Why lightweight models for pet commerce (UX/SEO tốc độ)
Dataset: cấu trúc thư mục, cân bằng lớp, split train/val
Model zoo timm + Augmentations (Albumentations)
Training loop (mixed precision), metrics (F1/macro)
Export ONNX + simple CPU benchmark
Gotchas: label leakage, imbalance, overfit lông/màu nền
Further reading + demo link (Hugging Face)
Opening snippet (Markdown)
Problem: Pet e-commerce needs fast, accurate breed recognition on low-spec devices.
Goal: <80ms CPU inference, >90% macro-F1 on 20–30 popular breeds.
Approach: timm backbone + strong aug + label smoothing + ONNX export.
model.eval()
dummy = torch.randn(1,3,224,224)
torch.onnx.export(model, dummy, "petbreed.onnx", input_names=["input"], output_names=["logits"], opset_version=17)