ASL Sign Language Recognition — Training Results

Trained by SharoonArshad

Results

Metric	Score
Overall macro-F1	77.24%
Accuracy	68.51%
Rare signs F1	99.47%
Medium signs F1	59.25%
Common signs F1	58.68%
Training time	52 minutes

Model Details

• Architecture: Transformer Encoder + Prototype Classifier
• Parameters: 3.57M
• Classes: 4,618 ASL signs
• Input: [60 frames × 204 features] body + hand landmarks
• Training: 3-phase curriculum (common → rare → fine-tune)

Files

File	Description
checkpoints/asl_v3_epoch050_score0.5728.pt	Best PyTorch checkpoint
checkpoints/asl_model.onnx	ONNX export for deployment
logs/training_history.json	Loss + F1 for all 50 epochs
logs/test_results.json	Final test set results
logs/train_v4.log	Full training log
label_map.json	Sign ID → Sign name mapping
tier_info.json	Rare / medium / common class splits
class_distribution.json	Samples per class
asl_results_complete.zip	All files in one zip

How to Load

import torch
from model_transformer import build_model
from config import CFG

ckpt = torch.load("asl_v3_epoch050_score0.5728.pt", map_location="cpu")
CFG.model.num_classes = ckpt["cfg"]["num_classes"]  # 4618

model = build_model(CFG, feature_dim_override=204)
model.load_state_dict(ckpt["model_state"])
model.eval()

# Inference
features     = torch.zeros(1, 60, 204)  # replace with real data
padding_mask = torch.zeros(1, 60, dtype=torch.bool)
logits, _    = model(features, padding_mask)
predicted    = logits.argmax(dim=-1).item()
print(f"Predicted sign ID: {predicted}")