README_libtorch1.5.md

Enabled to build with libtorch1.5

CUDA 9.2 is required for calculation using GPU.

also needs cuDNN v7.6.5 (November 5th, 2019), for CUDA 9.2

!!Currently, the cause is unknown, but the result is incorrect in CPU mode of libtorch 1.5.
It's not a smart method, but it has been fixed to give the correct result.

Infomation
[libtorch]Same model in CUDA and CPU got different result? #26456

can solve the problem that is not working on GPU (CUDA) (torch::cuda::is_available() returns false).

####How to resolve this issue
add the following argument -INCLUDE:?warp_size@cuda@at@@YAHXZ
Linker -> Command line -> Addtitional options
-INCLUDE:?warp_size@cuda@at@@YAHXZ

reference

Windows Libtorch Ver1.5 did't operate at the GPU mode , it operated at the CPU mode. #37568
Libtorch 1.5 failed to access CUDA devices #37481

The current code works with libtorch1.3, but the latest is libtorch1.4.

Support for libtorch1.3 to libtorch1.4 is being promoted. Please wait. The current changes that we learned from the work It cannot be compiled with visual studio2015.
visual studio2017 also requires updates to the latest patches.

with_bias-> bias transposed seems to be abolished in torch :: nn :: Conv2dOptions. Instead Must be changed to torch :: nn :: ConvTranspose2d.

torch :: nn :: FeatureDropout also seems to be abolished. instead of Use torch :: nn :: Dropout2d.

Other
torch :: Tensor & tensor = torch :: tensor ({vec [i]});
Is said to be a tensor and terminates abnormally.
torch :: Tensor & tensor = torch :: tensor (vec [i]);
Must be.

(/std:c++17)
std::byte compile error C2872
compile option -> /D _HAS_STD_BYTE=0

Linker -> Command line -> Addtitional options
-INCLUDE:?warp_size@cuda@at@@YAHXZ

---