SL

ASL Fingerspelling (A-Z) - Fine-tuned ResNet18

This repository/package contains your fine-tuned classifier weights for ASL fingerspelling (manual alphabet A–Z).

<!-- -->

Files

• models/finetuned_webcam.pth (recommended): your fine-tuned PyTorch checkpoint for the ASL classifier.

What this model expects (input)

Your Python pipeline (webcam_inference.py / api_server.py / train_finetune.py) uses:

• Input: a single hand crop resized to 224×224
• Color: RGB
• Normalization: ImageNet mean/std
  • mean = [0.485, 0.456, 0.406]
  • std = [0.229, 0.224, 0.225]
• Network output: 26 logits corresponding to letters A..Z

How to load from Hugging Face (Python)

Replace <your-username> and <your-repo-name> with your Hugging Face repo:

from pathlib import Path
import torch
from huggingface_hub import hf_hub_download
from asl_model import load_model_from_checkpoint

REPO_ID = "<your-username>/<your-repo-name>"
FILENAME = "models/finetuned_webcam.pth"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ckpt_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)

model, meta = load_model_from_checkpoint(str(ckpt_path), device=device)
print("Loaded checkpoint:", ckpt_path)

Recommended Hugging Face repo layout

When you upload, keep the same paths:

• models/finetuned_webcam.pth

This matches the filename used in the loader snippet above.