prabhuomkar / pytorch-cpp Goto Github PK
View Code? Open in Web Editor NEWC++ Implementation of PyTorch Tutorials for Everyone
License: MIT License
C++ Implementation of PyTorch Tutorials for Everyone
License: MIT License
Why does Libtorch use more memory than Pytorch does in Python? Such as:ESRGAN
Is your feature request related to a problem? Please describe.
Describe the solution you'd like
I need to predict discrete density functions that have ordinal information.
I would like an earth movers distance implementation for my work.
I'm currently using KL divergence
EDIT:
I rolled my own: https://github.com/wesboyt/wasserstein
Add tutorial with CIFAR10 dataset.
Hello there,
I'm having this problem:
[ 21%] Building CXX object tutorials/intermediate/convolutional_neural_network/CMakeFiles/convolutional-neural-network.dir/src/main.cpp.o
/root/ml/pytorch-cpp/tutorials/intermediate/convolutional_neural_network/src/main.cpp: In function 'int main()':
/root/ml/pytorch-cpp/tutorials/intermediate/convolutional_neural_network/src/main.cpp:106:12: error: 'InferenceMode' is not a member of 'torch'
torch::InferenceMode no_grad;
^~~~~~~~~~~~~
What could be the cause?
Thank you!
Is your feature request related to a problem? Please describe.
Lot of common code is available as of now, which is creating too much redundancy of same training/iterations/testing/data-loading/saving sections.
Describe the solution you'd like
Moving common pieces of code to utils
only specific to intermediate tutorials.
Additional context
Following modules can be extracted to utils:
Currently, lets stick to this CL to be in one PR.
To Reproduce
Steps to reproduce the behavior:
Run:
python model/create_vgg19_layers_scriptmodule.py
Traceback (most recent call last):
File "model/create_vgg19_layers_scriptmodule.py", line 19, in <module>
main()
File "model/create_vgg19_layers_scriptmodule.py", line 14, in main
vgg_19_layers.save(filename)
File "/usr/lib/python3.8/site-packages/torch/nn/modules/module.py", line 575, in __getattr__
raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'Sequential' object has no attribute 'save'
Expected behavior
The script module should be generated.
Desktop (please complete the following information):
When I use the CPU to train the model, as the number of iterations increases, Loss.Backward () this function is getting slower and slower ?However, the memory on vs2017 has been stable。
`for (int i = 0; i < num_epochs; i++) {
int batch_index = 0;
for (auto &batch : *dataloader) {
auto data = batch.data.to(device);
auto target = batch.target.to(device);
auto output= ae->forward(target);
Tensor LossToTal;
//Replace torch::nn:functional::mse_loss() is Same
s->caclSSIM(data, output, LossToTal);
auto Loss =1.0f- LossToTal.mean();
sts = clock();
optimizer.zero_grad();
Loss.backward();
optimizer.step();
endt = clock();
v1 = endt - sts;
if ((batch_index + 1) % 2 == 0) {
std::cout << "Epoch [" << i << "/" << num_epochs << "], Step [" << batch_index + 1 << "/"
<< num_samples / batch_size << "], MeanLoss " << Loss.item<double>() <<
"CostTime:"<<v1<<std::endl;
}
batch_index++;
}
}`
Hi @prabhuomkar @mfl28
While generating the build system, the process appeared to be stuck at the step "Fetching CIFAR10 dataset...".
I actually restarted the process midway, thinking that the process had frozen. The download process took approximately about 30 minutes on my computer, since the CIFAR-10 binary is a large file (of size about 162 MB).
A progress bar to indicate the download progress would really be helpful.
There is an existing cmake flag SHOW_PROGRESS. to achieve this.
In PR - #41
I was able to achieve these results:
Is your feature request related to a problem? Please describe.
Hello,
I is use Libtorch for process image, OpenCV very large so try avoid. I try use JCUE C++ framework.
Describe the solution you'd like
Not able to convert torch::Tensor to juce::image. I want to read a Video file, split into images on the fly and apply Deep Learning algorithm on image in Libtorch, then turn it back to image JUCE can understand. Anyone is have example make like this?
Describe alternatives you've considered
Is here someone make OpenCV to juce::image.
https://forum.juce.com/t/opencvs-cv-mat-to-juces-image/33518
Additional context
Try integrate code with JUCE c++ GUI.
I use torch.jit.trace()
and save my own CNN_model. The return of my network is a tuple contains 3 tensors.
