RdSean / DensePoseFnL

Fast and Light DensePose implementation

Making DensePose fast and light

Code for Making DensePose fast and light.

Original DensePose project Quick Start

See Getting Started

Training and Evaluation

0. The project dependencies:

• detectron2 Install it the following way:

git clone https://github.com/facebookresearch/detectron2.git && cd detectron2
git checkout b1fe5127e41b506cd3769180ad774dec0bfd56b0
pip install -e .

• timm == 0.1.16
• 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 s3x) 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 Use PyTorch nightly build (it is now containing the following commit https://github.com/pytorch/pytorch/pull/37233/commits/c8de10d2a394484ac58dd131878950b8ab7ac7a9)