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/mxnet/tensorflown 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.
26 May 2022. triple-Mu: YOLOv5 python script with CUDA Python API.23 May 2022. yhpark: Real-ESRGAN, Practical Algorithms for General Image/Video Restoration.19 May 2022. vjsrinivas: YOLOv3 TRT8 support and Python script.15 Mar 2022. sky_hole: Swin Transformer - Semantic Segmentation.19 Oct 2021. liuqi123123 added cuda preprossing for yolov5, preprocessing + inference is 3x faster when batchsize=8.18 Oct 2021. xupengao: YOLOv5 updated to v6.0, supporting n/s/m/l/x/n6/s6/m6/l6/x6.31 Aug 2021. FamousDirector: update retinaface to support TensorRT 8.0.27 Aug 2021. HaiyangPeng: add a python wrapper for hrnet segmentation.1 Jul 2021. freedenS: DE⫶TR: End-to-End Object Detection with Transformers. First Transformer model!10 Jun 2021. upczww: EfficientNet b0-b8 and l2.23 May 2021. SsisyphusTao: CenterNet DLA-34 with DCNv2 plugin.17 May 2021. ybw108: arcface LResNet100E-IR and MobileFaceNet.6 May 2021. makaveli10: scaled-yolov4 yolov4-csp.29 Apr 2021. upczww: hrnet segmentation w18/w32/w48, ocr branch also.28 Apr 2021. aditya-dl: mobilenetv2, alexnet, densenet121, mobilenetv3 with python API.
- Install the dependencies.
- A guide for quickly getting started, taking lenet5 as a demo.
- The .wts file content format
- 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)
- TensorRT 7.x
- TensorRT 8.x(Some of the models support 8.x)
Each folder has a readme inside, which explains how to run the models inside.
Following models are implemented.
| Name | Description |
|---|---|
| mlp | the very basic model for starters, properly documented |
| 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, v4 |
| mnasnet | MNASNet with depth multiplier of 0.5 from the paper |
| mobilenet | MobileNet v2, v3-small, v3-large |
| resnet | resnet-18, resnet-50 and resnext50-32x4d are implemented |
| senet | se-resnet50 |
| shufflenet | ShuffleNet v2 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 v1.0-v6.0, pytorch implementation from ultralytics/yolov5 |
| retinaface | resnet50 and mobilnet0.25, weights from biubug6/Pytorch_Retinaface |
| arcface | LResNet50E-IR, LResNet100E-IR and MobileFaceNet, 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 and hrnet-semantic-segmentation, pytorch implementation from HRNet-Image-Classification and HRNet-Semantic-Segmentation |
| psenet | PSENet Text Detection, tensorflow implementation from liuheng92/tensorflow_PSENet |
| ibnnet | IBN-Net, pytorch implementation from XingangPan/IBN-Net, ECCV2018 |
| unet | U-Net, pytorch implementation from milesial/Pytorch-UNet |
| repvgg | RepVGG, pytorch implementation from DingXiaoH/RepVGG |
| lprnet | LPRNet, pytorch implementation from xuexingyu24/License_Plate_Detection_Pytorch |
| refinedet | RefineDet, pytorch implementation from luuuyi/RefineDet.PyTorch |
| densenet | DenseNet-121, from torchvision.models |
| rcnn | FasterRCNN and MaskRCNN, model from detectron2 |
| tsm | TSM: Temporal Shift Module for Efficient Video Understanding, ICCV2019 |
| scaled-yolov4 | yolov4-csp, pytorch from WongKinYiu/ScaledYOLOv4 |
| centernet | CenterNet DLA-34, pytorch from xingyizhou/CenterNet |
| efficientnet | EfficientNet b0-b8 and l2, pytorch from lukemelas/EfficientNet-PyTorch |
| detr | DE⫶TR, pytorch from facebookresearch/detr |
| swin-transformer | Swin Transformer - Semantic Segmentation, only support Swin-T. The Pytorch implementation is microsoft/Swin-Transformer |
| real-esrgan | Real-ESRGAN. The Pytorch implementation is real-esrgan |
The .wts files can be downloaded from model zoo for quick evaluation. But it is recommended to convert .wts from pytorch/mxnet/tensorflow model, so that you can retrain your own model.
GoogleDrive | BaiduPan pwd: uvv2
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(darknet53) | Xeon E5-2620/GTX1080 | 1 | INT8 | 608x608 | 71.4 |
| 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 v3.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 142 |
| YOLOv5-s v3.0 | Xeon E5-2620/GTX1080 | 4 | FP32 | 608x608 | 173 |
| YOLOv5-s v3.0 | Xeon E5-2620/GTX1080 | 8 | FP32 | 608x608 | 190 |
| YOLOv5-m v3.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 71 |
| YOLOv5-l v3.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 43 |
| YOLOv5-x v3.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 29 |
| YOLOv5-s v4.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 142 |
| YOLOv5-m v4.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 71 |
| YOLOv5-l v4.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 40 |
| YOLOv5-x v4.0 | Xeon E5-2620/GTX1080 | 1 | FP32 | 608x608 | 27 |
| RetinaFace(resnet50) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 90 |
| RetinaFace(resnet50) | Xeon E5-2620/GTX1080 | 1 | INT8 | 480x640 | 204 |
| RetinaFace(mobilenet0.25) | Xeon E5-2620/GTX1080 | 1 | FP32 | 480x640 | 417 |
| 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.
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，备注：tensorrtx)