When I use C++ as front, I use auto output = module.forward(inputs);
to get my result. But how can I get the 3 tensors individually? and do some C++ operations?
Thank you!
Code to be made available here
Thanks for the nice code! Here is a Dockerfile to support CUDA-version pytorch-cpp. Hope it helps when you want to run the code with GPUs.
FROM nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04
LABEL maintainer="[email protected]"
# Fix the apt-get error from nvidia-docker
RUN rm /etc/apt/sources.list.d/cuda.list \
&& rm /etc/apt/sources.list.d/nvidia-ml.list \
&& apt-key del 7fa2af80 \
&& apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/3bf863cc.pub \
&& apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/7fa2af80.pub
# Install basics
RUN apt-get update -y \
&& apt-get install -y apt-utils git curl ca-certificates tree htop wget libssl-dev unzip \
&& rm -rf /var/lib/apt/lists/*
# Install g++-8 gcc-8
RUN apt-get update && apt-get install -y gcc-8 g++-8 \
&& update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-8 60 --slave /usr/bin/g++ g++ /usr/bin/g++-8 \
&& update-alternatives --config gcc \
&& rm -rf /var/lib/apt/lists/*
# Install cmake
RUN apt-get purge -y cmake \
&& mkdir /root/temp \
&& cd /root/temp \
&& wget https://github.com/Kitware/CMake/releases/download/v3.23.4/cmake-3.23.4.tar.gz \
&& tar -xzvf cmake-3.23.4.tar.gz \
&& cd cmake-3.23.4 \
&& bash ./bootstrap \
&& make \
&& make install \
&& cmake --version \
&& rm -rf /root/temp \
&& rm -rf /var/lib/apt/lists/*
# Install libtorch
RUN cd /root/ \
&& wget https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.12.1%2Bcu102.zip -O libtorch.zip \
&& unzip libtorch.zip
# Install pytorch-cpp
RUN cd /root \
&& wget https://github.com/prabhuomkar/pytorch-cpp/archive/refs/tags/v1.12.tar.gz \
&& tar -xzvf v1.12.tar.gz
RUN cd /root/pytorch-cpp-1.12 \
&& cmake -B build \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_PREFIX_PATH=/root/libtorch/share/cmake/Torch \
-D CREATE_SCRIPTMODULES=ON \
&& cmake --build build
WORKDIR /root
Describe the bug
There are two parts to this bug.
First off I am failing to build because CMAKE is telling me that the MD5sums of the 4 files from the classic source that is referenced are different to those in the CMakeLists file.
Bypassing that issue, as I reach the part of the example where the dataset is loaded, a std::bad_alloc exception is thrown.
To Reproduce
Steps to reproduce the behavior:
Expected behavior
The project would successfully pass through the input pipeline section of the example
Desktop (please complete the following information):
Additional context
I get the same issue using the auto-downloaded version of libtorch
Is your feature request related to a problem? Please describe.
C++ code seems to vary a bit in Beginner and Intermediate tutorials as of now. Need to make use of it at maximum potential for best performance.
Describe the solution you'd like
Following things need to be standardized:
"\n"
over std::endl
Additional context
Code should look consistent through out all set of tutorials.
Currently, lets stick to this CL to be in one PR.
I found that in this project, there is no code using ModuleList. I recently encountered some problems in using ModuleList.If you can give some examples, thank you very much.
Describe the bug
Users facing issues in running the code #26
Expected behavior
Users should be able to follow README and face no issues in running the code.
So, adding more detailed tutorial in README with all variations for running the code.
Additional context
Linking to installation of OS specific CMake, Libtorch will also be helpful.
Is your feature request related to a problem? Please describe.
I would like to train a CNN-classifier with my custom data using the widely-used models like ResNet series,I found it is useful to initialize the model weights with ImageNet pretrained weights, and it is easy to implement with the
torch::load
API when the image channels of my dataset is 3, the same as ImageNet,under which situation no change should be made to theConv1
layer.
It is the other situation when I try to train with gray images,as theConv1
weights is supposed to be ofin_channels=3
, In the python fronten, I guess this maybe solved but imdieatly repalce the model.conv1 like this:
model.conv1 = nn.Conv2d(in_channels=1, out_channels=64, kernel_size=7, stride=2, padding=3, bias=False)
but as for the C++ fronten, repalcing seems not work:
model->conv1 = torch::nn::Conv2d(torch::nn::Conv2dOptions(in_channels, 64, 7).stride(2).padding(3).bias(false).dilation(1));
Describe the solution you'd like
Correcting the API use to rightly loading the pretrained weights.
