allentdan / libtorchsegmentation Goto Github PK

It was very helpful to implement segmentation using libTorch with C++. Thank you.
I'm trying to implement Mask R-CNN for Instance Segmentation, but I'm not fully familiar with LibTorch yet, so I'm having a hard time.
I'd like to include it in your code, so could you give me some hints?

您好，我要问一下，提供的例子里面支持定位吗？比如给的image预测里面，图中有两个人，我能否把她们识别并且分别标注出来呢？

图中框出两人的大致轮廓。

Hi! Great repo!

I was exactly looking for a C++ interface to Qubvel pytorch models. Is there any possibility that you will provide the interface for the Unet++ model with SE-ResNet-50 encoder?

Thank you!

Hi, and thank you very much for making this code available. Very helpful!

I have the examples running great, but now I would like to train and use DeepLabV3+ model.

How do I generate weights and train using this model?

Thanks!

How to convert pt-model to ONNX-model? I need to load and run an ONNX-model using opencv.
Thanks!

你好，看了下unet结构的实施，几个问题请教下：

• 分割头上最后为什么还接了一层upsample，按道理conv出来的特征就是(b,n_class,h,w)已经恢复到原图大小了，那这层上采样貌似不是必要的。
• dice_loss的实现：发现dice_coef的公式用了dice_coef=inter/(pred_area + label_area - inter)，但这不是iou的公式吗？dice_coef不是应该等于2*inter/(pred_area + label_area)吗？

Hi, Great work guys!

This is a great work indeed and I would like to contribute into it by adding a support for another network. https://github.com/W-zx-Y/DANNet. Could you let me know the process and where is better to start? I'm sure you have done this before so I could use your experience instead of going completely on my own.

Thanks!

您好！学习了您的基于libtorch写的语义分割算法，我感觉收益匪浅。
在源码上还有一些问题希望能得到解答。在预测的时候看到initialize读入了resnet34.pt模型，然后在LoadWeight时又读取了segmentor.pt模型，这两个模型是一样的吗？在读入segmentor.pt时model这个类参数是不是会被覆盖，如果会被覆盖的话之前读入的resnet34这个模型好像没有必要？还有一个问题就是，训练的时候也要读入骨干网络的预训练模型，训练后保存的模型是不是已经包括了骨干网络，或者是只有一个分割头，骨干网络沿用预训练的网络参数？

还望您百忙之中能够解答，非常感谢！

非常感谢大佬开源这个项目，这对于想在C++上利用libtorch部署的人的帮助真的很大。
最近在研究solov2，但网上很多资料都是基于python的，而关于在C++上部署的非常少，不知道大佬可不可以出个相关的教程指导一下？可能这要求有点过分啦哈哈哈，但还是希望大佬能考虑一下

May you supply a multi-class semantic segmentation example, thanks!

Hi,

Firstly, thank you for sharing this project. It would be very helpful to work with Pytorch in C++ using this library.

I tried to build the project using the commands mentioned as follows:
cd build
cmake ..
make
./LibtorchSegmentation

The "cmake .." command is executed successfully. However, the "make" command shows an error message "make: *** No targets specified and no makefile found. Stop."
There is no makefile in the LibtorchSegmentation project directory. And therefore, there is no LibtorchSegmentation file generated in the build directory.

Please would you let me know how I can work around this?

Thanks!

Hi,

Thank you so much for your support with my last issue.
Now I can able to run run the given code in Linux platform also.

Please let me know how to do inference with GPU.
When try to do inference using video input it is very slow. Hope GPU usage will improve the efficiency.

It is possible to train multiple objects like people, animals etc.. in a model.

Thanks and regards,
Prabhakar M

Can training be accelerated using CUDA? If so, how can it be done?

1、加载的预训练模型resnet34 训练自己的数据集 train loss 和val loss 训练几轮就变成0
2、然后有训练集的图像进行测试 结果全黑

I am trying run this code in ubuntu OS but it is giving some file missing error. Kindly share me cmake file for the given post.

看了您写的Libtorch的博客，受益匪浅，慕名而来，star了一下您的项目。
Libtorch现阶段支持复数运算，博主什么时候能写下复数相关的运算的博客吗，等待更新。
博主有没有相关资料的推荐，官网只给了接口，啥也没讲……

(py37) ql@ql:~/projects/LibtorchSegmentation-main/test/build$ make
Scanning dependencies of target resnet34
[ 25%] Building CXX object CMakeFiles/resnet34.dir/resnet34.cpp.o
make[2]: *** 没有规则可制作目标“/usr/local/cuda-11.3/lib64/libcudart.so/usr/local/cuda-11.3/lib64/libcublas.so”，由“resnet34” 需求。 停止。
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/resnet34.dir/all' failed
make[1]: *** [CMakeFiles/resnet34.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

大佬，看你的博客收获巨多，谢谢大佬！
我自己用C++搭的模型加载python训练的参数pth文件一直报错，求教！

DiceLoss function in loss.h，just like iou.

Hi,
Thanks for your code.

Could you include your code impllemented by python ?
Thanks!

how can i use Quantization-aware training in this project

cannot run the CMakeLists file following the instructions.
did not find the src/main.cpp

Hi, my dataset has images and corresponding masks, but it has no json files. The dataset folders just like below:

Dataset
├── train
│   ├── 0.png
│   ├── 0_mask.png
│   └......
├── val
│   ├── 3.png
│   ├── 3_mask.png
│   └......

May I ask how to train on this dataset without json file?
Thanks!

Author, I have a question for you. I want to change the network to UNet, and change the FPN in the train.cpp and resnet34.cpp files to UNet, but the training error occurs. I changed to PAN and it works. Do you know what the reason is. thank you