TensorRTx aims to implement popular deep learning networks with tensorrt network definition APIs. As we know, tensorrt has builtin parsers, including caffeparser, uffparser, onnxparser, etc. But when we use these parsers, we often run into some "unsupported operations or layers" problems, especially some state-of-the-art models are using new type of layers.
So why don't we just skip all parsers? We just use TensorRT network definition APIs to build the whole network, it's not so complicated.
I wrote this project to get familiar with tensorrt API, and also to share and learn from the community.
All the models are implemented in pytorch or mxnet first, and export a weights file xxx.wts, and then use tensorrt to load weights, define network and do inference. Some pytorch implementations can be found in my repo Pytorchx, the remaining are from polular open-source implementations.
8 Oct 2020. ChrystleMyrnaLobo added ssd(mobilenetv2).21 Sep 2020. BaofengZan added hrnet classification and step by step tutorial(Chinese).16 Sep 2020. hwh-hit added ufld(Ultra-Fast-Lane-Detection, ECCV2020).13 Sep 2020. Add crnn, and got 1000fps on GTX1080.7 Sep 2020. Implement retinaface(mobilenet0.25), and got 333fps on GTX1080.28 Aug 2020. BaofengZan added a tutorial for compiling and running tensorrtx on windows.16 Aug 2020. upczww added a python wrapper for yolov5.14 Aug 2020. Update yolov5 to v3.0 release.3 Aug 2020. BaofengZan implemented yolov5 s/m/l/x (yolov5 v2.0 release).6 July 2020. Add yolov3-tiny, and got 333fps on GTX1080.28 May 2020. arcface LResNet50E-IR model from deepinsight/insightface implemented. We got 333fps on GTX1080.22 May 2020. A new branch trt4 created, which is using TensorRT 4 API. Now the master branch is using TensorRT 7 API. But onlyyolov4has been migrated to TensorRT 7 API for now. The rest will be migrated soon. And a tutorial formigarating from TensorRT 4 to 7provided.
- Install the dependencies.
- A guide for quickly getting started, taking lenet5 as a demo.
- Frequently Asked Questions (FAQ)
- Migrating from TensorRT 4 to 7
- How to implement multi-GPU processing, taking YOLOv4 as example
- Check if Your GPU support FP16/INT8
- How to Compile and Run on Windows
- Deploy YOLOv4 with Triton Inference Server
- From pytorch to trt step by step, hrnet as example(Chinese)
- GTX1080 / Ubuntu16.04 / cuda10.0 / cudnn7.6.5 / tensorrt7.0.0 / nvinfer7.0.0 / opencv3.3
Each folder has a readme inside, which explains how to run the models inside.
Following models are implemented.
| Name | Description |
|---|---|
| lenet | the simplest, as a "hello world" of this project |
| alexnet | easy to implement, all layers are supported in tensorrt |
| googlenet | GoogLeNet (Inception v1) |
| inception | Inception v3 |
| mnasnet | MNASNet with depth multiplier of 0.5 from the paper |
| mobilenetv2 | MobileNet V2 |
| mobilenetv3 | V3-small, V3-large. |
| resnet | resnet-18, resnet-50 and resnext50-32x4d are implemented |
| senet | se-resnet50 |
| shufflenet | ShuffleNetV2 with 0.5x output channels |
| squeezenet | SqueezeNet 1.1 model |
| vgg | VGG 11-layer model |
| yolov3-tiny | weights and pytorch implementation from ultralytics/yolov3 |
| yolov3 | darknet-53, weights and pytorch implementation from ultralytics/yolov3 |
| yolov3-spp | darknet-53, weights and pytorch implementation from ultralytics/yolov3 |
| yolov4 | CSPDarknet53, weights from AlexeyAB/darknet, pytorch implementation from ultralytics/yolov3 |
| yolov5 | yolov5-s/m/l/x v1.0 v2.0 v3.0, pytorch implementation from ultralytics/yolov5 |
| retinaface | resnet50 and mobilnet0.25, weights from biubug6/Pytorch_Retinaface |
| arcface | LResNet50E-IR, weights from deepinsight/insightface |
| retinafaceAntiCov | mobilenet0.25, weights from deepinsight/insightface, retinaface anti-COVID-19, detect face and mask attribute |
| dbnet | Scene Text Detection, weights from BaofengZan/DBNet.pytorch |
| crnn | pytorch implementation from meijieru/crnn.pytorch |
| ufld | pytorch implementation from Ultra-Fast-Lane-Detection, ECCV2020 |
| hrnet | hrnet-image-classification, pytorch implementation from HRNet-Image-Classification |
| ssd | ssd(mobilenetv2), pytorch implementation from qfgaohao/pytorch-ssd |
Some tricky operations encountered in these models, already solved, but might have better solutions.
