cuDNN 路径未解析

Question

我正在尝试运行 this theano test，但我的 cudnn 路径有问题。我记得安装 cudnn 后通过了测试，但是 theano 似乎找不到 cudnn 的路径。这是环境变量问题吗？有人有解决方案吗？

whereis cuda

给我

cuda: /usr/include/cuda.h /usr/local/cuda

运行 THEANO_FLAGS=device=cuda0 MKL_THREADING_LAYER=GNU python test.py 的测试给了我

ERROR (theano.gpuarray): Could not initialize pygpu, support disabled
Traceback (most recent call last):
  File "/home/lucas/miniconda2/lib/python2.7/site-packages/theano/gpuarray/__init__.py", line 220, in <module>
    use(config.device)
  File "/home/lucas/miniconda2/lib/python2.7/site-packages/theano/gpuarray/__init__.py", line 207, in use
    init_dev(device, preallocate=preallocate)
  File "/home/lucas/miniconda2/lib/python2.7/site-packages/theano/gpuarray/__init__.py", line 110, in init_dev
    context.cudnn_handle = dnn._make_handle(context)
  File "/home/lucas/miniconda2/lib/python2.7/site-packages/theano/gpuarray/dnn.py", line 124, in _make_handle
    cudnn = _dnn_lib()
  File "/home/lucas/miniconda2/lib/python2.7/site-packages/theano/gpuarray/dnn.py", line 111, in _dnn_lib
    config.dnn.base_path)
RuntimeError: Could not load cudnn library. Check your cudnn installation. Maybe using the Theano flag dnn.base_path can help you. Current value "/usr"
[Elemwise{exp,no_inplace}(<TensorType(float64, vector)>)]
Looping 1000 times took 3.294467 seconds
Result is [1.23178032 1.61879341 1.52278065 ... 2.20771815 2.29967753 1.62323285]
Used the cpu

Answer 1

由于出现错误，我不得不在 THEANO FLAGS 中包含 cudnn 路径。

这可以通过两种方式完成：

(1) 内联：THEANO_FLAGS="device=cuda0, dnn.base_path=/usr/local/cuda" MKL_THREADING_LAYER=GNU python test.py

或

(2) 在 theano 配置文件 ~/.theanorc 中包含以下行：

[dnn]
include_path = /usr/lib/cuda/include
library_path = /usr/lib/cuda/lib64

我希望这对其他人有帮助。

cuDNN 路径未解析

cuDNN path not resolving

python

theano

cudnn

theano-cuda