Describe alternatives you've considered
Maybe a pretrained model on gray image dataset would bypass the problem.
Additional context
Exception occurs during the model forward process.
* https://discuss.pytorch.org/t/libtorch-for-raspberry-pi/63107 This actually solved the build problem.
But now there are new Problems showing up:
cmake --build build
Scanning dependencies of target pytorch-cpp
[ 2%] Building CXX object CMakeFiles/pytorch-cpp.dir/main.cpp.o
[ 4%] Linking CXX executable pytorch-cpp
[ 4%] Built target pytorch-cpp
Scanning dependencies of target feedforward-neural-network
[ 7%] Building CXX object tutorials/basics/feedforward_neural_network/CMakeFiles/feedforward-neural-network.dir/src/main.cpp.o
/home/pi/Desktop/PytorchC++TestInternet/pytorch-cpp-master/tutorials/basics/feedforward_neural_network/src/main.cpp: In function ‘int main()’:
/home/pi/Desktop/PytorchC++TestInternet/pytorch-cpp-master/tutorials/basics/feedforward_neural_network/src/main.cpp:71:48: error: ‘cross_entropy’ is not a member of ‘torch::nn::functional’
auto loss = torch::nn::functional::cross_entropy(output, target);
^~~~~~~~~~~~~
/home/pi/Desktop/PytorchC++TestInternet/pytorch-cpp-master/tutorials/basics/feedforward_neural_network/src/main.cpp:74:39: error: expected primary-expression before ‘double’
running_loss += loss.item() * data.size(0);
^~~~~~
/home/pi/Desktop/PytorchC++TestInternet/pytorch-cpp-master/tutorials/basics/feedforward_neural_network/src/main.cpp:111:44: error: ‘cross_entropy’ is not a member of ‘torch::nn::functional’
auto loss = torch::nn::functional::cross_entropy(output, target);
^~~~~~~~~~~~~
/home/pi/Desktop/PytorchC++TestInternet/pytorch-cpp-master/tutorials/basics/feedforward_neural_network/src/main.cpp:113:35: error: expected primary-expression before ‘double’
running_loss += loss.item() * data.size(0);
^~~~~~
make[2]: *** [tutorials/basics/feedforward_neural_network/CMakeFiles/feedforward-neural-network.dir/build.make:63: tutorials/basics/feedforward_neural_network/CMakeFiles/feedforward-neural-network.dir/src/main.cpp.o] Fehler 1
make[1]: *** [CMakeFiles/Makefile2:354: tutorials/basics/feedforward_neural_network/CMakeFiles/feedforward-neural-network.dir/all] Fehler 2
make: *** [Makefile:84: all] Fehler 2Is this a Problem of Libtorch 1.3 instead of the leatest Version?
Do you know that?
I have a similar issue while trying to build convolutional_neural_network from repo.
cmake --build . --config Release Scanning dependencies of target freespace_torch [ 33%] Building CXX object CMakeFiles/freespace_torch.dir/src/convnet.cpp.o In file included from /home/fugurcal/freespace_torch/src/convnet.cpp:2:0: /home/fugurcal/freespace_torch/include/convnet.h:13:20: error: ‘BatchNorm2d’ is not a member of ‘torch::nn’ torch::nn::BatchNorm2d(16), ^~~~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:13:20: note: suggested alternative: ‘BatchNorm’ torch::nn::BatchNorm2d(16), ^~~~~~~~~~~ BatchNorm /home/fugurcal/freespace_torch/include/convnet.h:14:20: error: ‘ReLU’ is not a member of ‘torch::nn’ torch::nn::ReLU(), ^~~~ /home/fugurcal/freespace_torch/include/convnet.h:15:20: error: ‘MaxPool2d’ is not a member of ‘torch::nn’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:15:41: error: ‘MaxPool2dOptions’ is not a member of ‘torch::nn’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:15:41: note: suggested alternative: ‘Conv2dOptions’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~~~~~~~~ Conv2dOptions /home/fugurcal/freespace_torch/include/convnet.h:16:5: error: could not convert ‘{torch::nn::Conv2d((* &(& torch::nn::ConvOptions<2>(1, 16, torch::ExpandingArray<2, long int>(5)).torch::nn::ConvOptions<2>::stride(torch::ExpandingArray<2, long int>(1)))->torch::nn::ConvOptions<2>::padding(torch::ExpandingArray<2, long int>(2)))), <expression error>, <expression error>, <expression error>}’ from ‘<brace-enclosed