2

在我尝试导入 cuml 库之前，我已经在 Colab 中安装了 RAPIDS，没有任何问题。幸运的是，我有 Tesla 4 作为 GPU。

这就是我安装 RAPIDS 的方式

# clone RAPIDS AI rapidsai-csp-utils scripts repo
>> !git clone https://github.com/rapidsai/rapidsai-csp-utils.git    
# install RAPIDS
>> !bash rapidsai-csp-utils/colab/rapids-colab.sh
>> import sys, os
# set necessary environment variables 
>> dist_package_index = sys.path.index('/usr/local/lib/python3.6/dist-packages')
>> sys.path = sys.path[:dist_package_index] + ['/usr/local/lib/python3.6/site-packages']+sys.path[dist_package_index:]
>> sys.path    
# update pyarrow & modules 
>> exec(open('rapidsai-csp-utils/colab/update_modules.py').read(), globals())
Enjoy using RAPIDS!
RAPIDS Version to install is 0.11
Checking for GPU type:
***********************************************************************
Woo! Your instance has the right kind of GPU, a 'Tesla T4'!
***********************************************************************

************************************************
Your Google Colab instance has RAPIDS installed!
************************************************
***********************************************************************
Let us check on those pyarrow and cffi versions...
***********************************************************************
You're running pyarrow 0.15.0 and are good to go!
unloaded cffi 1.11.5
loaded cffi 1.11.5

当我尝试导入时：

>> import cuml

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-31-a450aff8eac6> in <module>()
----> 1 import cuml

5 frames
/usr/local/lib/python3.6/site-packages/cudf/core/dataframe.py in <module>()
     22 
     23 import cudf
---> 24 import cudf._lib as libcudf
     25 import cudf._libxx as libcudfxx
     26 from cudf._libxx.null_mask import MaskState, create_null_mask

AttributeError: module 'cudf' has no attribute '_lib'

我也放了这个：

>> ! conda list | grep cudf
>> ! conda list | grep cuml
>> ! conda list | grep cugraph
>> ! conda list | grep numpy
>> ! conda list | grep pandas

cudf                      0.13.0                   py36_0    rapidsai/label/main
dask-cudf                 0.13.0                   py36_0    rapidsai/label/main
libcudf                   0.13.0               cuda10.0_0    rapidsai/label/main
cuml                      0.13.0          cuda10.0_py36_0    rapidsai/label/main
libcuml                   0.13.0               cuda10.0_0    rapidsai/label/main
libcumlprims              0.13.0               cuda10.0_0    nvidia
cugraph                   0.13.0                   py36_0    rapidsai/label/main
libcugraph                0.13.0               cuda10.0_0    rapidsai/label/main
numpy                     1.17.5           py36h95a1406_0    conda-forge
pandas                    0.25.3           py36hb3f55d8_0    conda-forge

这是我的问题，我正在尝试使用 RAPIDS 使用他的 T-SNE，它比 Scipy T-SNE 更快。

4

1 回答 1

2

谢谢你分享这个。此问题是由于对 numba (0.48.0 -- 0.49.0) 的更新导致它与 cudf 不兼容所致。这个 PR https://github.com/rapidsai/rapidsai-csp-utils/pull/18解决了这个问题，它将 numba 锁定为 0.13 及以下的 0.48.0。

于 2020-06-09T19:44:51.517 回答