peterzs / denseposefnl

Fast and Light DensePose implementation

Making DensePose fast and light

Original DensePose project Quick Start

See Getting Started

Training and Evaluation

0. The project dependencies:

• detectron2==0.1.0
• pytorch >= 1.4.0

1. You can train a network from scratch using configs in ./configs folder and train_net.py script.

• s0_bv2_bifpn_f64_s3x.yaml config corresponds to the Mobile-RCNN (B s3x) model,
• s0_bv2_bifpn_f64.yaml config corresponds to the Mobile-RCNN (B s1x) model,
• densepose_parsing_rcnn_spnasnet_100_FPN_s3x.yaml config corresponds to the Mobile-RCNN (A s1x) model,
• densepose_parsing_rcnn_R_50_FPN_s1x.yaml config corresponds to the Parsing RCNN model Then evaluate the model with --eval_only flag.

2. You can run QAT of the Mobile-RCNN (B s3x) using train_net.py with --qat flag then evaluate it with --quant-eval flag. To use proposed hooks preserving mechanism it is needed to modify PyTorch source code according to files inside modify_pytorch directroy OR Build PyTorch from source using the following commit https://github.com/pytorch/pytorch/pull/37233/commits/c8de10d2a394484ac58dd131878950b8ab7ac7a9