warp-ctc
warp-ctc copied to clipboard
SeanNaren
支持CUDA11.1吗
我使用 from warp_ctc import CTCLoss 之后，报错：
ImportError: ....... warp-ctc-pytorch_bindings/pytorch_binding/warpctc_pytorch/_warp_ctc.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN3c1011CPUTensorIdEv
我之前是能用的。然后我用nvidia-smi查看cuda版本为11.1 ，请问是否不支持cuda11.1？期待解答，不胜感激！
I add some lines in CMakeLists for RTX3090+CUDA11.1+pytorch1.7.1.
https://github.com/SeanNaren/warp-ctc/blob/e2609d83b4de81c7c9c4804de490d1084655d3d9/CMakeLists.txt#L50-L57
IF (CUDA_VERSION GREATER 8.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70")
ENDIF()
IF (CUDA_VERSION GREATER 9.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75")
ENDIF()
IF (CUDA_VERSION GREATER 10.9)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80")
ENDIF()
if (NOT APPLE)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --std=c++14")
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}")
ENDIF()
It works for the following testing codes,
import torch
import warpctc_pytorch as warp_ctc
def test_simple():
probs = torch.FloatTensor([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]]]).transpose(0, 1).contiguous()
grads = torch.zeros(probs.size())
labels = torch.IntTensor([1, 2])
label_sizes = torch.IntTensor([2])
sizes = torch.IntTensor(probs.size(1)).fill_(probs.size(0))
minibatch_size = probs.size(1)
costs = torch.zeros(minibatch_size)
warp_ctc.cpu_ctc(probs,
grads,
labels,
label_sizes,
sizes,
minibatch_size,
costs,
0)
print('CPU_cost: %f' % costs.sum())
probs = probs.clone().cuda()
grads = torch.zeros(probs.size()).cuda()
costs = torch.zeros(minibatch_size)
warp_ctc.gpu_ctc(probs,
grads,
labels,
label_sizes,
sizes,
minibatch_size,
costs,
0)
print('GPU_cost: %f' % costs.sum())
print(grads.view(grads.size(0) * grads.size(1), grads.size(2)))
if __name__ == '__main__':
test_simple()
Output:
CPU_cost: 2.462858
GPU_cost: 2.462858
tensor([[ 0.1770, -0.7081, 0.1770, 0.1770, 0.1770],
[ 0.1770, 0.1770, -0.7081, 0.1770, 0.1770]], device='cuda:0')
I add some lines in CMakeLists for RTX3090+CUDA11.1+pytorch1.7.1.
https://github.com/SeanNaren/warp-ctc/blob/e2609d83b4de81c7c9c4804de490d1084655d3d9/CMakeLists.txt#L50-L57
IF (CUDA_VERSION GREATER 8.9) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70") ENDIF() IF (CUDA_VERSION GREATER 9.9) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75") ENDIF() IF (CUDA_VERSION GREATER 10.9) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80") ENDIF() if (NOT APPLE) set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} --std=c++14") set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS}") ENDIF()It works for the following testing codes,
import torch import warpctc_pytorch as warp_ctc def test_simple(): probs = torch.FloatTensor([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]]]).transpose(0, 1).contiguous() grads = torch.zeros(probs.size()) labels = torch.IntTensor([1, 2]) label_sizes = torch.IntTensor([2]) sizes = torch.IntTensor(probs.size(1)).fill_(probs.size(0)) minibatch_size = probs.size(1) costs = torch.zeros(minibatch_size) warp_ctc.cpu_ctc(probs, grads, labels, label_sizes, sizes, minibatch_size, costs, 0) print('CPU_cost: %f' % costs.sum()) probs = probs.clone().cuda() grads = torch.zeros(probs.size()).cuda() costs = torch.zeros(minibatch_size) warp_ctc.gpu_ctc(probs, grads, labels, label_sizes, sizes, minibatch_size, costs, 0) print('GPU_cost: %f' % costs.sum()) print(grads.view(grads.size(0) * grads.size(1), grads.size(2))) if __name__ == '__main__': test_simple()Output:
CPU_cost: 2.462858 GPU_cost: 2.462858 tensor([[ 0.1770, -0.7081, 0.1770, 0.1770, 0.1770], [ 0.1770, 0.1770, -0.7081, 0.1770, 0.1770]], device='cuda:0')
(yu_tf_py) root@gao-PowerEdge-T640:/yu_ex/warp-ctc3333/pytorch_binding# python setup.py install Traceback (most recent call last): File "setup.py", line 6, in from torch.utils.cpp_extension import BuildExtension, CppExtension ModuleNotFoundError: No module named 'torch.utils' (yu_tf_py) root@gao-PowerEdge-T640:/yu_ex/warp-ctc3333/pytorch_binding# pip install torch.utils Requirement already satisfied: torch.utils in /root/ana3/envs/yu_tf_py/lib/python3.6/site-packages (0.1.2) Requirement already satisfied: torch in /root/ana3/envs/yu_tf_py/lib/python3.6/site-packages (from torch.utils) (1.7.1+cu110) Requirement already satisfied: numpy in /root/ana3/envs/yu_tf_py/lib/python3.6/site-packages (from torch->torch.utils) (1.19.2) Requirement already satisfied: dataclasses in /root/ana3/envs/yu_tf_py/lib/python3.6/site-packages (from torch->torch.utils) (0.8) Requirement already satisfied: typing-extensions in /root/ana3/envs/yu_tf_py/lib/python3.6/site-packages (from torch->torch.utils) (3.7.4.3) I encountered such a problem when running install. Would you like to ask me if you have encountered any? cuda11.1 torch 1.7.1
