One loading dock for PyTorch’s scattered ConvNets

What it does

pretrained-models.pytorch is a model zoo that corrals dozens of ConvNets—NASNet, SENet, Xception, ResNeXt, and others—under a single loading interface inspired by torchvision. Instead of chasing down separate repos for each architecture, you call one API and get the model plus its correct input size, normalization constants, and preprocessing logic. A companion utility handles rescaling, cropping, and color-space quirks automatically.

The interesting bit

The real value is normalization diplomacy. Every model exposes metadata like input_range, mean, and std, and the TransformImage utility uses it to reconcile idiosyncratic expectations—BGR versus RGB, 0–1 versus 0–255—without you memorizing them. That makes swapping architectures or reproducing paper results far less tedious.

Key highlights

• Covers models that were hard to find in PyTorch at the time, including NASNet, PNASNet, SENet154, and DualPathNetworks.
• Exposes model metadata (input_size, input_space, mean, std) so preprocessing stays in sync with the weights.
• Splits inference into model.features() and model.logits() for transfer-learning setups, though not every network supports it.
• Ships with scripts to compute ImageNet logits and validation metrics out of the box.
• Downloads weights automatically from hosted URLs.

Caveats

• The README still carries a “Work in progress” warning.
• The features/logits API is explicitly noted as unavailable for some networks.
• Weights are pulled from external hosts like data.lip6.fr, so availability depends on those endpoints staying up.

Verdict

Worth a look if you need older or exotic ImageNet architectures in PyTorch and want to avoid porting weights yourself. Skip it if you are already happy with the modern model zoo baked into current torchvision.