Imbatmann/nepali-banknote-embedder

Banknote Embedder

Fine-tuned visual embedding model for Nepali banknotes.

It maps an input banknote image to a 512-d normalized embedding vector.
You can compare two notes using cosine similarity:

• higher score -> more similar
• lower score -> less similar

Model Files

• banknote_embedder_finetuned.pth: training state_dict
• banknote_embedder_fp16.safetensors: compressed inference weights (recommended)
• banknote_embedder_int8_dynamic.pth: dynamic int8 export for supported CPU backends

Quantize / Export

Generate compressed artifacts from your fine-tuned weights:

python quantize_model.py --input-model banknote_embedder_finetuned.pth --skip-int8

Outputs:

• banknote_embedder_fp16.safetensors

Optional (if your CPU backend supports it):

python quantize_model.py --input-model banknote_embedder_finetuned.pth

Additional output:

• banknote_embedder_int8_dynamic.pth

Quick Test

Run similarity test:

python test.py --weights banknote_embedder_fp16.safetensors --img1 notes/20.jpg --img2 notes/20.jpg --img3 notes/22.jpg

Expected behavior:

• img1 vs img2 should be high (often close to 1.0 for identical image)
• img1 vs img3 should be lower

Usage Notes

• Input preprocessing uses ImageNet normalization and 224x224 resize.
• Embeddings are L2-normalized in the model forward pass.
• xFormers is not available warnings are expected and non-fatal.