Skip to content

A c++ trainable semantic segmentation library based on libtorch (pytorch c++). Backbone: VGG, ResNet, ResNext. Architecture: FPN, U-Net, PAN, LinkNet, PSPNet, DeepLab-V3, DeepLab-V3+ by now.

License

Notifications You must be signed in to change notification settings

AllentDan/LibtorchSegmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

English | δΈ­ζ–‡

logo
C++ library with Neural Networks for Image
Segmentation based on LibTorch.

⭐Please give a star if this project helps you.⭐

The main features of this library are:

  • High level API (just a line to create a neural network)
  • 7 models architectures for binary and multi class segmentation (including legendary Unet)
  • 15 available encoders
  • All encoders have pre-trained weights for faster and better convergence
  • 35% or more inference speed boost compared with pytorch cuda, same speed for cpu. (Unet tested in rtx 2070s).

Visit Libtorch Tutorials Project if you want to know more about Libtorch Segment library.

πŸ“‹ Table of content

  1. Quick start
  2. Examples
  3. Train your own data
  4. Models
    1. Architectures
    2. Encoders
  5. Installation
  6. Thanks
  7. To do list
  8. Citing
  9. License
  10. Related repository

⏳ Quick start

1. Create your first Segmentation model with Libtorch Segment

A resnet34 trochscript file is provided here. Segmentation model is just a LibTorch torch::nn::Module, which can be created as easy as:

#include "Segmentor.h"
auto model = UNet(1, /*num of classes*/
                  "resnet34", /*encoder name, could be resnet50 or others*/
                  "path to resnet34.pt"/*weight path pretrained on ImageNet, it is produced by torchscript*/
                  );
  • see table with available model architectures
  • see table with available encoders and their corresponding weights

2. Generate your own pretrained weights

All encoders have pretrained weights. Preparing your data the same way as during weights pre-training may give your better results (higher metric score and faster convergence). And you can also train only the decoder and segmentation head while freeze the backbone.

import torch
from torchvision import models

# resnet34 for example
model = models.resnet34(pretrained=True)
model.eval()
var=torch.ones((1,3,224,224))
traced_script_module = torch.jit.trace(model, var)
traced_script_module.save("resnet34.pt")

Congratulations! You are done! Now you can train your model with your favorite backbone and segmentation framework.

πŸ’‘ Examples

  • Training model for person segmentation using images from PASCAL VOC Dataset. "voc_person_seg" dir contains 32 json labels and their corresponding jpeg images for training and 8 json labels with corresponding images for validation.
Segmentor<FPN> segmentor;
segmentor.Initialize(0/*gpu id, -1 for cpu*/,
                    512/*resize width*/,
                    512/*resize height*/,
                    {"background","person"}/*class name dict, background included*/,
                    "resnet34"/*backbone name*/,
                    "your path to resnet34.pt");
segmentor.Train(0.0003/*initial leaning rate*/,
                300/*training epochs*/,
                4/*batch size*/,
                "your path to voc_person_seg",
                ".jpg"/*image type*/,
                "your path to save segmentor.pt");
  • Predicting test. A segmentor.pt file is provided in the project here. It is trained through a FPN with ResNet34 backbone for a few epochs. You can directly test the segmentation result through:
cv::Mat image = cv::imread("your path to voc_person_seg\\val\\2007_004000.jpg");
Segmentor<FPN> segmentor;
segmentor.Initialize(0,512,512,{"background","person"},
                      "resnet34","your path to resnet34.pt");
segmentor.LoadWeight("segmentor.pt"/*the saved .pt path*/);
segmentor.Predict(image,"person"/*class name for showing*/);

the predicted result shows as follow:

πŸ§‘β€πŸš€ Train your own data

  • Create your own dataset. Using labelme through "pip install" and label your images. Split the output json files and images into folders just like below:
Dataset
β”œβ”€β”€ train
β”‚   β”œβ”€β”€ xxx.json
β”‚   β”œβ”€β”€ xxx.jpg
β”‚   β””......
β”œβ”€β”€ val
β”‚   β”œβ”€β”€ xxxx.json
β”‚   β”œβ”€β”€ xxxx.jpg
β”‚   β””......
  • Training or testing. Just like the example of "voc_person_seg", replace "voc_person_seg" with your own dataset path.
  • Refer to training tricks to improve your final training performance.

