本文主要是介绍装CTC，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！
https://github.com/SeanNaren/warp-ctc
下载下来，传到服务器上
解压
更名
cd warp-ctc
mkdir build; cd build
cmake ..
make上面这些过程缺啥装啥，失败了,把缺的东西装上，重复下面这个过程，直到成功
cd build
make clean
cmake .. make
下一步开始安装
cd pytorch_binding
python setup.py install
报下面这个错：
generating build/warpctc_pytorch/_warp_ctc/__warp_ctc.c
(already up-to-date)
not modified: 'build/warpctc_pytorch/_warp_ctc/__warp_ctc.c'
running install
running bdist_egg
running egg_info
writing warpctc_pytorch.egg-info/PKG-INFO
writing dependency_links to warpctc_pytorch.egg-info/dependency_links.txt
writing top-level names to warpctc_pytorch.egg-info/top_level.txt
reading manifest file 'warpctc_pytorch.egg-info/SOURCES.txt'
writing manifest file 'warpctc_pytorch.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_py
copying warpctc_pytorch/_warp_ctc/__init__.py -> build/lib.linux-x86_64-3.6/warpctc_pytorch/_warp_ctc
running build_ext
building 'warpctc_pytorch._warp_ctc.__warp_ctc' extension
gcc -pthread -B /home1/fzp/anaconda3/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-pib/include -I/home1/fzp/anaconda3/lib/python3.6/site-packages/torch/utils/ffi/../../lib/include/TH -I/home1/fzp/anaconda3/li-I/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/include -I/home1/fzp/anaconda3/include/python3.6m -c build/warpctc_pyto__warp_ctc.o -std=c++11 -fPIC -std=c99 -DWARPCTC_ENABLE_GPU
cc1: warning: command line option ‘-std=c++11’ is valid for C++/ObjC++ but not for C [enabled by default]
gcc -pthread -B /home1/fzp/anaconda3/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-pib/include -I/home1/fzp/anaconda3/lib/python3.6/site-packages/torch/utils/ffi/../../lib/include/TH -I/home1/fzp/anaconda3/li-I/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/include -I/home1/fzp/anaconda3/include/python3.6m -c /home1/fzp/attenti4-3.6/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.o -std=c++11 -fPIC -std=c99 -DWARPCTC_EN
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++ [enabled by default]
cc1plus: warning: command line option ‘-std=c99’ is valid for C/ObjC but not for C++ [enabled by default]
/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.cpp: In function ‘int gpu_ctc(THCudaTensor*,
/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.cpp:92:49: error: cannot convert ‘THCudaTensotTensor*, int)’
int probs_size = THFloatTensor_size(probs, 2);
^
/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.cpp:105:61: error: invalid conversion from ‘s
void* gpu_workspace = THCudaMalloc(state, gpu_size_bytes);
^
/home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.cpp:105:61: error: too few arguments to funct
In file included from /home1/fzp/anaconda3/lib/python3.6/site-packages/torch/utils/ffi/../../lib/include/THC/THC.h:4:0,
from /home1/fzp/attention-OCR/warp-ctc-pytorch_bindings/pytorch_binding/src/binding.cpp:9:
/home1/fzp/anaconda3/lib/python3.6/site-packages/torch/utils/ffi/../../lib/include/THC/THCGeneral.h:209:21: note: declared h
THC_API cudaError_t THCudaMalloc(THCState *state, void **ptr, size_t size);
^
error: command 'gcc' failed with exit status 1
解决方案：
改了之后，执行成功
开始测试是否安装成功：
-
cd /home/xxx/warp-ctc/pytorch_binding/tests
-
python test_gpu.py
报错1：
Traceback (most recent call last):
File "test_gpu.py", line 1, in <module>
import torch
ImportError: No module named torch
装
尽管pytorch到我写下这篇文档时已经更新到0.5.0+版本了，但是SeanNaren/warp-ctc项目目前只兼容pytorch 0.4.0版本。所以安装过程中，请不要使用git clone法安装 pytorch，而是选用conda进行安装。我试过0.5+的版本报错，所以最好还是0.4版本
pytorch 0.4.0版本安装
报错2：
Traceback (most recent call last):
File "test_gpu.py", line 3, in <module>
import pytest
ImportError: No module named pytest
conda install pytest
装了下列两个东西
py: 1.4.34-py27_0
pytest: 3.2.1-py27_0
(crnn) [XXXXX@JXQ-240-26-65 tests]$ python test_cpu.py
===================================================================================================================== test session starts ======================================================================================================================
platform linux2 -- Python 2.7.13, pytest-3.2.1, py-1.4.34, pluggy-0.4.0
rootdir: /export/gpudata/fujingling/projects/warp-ctc/pytorch_binding, inifile: setup.cfg
collected 4 items
test_cpu.py ....
=================================================================================================================== 4 passed in 0.09 seconds ===================================================================================================================
看了其他人的博客：
在build下建一个test.py，加入以下代码，运行并没有报错呢
import torch
from warpctc_pytorch import CTCLoss
ctc_loss = CTCLoss()
# expected shape of seqLength x batchSize x alphabet_size
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()
labels = torch.IntTensor([1, 2])
label_sizes = torch.IntTensor([2])
probs_sizes = torch.IntTensor([2])
probs.requires_grad_(True) # tells autograd to compute gradients for probs
cost = ctc_loss(probs, labels, probs_sizes, label_sizes)
cost.backward()
于是，我就想着可以跑我的CRNN了，
cd crnn.pytorch-master
cp -r /xxx/xxx/XXX/projects/warp-ctc/pytorch_binding/warpctc_pytorch .
