I am unable to test out the channels-first usecase.

Could you please share me a notebook with channels_first image data using the WandBImageClassificationCallback?
It would be of great help. thanks!

commented

@soumik12345 I am unable to test the channels-first pipeline.

I feel WandBImageClassificationCallback callback is designed to work with TensorFlow datasets (tf.data.Dataset) or NumPy arrays (np.array). And not with the PyTorch's DataLoader.

more details

Here are some inputs I was using to debug:

  1. For channels-last notebook:

    Use this example to test for the channels-last pipeline.

    • type(test_dataset) --> tensorflow.python.data.ops.prefetch_op._PrefetchDataset
    • type(config.classes) --> list

  2. For channels-first notebook:

    Use this example to test for the channels-first pipeline.

    • type(image_datasets) --> dict
    • type(image_datasets["val"]) --> torchvision.datasets.folder.ImageFolder
    • type(class_names) --> list

Hence, I believe there was no WandBImageClassificationCallback in the channels-first notebook, and hence I made the above conclusion from a little bit of debugging.

I believe this could be a good feature request.
As per the current issue, please help me with how I should proceed and close the PR.