MMDet to tensorrt

This project aims to convert the mmdetection model to tensorrt model end2end.
Focus on object detection for now, instance segmentation is not fully supported.

support:

• fp16
• int8(experiment)
• batched input
• dynamic input shape
• combination of different modules
• deepstream support

Any advices, bug reports and stars are welcome.

License

This project is released under the Apache 2.0 license.

Requirement

• mmdet>=2.3.0
• https://github.com/grimoire/torch2trt_dynamic
• https://github.com/grimoire/amirstan_plugin

Important!

Set the envoirment variable(in ~/.bashrc):

export AMIRSTAN_LIBRARY_PATH=${amirstan_plugin_root}/build/lib

Installation

Host

git clone https://github.com/grimoire/mmdetection-to-tensorrt.git
cd mmdetection-to-tensorrt
python setup.py develop

Docker

Build docker image

# cuda10.2 tensorrt7.0 pytorch1.6
sudo docker build -t mmdet2trt_docker:v1.0 docker/

Run

sudo docker run --gpus all -it --rm -v ${your_data_path}:${bind_path}  mmdet2trt_docker:v1.0

Usage

how to create a tensorrt model from mmdet model (converting might take few minutes)(Might have some warning when converting.)
detail can be found in getting_started.md

CLI

mmdet2trt ${CONFIG_PATH} ${CHECKPOINT_PATH} ${OUTPUT_PATH}

Run mmdet2trt -h for help on optional arguments.

Python

opt_shape_param=[
    [
        [1,3,320,320],      # min shape
        [1,3,800,1344],     # optimize shape
        [1,3,1344,1344],    # max shape
    ]
]
max_workspace_size=1<<30    # some module and tactic need large workspace.
trt_model = mmdet2trt(cfg_path, weight_path, opt_shape_param=opt_shape_param, fp16_mode=True, max_workspace_size=max_workspace_size)
torch.save(trt_model.state_dict(), save_path)

how to use the converted model

trt_model = init_detector(save_path)
num_detections, trt_bbox, trt_score, trt_cls = inference_detector(trt_model, image_path, cfg_path, "cuda:0")

how to save the tensorrt engine

with open(engine_path, mode='wb') as f:
    f.write(model_trt.state_dict()['engine'])

note that the bbox inference result did not divided by scale factor, divided by you self if needed.

play demo in demo/inference.py

getting_started.md for more detail

How does it works?

Most other project use pytorch=>ONNX=>tensorRT route, This repo convert pytorch=>tensorRT directly, avoid unnecessary ONNX IR. read https://github.com/NVIDIA-AI-IOT/torch2trt#how-does-it-work for detail.

Support Model/Module

• Faster R-CNN
• Cascade R-CNN
• Double-Head R-CNN
• Group Normalization
• Weight Standardization
• DCN
• SSD
• RetinaNet
• Libra R-CNN
• FCOS
• Fovea
• CARAFE
• FreeAnchor
• RepPoints
• NAS-FPN
• ATSS
• PAFPN
• FSAF
• GCNet
• Guided Anchoring
• Generalized Attention
• Dynamic R-CNN
• Hybrid Task Cascade
• DetectoRS
• Side-Aware Boundary Localization
• YOLOv3
• PAA
• CornerNet(WIP)
• Generalized Focal Loss
• Grid RCNN
• VFNet

Tested on:

• torch=1.6.0
• tensorrt=7.1.3.4
• mmdetection=2.5.0
• cuda=10.2
• cudnn=8.0.2.39

If you find any error, please report in the issue.

FAQ

read this page if you meet any problem.

Contact

This repo is maintained by @grimoire

Discuss group: QQ:1107959378