Theano is a library that CUDA can easily use in Python, but it seems that the new information is neither Japanese nor English, so I will share what I confirmed yesterday.
-Theano 0.6rc2 is usually introduced as the latest version, but it only supports up to 0.5 in the Windows environment. When using 0.6rc2 on Windows, it works in CPU mode, but when it is in mode that uses the essential GPU, nvcc (actually VC ++) called from Theano as a subcontractor gives an error and cannot compile (Theano's It is written in ML that the CPU mode has been confirmed to work, but it seems that the GPU mode is not written as bad ...).
-If you run Theano 0.5 in a Windows environment, you can run it in GPU mode even with CUDA v5.0. Both Liner Regression and Multilayer Perceptron in the deeplearning.net tutorial work as they are, and the code in the 0.6rc2 tutorial works fine with a few modifications (comment out of the automatic judgment function of CPU mode or GPU mode, etc.). , 0.5 and 0.6rc2 test is unconfirmed to what extent it passes.
Recommended Posts