Bubunur/animals-classifier

Animals Classifier: Dog / Cat / Panda

A Vision Transformer fine-tuned to classify images into 3 classes: cat, dog, and panda (giant panda).

Base model: google/vit-base-patch16-224-in21k

Usage

from transformers import pipeline

clf = pipeline("image-classification", model="Bubunur/animals-classifier")
clf("photo.jpg")
# [{'label': 'cat', 'score': 0.93}, ...]

Results

Evaluated on the held-out test split (369 images, 123 per class):

Class	Precision	Recall	F1
Cat	0.9667	0.9431	0.9547
Dog	0.9444	0.9675	0.9558
Panda	1.0000	1.0000	1.0000
Overall accuracy			0.9702

Confusion matrix (rows = true, columns = predicted):

	Cat	Dog	Panda
Cat	116	7	0
Dog	4	119	0
Panda	0	0	123

Training data

Bubunur/dog-cat-panda: photos collected from iNaturalist (Creative Commons licenses only). The raw data contained many non-photo observations (skulls, bones, tracks, empty camera-trap frames), so the dataset was cleaned with zero-shot filtering by CLIP ViT-L/14 (confidence threshold 0.6) and balanced to 614 images per class.

Split: 1,252 train / 221 validation / 369 test.

Training procedure

• Epochs: 8 (best checkpoint: epoch 6, selected by validation accuracy)
• Learning rate: 5e-5, warmup ratio 0.06, weight decay 0.01
• Batch size: 16
• Label smoothing: 0.1
• Mixed precision (fp16) on a single T4 GPU
• Augmentations (train only): random resized crop (70–100%), horizontal flip, rotation (±15°), color jitter (brightness / contrast / saturation 0.3)

Limitations

• Only 3 classes. Every input is assigned to cat, dog, or panda, even if it shows something else (e.g. a fox will be classified as a dog or a cat).
• Confidence scores are capped at ~0.93 because of label smoothing (0.1). A score of 0.93 means maximum confidence, not uncertainty.
• Cat/dog confusion remains on hard images: small or distant animals, partially hidden animals, fluffy small dog breeds.
• The test set was filtered by CLIP, so it may be easier than real-world images, and the reported accuracy may be optimistic.

License

Released under CC BY-NC-SA 4.0 to match the license of the training data (which includes CC BY-NC and CC BY-SA images). Non-commercial use only.