| Name | Description |
|---|---|
| BatchNorm | Implement by a scale layer, used in resnet, googlenet, mobilenet, etc. |
| MaxPool2d(ceil_mode=True) | use a padding layer before maxpool to solve ceil_mode=True, see googlenet. |
| average pool with padding | use setAverageCountExcludesPadding() when necessary, see inception. |
| relu6 | use Relu6(x) = Relu(x) - Relu(x-6), see mobilenet. |
| torch.chunk() | implement the 'chunk(2, dim=C)' by tensorrt plugin, see shufflenet. |
| channel shuffle | use two shuffle layers to implement channel_shuffle, see shufflenet. |
| adaptive pool | use fixed input dimension, and use regular average pooling, see shufflenet. |
| leaky relu | I wrote a leaky relu plugin, but PRelu in NvInferPlugin.h can be used, see yolov3 in branch trt4. |
| yolo layer v1 | yolo layer is implemented as a plugin, see yolov3 in branch trt4. |
| yolo layer v2 | three yolo layers implemented in one plugin, see yolov3-spp. |
| upsample | replaced by a deconvolution layer, see yolov3. |
| hsigmoid | hard sigmoid is implemented as a plugin, hsigmoid and hswish are used in mobilenetv3 |
| retinaface output decode | implement a plugin to decode bbox, confidence and landmarks, see retinaface. |
| mish | mish activation is implemented as a plugin, mish is used in yolov4 |
| prelu | mxnet's prelu activation with trainable gamma is implemented as a plugin, used in arcface |
| HardSwish | hard_swish = x * hard_sigmoid, used in yolov5 v3.0 |
| LSTM | Implemented pytorch nn.LSTM() with tensorrt api |
| Models | Device | BatchSize | Mode | Input Shape(HxW) | FPS |
|---|---|---|---|---|---|
| YOLOv3-tiny | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 333 |
| YOLOv3(darknet53) | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 39.2 |
| YOLOv3-spp(darknet53) | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 38.5 |
| YOLOv4(CSPDarknet53) | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 35.7 |
| YOLOv4(CSPDarknet53) | Xeon E5-2620/GTX1080 | 4 | FP32 | 608x608 | 40.9 |
| YOLOv4(CSPDarknet53) | Xeon E5-2620/GTX1080 | 8 | FP32 | 608x608 | 41.3 |
| YOLOv5-s | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 142 |
| YOLOv5-s | Xeon E5-2620/GTX1080 | 4 | FP32 | 608x608 | 173 |
| YOLOv5-s | Xeon E5-2620/GTX1080 | 8 | FP32 | 608x608 | 190 |
| YOLOv5-m | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 71 |
| YOLOv5-l | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 43 |
| YOLOv5-x | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 29 |
| RetinaFace(resnet50) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 90 |
| RetinaFace(mobilenet0.25) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 333 |
| ArcFace(LResNet50E-IR) | Xeon E5-2620/GTX1080 | 1 | FP32 | 112x112 | 333 |
| CRNN | Xeon E5-2620/GTX1080 | 1 | FP32 | 32x100 | 1000 |
Help wanted, if you got speed results, please add an issue or PR.
Currently, This repo is funded by Alleyes-THU AI Lab(aboutus in Chinese). We are based in Tsinghua University, Beijing, and seeking for talented interns for CV R&D. Contact me if you are interested.
Any contributions, questions and discussions are welcomed, contact me by following info.
E-mail: wangxinyu_es@163.com
WeChat ID: wangxinyu0375 (可加我微信进tensorrtx交流群)