gofynd/mildnet

Implements lightweight CNN architectures (MILDNet, RankNet, Visnet) with skip connections for e-commerce image embedding and retrieval, using various loss functions (contrastive, hinge, angular) to train visual similarity models. Provides 25 pre-configured experiment variants enabling ablation studies across architectures (VGG16/19, MobileNet), loss functions, and feature dimensions. Supports both local and Google Cloud ML Engine training with pre-trained model weights available, plus Jupyter notebooks for end-to-end training and inference workflows.

No commits in the last 6 months.