header only, deep learning framework with no dependencies other than libtorch
Enabled to build with libtorch1.6 -> libtorch1.7.1 -> libtorch1.9.0 -> libtorch1.12.1 -> ** libtorch2.5.0 **
also needs cuDNN
I am currently trying to upgrade libtorch to version 2.x. The current libtorch version of cpp_torch is 1.12.1, but I was hesitant to work on it because I thought upgrading to 2.0 would be a major change.
Notes.
Most of the code in cpp_torch did not require modification, but some functions are no longer available.
mnist datasets can no longer be downloaded, so examples of downloading and running datasets are not available.
torch_jit and some functions in torchvision are also Some functions of torch_jit and torchvision have also been moved to python, so it is no longer possible to do everything in C++ as in the examples provided.
The OpenCV version has also been changed.
The following are no longer required in libtorch-2.5.
LinkOption(libtorch1.9.0):
-INCLUDE:?warp_size@cuda@at@@YAHXZ
LinkOption(libtorch1.12.1):
-INCLUDE:?warp_size@cuda@at@@YAHXZ -INCLUDE:?_torch_cuda_cu_linker_symbol_op_cuda@native@at@@YA?AVTensor@2@AEBV32@@Z
Enabled to build with libtorch1.5
Directory structure
This project aims to be a wrapper for libtorch to make tiny-dnn compatible with GPU. tiny-dnn really great, but unfortunately it can not be calculated on a GPU. At first glance, this header-only framework aims to be used as written in tiny-dnn.
Include path settings
Libtorch_path/include
Libtorch_path/include/torch/csrc/api/include
cpp_torch_path
cpp_torch_path/include
Library path setting
Libtorch_path/lib
cpp_torch_path
Minimum include file
#include "cpp_torch.h"
progress
tiny_dnn
tiny_dnn::progress_display disp(train_images.size());
cpp_torch
cpp_torch::progress_display disp(train_images.size());
cpp_torch::progress_display disp(train_images.size());
data set download
What you can do is still limited.
| options | description | default | |
|---|---|---|---|
| USE_WINDOWS | ON | ||
| USE_COLOR_CONSOLE | ON | ||
| USE_ZLIB | ON | ||
| USE_IMAGE_UTIL | ON | ||
| USE_OPENCV_UTIL | OpenCV >= 2.3 | OFF | ex. C:\dev\opencv-3.4.0 |
MNIS
CIFAR10
DCGAN
Do you have a favorite cute person?
This app was created with C # for GUI and C ++ only for core processing. Python is not required.
visual studio 2015,2017,2019
libtorch Please adapt the version of cuda to your environment
BSD 3-Clause License Copyright (c) 2013, Taiga Nomi
tiny_dnn was good, but unfortunately development has clearly stopped. Therefore, we created cpp_torch that can be used instead.
If you are building in C++, Even if Python or pytorch (libtorch) is changed Should work. Will the Python app function correctly next month? Is there a guarantee that customers will not update python, etc.?