PyTorch SSD that treats backbones like LEGO bricks

What it does

This is a PyTorch implementation of the Single Shot MultiBox Detector (SSD) for object detection. It handles training and inference on Pascal VOC and COCO, ships with a model zoo spanning VGG to EfficientNet, and targets researchers who need a baseline they can modify without unraveling the entire codebase.

The interesting bit

The codebase is split into registered, swappable components—backbone, Detector, BoxHead, BoxPredictor—so adding EfficientNet (already bundled) is just a new file plus a config change. It also checkpoints the full training state—model, optimizer, scheduler, and current iteration—so resuming is exact, not a rough restart.

Key highlights

• Modular architecture: swap backbones, heads, or predictors via registration and YAML configs.
• Multi-GPU training and inference through DistributedDataParallel.
• Exact resumption from saved optimizer and scheduler states, not just model weights.
• TensorBoard logging of granular metrics, including per-category AP for VOC and AP/APL/APM/APS for COCO.
• CPU inference and batched multi-image inference per GPU.
• Auto-download and cache for pre-trained weights from release URLs.

Caveats

• Provided configs assume single-GPU training; moving to multi-GPU requires manually rescaling learning rate and iterations per the large-minibatch SGD paper.
• Dataset setup expects rigid directory structures or symlinks and environment variables rather than automatic discovery.

Verdict

Worth a look for CV researchers who want a hackable SSD baseline with modern backbone support. Look elsewhere if you need a fully managed, zero-config training platform.