然而，万万没想到又报错了，应该是没装成功
python train.py
Traceback (most recent call last):
File "train.py", line 12, in <module>
from warpctc_pytorch import CTCLoss
File "/export/gpudata/fujingling/projects/crnn.pytorch-master/warpctc_pytorch/__init__.py", line 6, in <module>
from ._warp_ctc import *
File "/export/gpudata/fujingling/projects/crnn.pytorch-master/warpctc_pytorch/_warp_ctc/__init__.py", line 3, in <module>
from .__warp_ctc import lib as _lib, ffi as _ffi
ImportError: No module named __warp_ctc
解决方案：
>>> import torch
print torch.__version__
>>> print torch.__version__
0.1.12
估计是这里的问题
conda uninstall pytorch
Fetching package metadata .........................
Solving package specifications: .
Package plan for package removal in environment /export/gpudata/fujingling/conda/envs/crnn:
The following packages will be REMOVED:
pytorch: 0.1.12-py27cuda7.5cudnn5.1_1 http://conda.jdfin.local/conda/free
torchvision: 0.1.8-py27_0 http://conda.jdfin.local/conda/free
Proceed ([y]/n)? y
又报错，cudnn的版本太低
conda install pytorch==0.4
Fetching package metadata .........................
Solving package specifications: .
UnsatisfiableError: The following specifications were found to be in conflict:
- pytorch ==0.4 -> cudnn 7.*
- pytorch-gpu -> cudnn ==5.1
Use "conda info <package>" to see the dependencies for each package.
卸载，继续装
conda uninstall cudnn
Fetching package metadata .........................
Solving package specifications: .
Package plan for package removal in environment /export/gpudata/fujingling/conda/envs/crnn:
The following packages will be REMOVED:
cudnn: 5.1-0 http://conda.jdfin.local/conda/free
pytorch-gpu: 0.1.12-py27_0 http://conda.jdfin.local/conda/free
Proceed ([y]/n)? y
重新安装cudnn之前，先查看以下我的cuda版本，看看硬件能不能支持cudnn7.*
cat /usr/local/cuda/version.txt
CUDA Version 8.0.61
下面是CUDA对应能支持的cudnn列表
7.0.5和7.1.3版本的cudnn比较合适我的机器
(crnn) [xxx@xxxxxx crnn.pytorch-master]$ conda install cudnn==7.1.3
Fetching package metadata .........................
Solving package specifications: .
UnsatisfiableError: The following specifications were found to be in conflict:
- cudnn ==7.1.3 -> cudatoolkit 8.0*
- libtorch-gpu -> cudatoolkit ==7.5
Use "conda info <package>" to see the dependencies for each package.
报错，版本不对应
再卸载再安装：
conda uninstall cudatoolkit
Fetching package metadata .........................
Solving package specifications: .
Package plan for package removal in environment /export/gpudata/fujingling/conda/envs/crnn:
The following packages will be REMOVED:
cudatoolkit: 7.5-2 http://conda.jdfin.local/conda/free
libtorch-gpu: 0.1.12-0 http://conda.jdfin.local/conda/free
nccl: 1.3.4-cuda7.5_1 http://conda.jdfin.local/conda/free
一个一个的装回去
conda install cudatoolkit==8.0
Fetching package metadata .........................
Solving package specifications: .
Package plan for installation in environment /export/gpudata/fujingling/conda/envs/crnn:
The following NEW packages will be INSTALLED:
cudatoolkit: 8.0-3 http://conda.jdfin.local/conda/free
Proceed ([y]/n)? y
装cudnn
conda install cudnn==7.1.3(省略日志)
装pytorch
conda install pytorch==0.4
Fetching package metadata .........................
Solving package specifications: .
Package plan for installation in environment /export/gpudata/fujingling/conda/envs/crnn:
The following NEW packages will be INSTALLED:
intel-openmp: 2019.0-118 http://repos.jd.com/conda/main
libgcc-ng: 8.2.0-hdf63c60_1 http://repos.jd.com/conda/main
libgfortran: 3.0.0-1 http://conda.jdfin.local/conda/free
libstdcxx-ng: 8.2.0-hdf63c60_1 http://repos.jd.com/conda/main
nccl: 1.3.5-cuda9.0_0 http://repos.jd.com/conda/main
ninja: 1.7.2-0 http://conda.jdfin.local/conda/free
openblas: 0.2.14-4 http://conda.jdfin.local/conda/free
pytorch: 0.4.0-py27hdf912b8_0 http://repos.jd.com/conda/main
The following packages will be UPDATED:
cudatoolkit: 8.0-3 http://conda.jdfin.local/conda/free --> 9.0-h13b8566_0 http://repos.jd.com/conda/main
mkl: 2017.0.3-0 http://conda.jdfin.local/conda/free --> 2019.0-118 http://repos.jd.com/conda/main
The following packages will be DOWNGRADED:
cudnn: 7.1.3-cuda8.0_0 http://repos.jd.com/conda/main --> 7.1.2-cuda9.0_0 http://repos.jd.com/conda/main
numpy: 1.13.1-py27_0 http://conda.jdfin.local/conda/free --> 1.10.2-py27_0 http://conda.jdfin.local/conda/free
Proceed ([y]/n)? y
最后经过一番折腾，我发现，把warp-ctc/pytorch_binding/warpctc_pytorch 这个文件夹移动到crnn工程的上一级目录下就好了
这篇关于装CTC的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！
