🐱🐶 Cat vs. Dog Classifier (EfficientNet-B0, Keras/TensorFlow)

A lightweight CNN that predicts whether an image contains a cat or a dog. The backbone is EfficientNetB0 pre-trained on ImageNet and fine-tuned on the microsoft/cats_vs_dogs training split (23 410 images).

Model Details

	Value
Backbone	EfficientNet-B0 (`include_top=False`)
Input size	`128×128×3`
Extra layers	GlobalAvgPool ➜ Dropout(0.2) ➜ Dense(1, sigmoid)
Precision	Mixed-precision (`float16` activations / `float32` dense)
Optimizer	AdamW with cosine-decay-restarts schedule
Loss	Binary Cross-Entropy
Epochs	25 (frozen backbone) + 5 (fine-tune full network)
Batch size	16
Class weighting	Balanced weights computed from training labels

Validation Metrics

Metric	Value
Accuracy	97.2 %
AUC	0.9967
Loss (BCE)	0.079

(computed on 15 % stratified validation split – 3 512 images)

Intended Uses & Limitations

Intended : quick demos, tutorials, educational purposes, CAPTCHA-like tasks.
Not intended : production-grade pet breed classification, safety-critical applications.
The model only distinguishes cats vs dogs; images with neither are undefined behaviour.
Trained on 128×128 crops; very large images might require resizing first.

Dataset Credits

The training data is the publicly available microsoft/cats_vs_dogs dataset (originally the Asirra CAPTCHA dataset). Huge thanks to Microsoft Research and Petfinder.com for releasing the images!

@misc{microsoftcatsdogs,
  title  = {Cats vs. Dogs Image Dataset},
  author = {Microsoft Research & Petfinder.com},
  howpublished = {HuggingFace Hub},
  url    = {https://huggingface.co/datasets/microsoft/cats_vs_dogs}
}

Acknowledgements

TensorFlow/Keras team for the excellent deep-learning framework.
Mingxing Tan & Quoc V. Le for EfficientNet.
The Hugging Face community for the awesome Model & Dataset hubs.

deruppu
/

catndog