initializer list>’ to ‘torch::nn::Sequential’ }; ^ /home/fugurcal/freespace_torch/include/convnet.h:20:20: error: ‘BatchNorm2d’ is not a member of ‘torch::nn’ torch::nn::BatchNorm2d(32), ^~~~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:20:20: note: suggested alternative: ‘BatchNorm’ torch::nn::BatchNorm2d(32), ^~~~~~~~~~~ BatchNorm /home/fugurcal/freespace_torch/include/convnet.h:21:20: error: ‘ReLU’ is not a member of ‘torch::nn’ torch::nn::ReLU(), ^~~~ /home/fugurcal/freespace_torch/include/convnet.h:22:20: error: ‘MaxPool2d’ is not a member of ‘torch::nn’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:22:41: error: ‘MaxPool2dOptions’ is not a member of ‘torch::nn’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~~~~~~~~ /home/fugurcal/freespace_torch/include/convnet.h:22:41: note: suggested alternative: ‘Conv2dOptions’ torch::nn::MaxPool2d(torch::nn::MaxPool2dOptions(2).stride(2)) ^~~~~~~~~~~~~~~~ Conv2dOptions /home/fugurcal/freespace_torch/include/convnet.h:23:5: error: could not convert ‘{torch::nn::Conv2d((* &(& torch::nn::ConvOptions<2>(16, 32, torch::ExpandingArray<2, long int>(5)).torch::nn::ConvOptions<2>::stride(torch::ExpandingArray<2, long int>(1)))->torch::nn::ConvOptions<2>::padding(torch::ExpandingArray<2, long int>(2)))), <expression error>, <expression error>, <expression error>}’ from ‘<brace-enclosed initializer list>’ to ‘torch::nn::Sequential’ }; ^ CMakeFiles/freespace_torch.dir/build.make:62: recipe for target 'CMakeFiles/freespace_torch.dir/src/convnet.cpp.o' failed make[2]: *** [CMakeFiles/freespace_torch.dir/src/convnet.cpp.o] Error 1 CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/freespace_torch.dir/all' failed make[1]: *** [CMakeFiles/freespace_torch.dir/all] Error 2 Makefile:83: recipe for target 'all' failed make: *** [all] Error 2
I am using torch 1.5, any idea how solve this problem?
Originally posted by @FarukUgurcali in #37 (comment)
I'm always frustrated when google spits out only python and tensorflow human detection implementation 🤦♂️
It would be great if this libtroch tutorial can show some human detection implementation which have high accuracies.
Alternatives you've considered:
OpenPose (Super Slow, low FPS even on RTX 3090).
Would like to use OpenCL for AMD and ARM Mali GPUs.
Thanks.
This is g great project!
Recently, I find using '.to(at::kCPU)' from CUDA takes a very long time, when I use my model to inference on GPU. Like this:
vector<torch::jit::IValue> inputs = {input_data.input_ids_pt_,
input_data.attention_mask_pt_,
input_data.token_type_ids_pt_};
auto pred_res = (static_cast<torch::jit::script::Module*>(bert_model_))->forward(inputs); // About 12 milliseconds
auto logits = pred_res.toTensor().to(at::kCPU); // About 80 milliseconds
So what should I do to reduce the time consuming?
Looking forward to your reply, thanks!
Is your feature request related to a problem? Please describe.
I am the author of Siv3DTorch (https://github.com/QuantScientist/Siv3DTorch) which integrates OpenSiv3d with Libtorch C++.
At the moment Siv3D does not support CMake and therefore all integration efforts are on VC 19.
One main burning issue that I have is reading and writing Images/video frames from and to Siv3D without using OpenCV.
Describe the solution you'd like
I would love to have an Image conversation method between the two frameworks, either using stb_image or libpng (used by Siv3D).
Additional context
The whole scenario is described here:
Siv3D/OpenSiv3D#534
The source code is here:
https://github.com/QuantScientist/Siv3DTorch/blob/master/src/loadmodel003.cpp
Many thanks for your help,
@mfl28 can we add some instructions on how to build a single tutorial and if necessary its datasets only?
I believe many users want to try "X" tutorial from all and if there is an easy to follow guide that would be great :)
PS: Added this as an issue for visibility and referencing.
