Vecrist/resnet18-handsign-classifier

ResNet-18 American Sign Language (ASL) Classifier

This repository contains pre-trained PyTorch weights for a ResNet-18 model fine-tuned for American Sign Language (ASL) alphabetic hand sign classification (static letters A through Y, excluding motion-based letters J and Z).

• Overall Test Accuracy: 98.48% on 7,172 test images.
• Framework: PyTorch
• Base Architecture: ResNet-18 (ImageNet pre-trained)
• Transfer Learning Strategy: Frozen layer1-layer3, fine-tuned layer4 + custom head:
  • Linear(512, 256) -> BatchNorm1d(256) -> ReLU -> Dropout(0.3) -> Linear(256, 24)

---

Model Usage

import torch
import torchvision.models as models
import torch.nn as nn
from PIL import Image
from torchvision import transforms
from huggingface_hub import hf_hub_download

# 1. Define Model Architecture
model = models.resnet18()
model.fc = nn.Sequential(
    nn.Linear(512, 256),
    nn.BatchNorm1d(256),
    nn.ReLU(),
    nn.Dropout(0.3),
    nn.Linear(256, 24)
)

# 2. Download and Load Model Weights
weights_path = hf_hub_download(repo_id="Vecrist/resnet18-handsign-classifier", filename="ResNet-18_9848AccModel_weights.pth")
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

# 3. Preprocess Image
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# 4. Predict
# img = Image.open("path_to_handsign_image.jpg").convert("RGB")
# outputs = model(transform(img).unsqueeze(0))
# predicted_class_idx = outputs.argmax(dim=1).item()

---

Dataset & Training Details

• Dataset: Kaggle Hand Sign Images dataset (static alphabet signs).
• Optimizer: Adam with differential learning rates (0.0001 for layer4, 0.001 for FC head).
• Batch Size: 64
• Loss Function: CrossEntropyLoss
• Data Augmentations: RandomResizedCrop(224), RandomHorizontalFlip, ImageNet Normalization.