✨ Add PyTorch image classification example #13134
Conversation
nateraw
changed the title
nateraw
changed the title
|
Nice!! Relevant for #13080 |
nateraw
force-pushed
the
torch-image-classification-ex
branch
from
459e24e
to
b61ab72
Compare
|
I'll review this PR in detail (thanks for working on this!). Regarding the fixtures for the tests, I've recently moved these files to the hf-internal-testing organization on the hub. This makes it more clear, as otherwise these fixture files are also downloaded when people do a |
| run_image_classification.py | ||
| --output_dir {tmp_dir} | ||
| --model_name_or_path google/vit-base-patch16-224-in21k | ||
| --train_dir tests/fixtures/tests_samples/cats_and_dogs/ |
There was a problem hiding this comment.
Can we add the cats and dogs as a dataset to the hub under the hf-internal-testing organization?
There was a problem hiding this comment.
We can, but this is actually testing the fact that the script works on local image folders. I can issue another PR that directly pulls them down into an imagefolder-like cache dir to test. But I'll leave this as-is for now if its not a big deal.
| "value if set." | ||
| }, | ||
| ) | ||
| image_size: Optional[int] = field(default=224, metadata={"help": " The size (resolution) of each image."}) |
There was a problem hiding this comment.
Will images be squared? Or is it just the smaller edge of the image that will be matched to this number (which torchvision's Resize does if you only provide an integer)?
There was a problem hiding this comment.
You're correct, this value is passed directly to torchvision's Resize
| train_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the training data."}) | ||
| validation_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the validation data."}) | ||
| train_val_split: Optional[float] = field( | ||
| default=0.15, metadata={"help": "Percent to split off of train for validation."} | ||
| ) |
There was a problem hiding this comment.
Hmm so you have a training dataset, a validation dataset and a test set. The validation_dir is actually the test set? And it only makes sense to add a train_val_split if you don't provide a validation dataset yourself?
There was a problem hiding this comment.
there's no test set here. There's just train + validation. If 'validation' key is not found in dataset, we create a split off of train for validation and set it to the 'validation' key of the dataset dict.
|
Last nit on my side: can we move the vision folder to be |
nateraw
force-pushed
the
torch-image-classification-ex
branch
from
21231cb
to
d979eeb
Compare
nateraw
changed the title
nateraw
changed the title
|
Ok, addressed most of the comments. Merging as-is for now. @NielsRogge I did not address these two items, however I can in future PRs (if need be):
|
|
13134 |
What does this PR do?
Adds PyTorch image classification example. For now, it uses
torchvision.datasets.ImageFolderto load local image folders (just like the flax image classification example). In the future, we will switch to using thedatasetspackage's image folder (once it exists).Marking as draft for now as I'm still working through cleaning up changes I made from this example I wrote earlier that uses
datasetsinstead.Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.