# CS604 Course Project Dataset

For the task "ViT Model Adaptation".

## Overview
An anonymized fine-grained image classification dataset with 10,000 classes.

## Dataset Statistics
- **Number of classes**: 10,000
- **Training samples**: 450,000 (45 per class, with labels)
- **Validation samples**: 50,000 (5 per class, **without labels**) — Phase 1
- **Test samples**: 100,000 (10 per class, **without labels**) — released when Phase 2 opens

## Files
- `train.tar.gz` + `train.txt`: training images and labels (format: `train/class_XXXX/filename.jpg class_id`)
- `val.tar.gz` + `val.txt`: validation images and image paths only (format: `val/image_XXXXXX.jpg`)
- `test.tar.gz` + `test.txt`: added when Phase 2 opens
- `metadata.json`: dataset metadata

```
train/
  class_0000/  # 45 images
  ...
  class_9999/
val/
  image_000000.jpg
  ...
  image_049999.jpg
```

File names in `val/` and `test/` are random and carry no information about the class.

## Download
```bash
pip install -U huggingface_hub
hf download doem1997/cs604-course-data --repo-type dataset --local-dir cs604-course-data
cd cs604-course-data && tar -xzf train.tar.gz && tar -xzf val.tar.gz
```

## Submission
Predict a class for every image of the split and submit to the CodaBench competition (link in the course slides).

One line per image, `<image_path> <predicted_class_id>`, with `class_id` in [0, 9999]:
```
val/image_000000.jpg 1234
val/image_000001.jpg 5678
...
```
Name the file `val_team_XX.txt` (Phase 1) or `test_team_XX.txt` (Phase 2), zip it, and upload the zip.
Every image of the split must appear exactly once.

Validation and test labels are withheld. Do not label validation/test images by hand or train on them.