models

Just Another PyTorch Model Zoo

All models here were either written from scratch or refactored from open-source implementations. All models here use Activated Normalization layers instead of traditional Normalization followed by Activation. It makes changing activation function and normalization layer easy and convenient. It also allows using Inplace Activated Batch Norm from the box, which is essential for reducing memory footprint in segmentation tasks.

Pretrained models

All default weights from TorchVision repository are supported. There are also weights for modified Resnet family models trained on Imagenet 2012. It's hard to keep this README up to date with new weights, so check the code for all available weight for particular model. All models have pretrained_settings attribute with training size, mean, std and other useful information about the weights.

Encoders

All models from this repo could be used as feature extractors for both object detection and semantic segmentation. Passing encoder=True arg will overwrite forward method of the model to return features at 5 different resolutions starting from 1/32 to 1/2.

Features

• Unified API. Create resnet, efficientnet, hrnet models using the same code
• Low memory footprint dy to heavy use of inplace operations. Could be reduced even more by using norm_layer='inplaceabn'
• Fast models. As of 04.20 Efficient net's in this repo are the fastest available on GitHub (afaik)
• Support for custom number of input channels in pretrained models. Try with resnet34(pretrained='imagenet', in_channels=7)
• All core functionality covered with tests

Repositories used

• Torch Vision Main Repo
• Cadene pretrained models
• Ross Wightman models
• Inplace ABN
• Efficient Densenet
• Official Efficient Net
• Original HRNet for Classification