Closed vadimkantorov closed 5 years ago
CMakeLists.txt needs to be modified. Also suggested in https://github.com/SeanNaren/warp-ctc/issues/116#issuecomment-475145386 .
Otherwise test_gpu does not pass and CTCLoss returns 0.0.
test_gpu
IF (CUDA_VERSION GREATER 7.6) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_60,code=sm_60") set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_61,code=sm_61") set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_62,code=sm_62") set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70") ## new line ENDIF()
you can use CTC Loss in pytorch 1.0.1 directly without warp_ctc
warp-ctc does support this compute architecture already, the change was made here
CMakeLists.txt needs to be modified. Also suggested in https://github.com/SeanNaren/warp-ctc/issues/116#issuecomment-475145386 .
Otherwise
test_gpu
does not pass and CTCLoss returns 0.0.