πŸ“¦ Models

Architectures

Encoders

  • ResNet
  • ResNext
  • VGG

The following is a list of supported encoders in the Libtorch Segment. All the encoders weights can be generated through torchvision except resnest. Select the appropriate family of encoders and click to expand the table and select a specific encoder and its pre-trained weights.

ResNet
Encoder Weights Params, M
resnet18 imagenet 11M
resnet34 imagenet 21M
resnet50 imagenet 23M
resnet101 imagenet 42M
resnet152 imagenet 58M
ResNeXt
Encoder Weights Params, M
resnext50_32x4d imagenet 22M
resnext101_32x8d imagenet 86M
ResNeSt
Encoder Weights Params, M
timm-resnest14d imagenet 8M
timm-resnest26d imagenet 15M
timm-resnest50d imagenet 25M
timm-resnest101e imagenet 46M
timm-resnest200e imagenet 68M
timm-resnest269e imagenet 108M
timm-resnest50d_4s2x40d imagenet 28M
timm-resnest50d_1s4x24d imagenet 23M
SE-Net
Encoder Weights Params, M
senet154 imagenet 113M
se_resnet50 imagenet 26M
se_resnet101 imagenet 47M
se_resnet152 imagenet 64M
se_resnext50_32x4d imagenet 25M
se_resnext101_32x4d imagenet 46M
VGG
Encoder Weights Params, M
vgg11 imagenet 9M
vgg11_bn imagenet 9M
vgg13 imagenet 9M
vgg13_bn imagenet 9M
vgg16 imagenet 14M
vgg16_bn imagenet 14M
vgg19 imagenet 20M
vgg19_bn imagenet 20M

πŸ›  Installation

Dependency:

Windows:

Configure the environment for libtorch development. Visual studio and Qt Creator are verified for libtorch1.7x release.

Linux && MacOS:

Install libtorch and opencv.

For libtorch, follow the official pytorch c++ tutorials here.

For opencv, follow the official opencv install steps here.

If you have already configured them both, congratulations!!! Download the pretrained weight here and a demo .pt file here into weights.

Building shared or static library -DBUILD_SHARED=<TRUE/FALSE>:

export Torch_DIR='/path/to/libtorch'
cd build
cmake -DBUILD_SHARED=TRUE ..
make
sudo make install

Building tests:

cd test
mkdir build && cd build
cmake ..
make
./resnet34 ../../voc_person_seg/val/2007_003747.jpg ../../weights/resnet34.pt ../../weights/segmentor.pt

⏳ ToDo

  • More segmentation architectures and backbones
    • UNet++ [paper]
    • ResNest
    • Se-Net
    • ...
  • Data augmentations
    • Random horizontal flip
    • Random vertical flip
    • Random scale rotation
    • ...
  • Training tricks
    • Combined dice and cross entropy loss
    • Freeze backbone
    • Multi step learning rate schedule
    • ...

🀝 Thanks

By now, these projects helps a lot.

πŸ“ Citing

@misc{Chunyu:2021,
  Author = {Chunyu Dong},
  Title = {Libtorch Segment},
  Year = {2021},
  Publisher = {GitHub},
  Journal = {GitHub repository},
  Howpublished = {\url{https://github.com/AllentDan/SegmentationCpp}}
}

πŸ›‘οΈ License

Project is distributed under MIT License.

Related repository

Based on libtorch, I released following repositories:

Last but not least, don't forget your star...

Feel free to commit issues or pull requests, contributors wanted.

stargazers over time