Hi, your work is very awesome! And I hope you can add support for yolov3 model if you have time because yolov3 is very popular in object detection.
@prabhuomkar thanks for the this codebase , just wanted to knwo run your tutorials ??
Is your feature request related to a problem? Please describe.
There is (afaik) no pre-c++17 way to do file operations such as iterating over files in a folder in a way that is portable across operating systems without additional external dependencies such as boost.Filesystem
. This makes it hard to implement e.g. custom datasets such as an ImageFolderDataset
.
Describe the solution you'd like
The <filesystem>
header - officially part of C++ as of version 17 - makes it possible to easily handle filesystem related operations in an OS-independent way. Compiling for c++17 requires a change in the CMake configurations. C++17 is supported by most compiler versions of the last years - that includes all compilers currently used in our CI-workflows.
Hi,
I have found that current examples are on existing pytorch datasets. There are no examples on a custom dataset. I have a tutorial of sentence classification on a custom dataset.
https://github.com/avinashsai/pytorch_cpp/tree/master/sentence_classification_cnn
Please, let me know if it can be added.
Hi,
I am unable to read the images in the MNIST dataset.
below I have listed the error that I am receiving. Can you please help me with it.
Thanks
: ~/Projects/cnn/build$ ./cnn
Convolutional Neural Network
CUDA available. Training on GPU.
terminate called after throwing an instance of 'c10::Error'
what(): Error opening images file at ../../../../data/mnist/train-images-idx3-ubyte
Exception raised from read_images at ../torch/csrc/api/src/data/datasets/mnist.cpp:67 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits, std::allocator >) + 0x6b (0x7f9f8f93e0db in /home/paras/Libraries/libtorch/lib/libc10.so)
frame #1: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&) + 0xce (0x7f9f8f939d2e in /home/paras/Libraries/libtorch/lib/libc10.so)
frame #2: + 0x43847a2 (0x7f9f0b77e7a2 in /home/paras/Libraries/libtorch/lib/libtorch_cpu.so)
frame #3: torch::data::datasets::MNIST::MNIST(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&, torch::data::datasets::MNIST::Mode) + 0x46 (0x7f9f0b77f846 in /home/paras/Libraries/libtorch/lib/libtorch_cpu.so)
frame #4: main + 0x121 (0x55670d49790c in ./cnn)
frame #5: __libc_start_main + 0xf3 (0x7f9ec30140b3 in /lib/x86_64-linux-gnu/libc.so.6)
frame #6: _start + 0x2e (0x55670d49752e in ./cnn)
Aborted (core dumped)
Required via pytorch/pytorch#34862 (comment)
Is your feature request related to a problem? Please describe.
N/A
Describe the solution you'd like
Reference
Describe alternatives you've considered
N/A
Additional context
Bullet proofing the tutorials by testing the builds on Windows and available C++ compilers
Code to be made available here
Problem
Scanning dependencies of target pytorch-cpp
[ 2%] Building CXX object CMakeFiles/pytorch-cpp.dir/main.cpp.o
[ 4%] Linking CXX executable pytorch-cpp
/home/pi/Desktop/LibTorch/libtorch/lib/libtorch.so: file not recognized: file format not recognized
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/pytorch-cpp.dir/build.make:87: pytorch-cpp] Fehler 1
make[1]: *** [CMakeFiles/Makefile2:327: CMakeFiles/pytorch-cpp.dir/all] Fehler 2
make: *** [Makefile:84: all] Fehler 2
Help Me Pls
I think this is a Problem of my specific System, but I have no Idea what the issue is
Desktop:
Initialize Intermediate Tutorials with Convolutional Neural Network
windows-2016
, macos-10.15
) and must be updated.checkout
and setup-cmake
) and must be updated to their latest versions.Describe the bug
The classical MNIST source (http://yann.lecun.com/exdb/mnist/) seems to be very unreliable lately. We should switch to an alternative mirror as used in pytorch MNIST dataset (https://ossci-datasets.s3.amazonaws.com/mnist).
This has also been reported in #83 .
To Reproduce
Steps to reproduce the behavior:
Expected behavior
MNIST dataset is downloaded correctly.
Is your feature request related to a problem? Please describe.
I want have load ImageNet, not MNIST.
Describe the solution you'd like
To have a dataloader for ImageNet that works with training one of the models.
Thnaks.
Is your feature request related to a problem? Please describe.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.