Unit 3: bug fixes for transfer-learning-image-classification.ipynb

[tyleryy]

Bug Description

Ran the notebook on Google Colab and noticed a few bugs. I managed to fix them, so if this issue is approved, I'd like to work on it. It would is also my first contribution to open-source :)

1. Loading the pcuenq/oxford-pets dataset

• NotImplementedError: Loading a dataset cached in a LocalFileSystem is not supported. is thrown
• fixed by adding !pip install -U datasets beforehand

2. Functions processing images for display

• show_samples and show_predictions, try to convert samples[i]['image']['bytes'] into a PIL Image File. However, samples[i]['image'] is already a PIL.JpegImagePlugin.JpegImageFile. TypeError: 'JpegImageFile' object is not subscriptable is thrown.
• Fixed by removing code that converts bytes to PIL image

3. TrainingArguments has a deprecated evaluation_strategy argument

• Should be eval_strategy. Fixed by renaming.

Existing Issues

Found a similar existing issue #375, but I believe it is saying the opposite of what I am observing. The internal image attribute is not a bytes object and is already a PIL image.

Reproduction

Run the existing notebook on Colab.

Logs/Screenshots

Loading the pcuenq/oxford-pets dataset

Functions processing images for display

TrainingArguments has a deprecated evaluation_strategy argument