Jan Schlüter

Perfect, thanks. Can you squash this into a single commit? ``` git reset --soft 41ed700 git commit --amend --no-edit git push --force ```

> I downloaded the cudamat-master folder from git into my anaconda/lib/site-packages folder I think you shouldn't do that, `setup.py` will try to install it into that directory on its own....

> UPDATE: issue resolved. This is crazy. Thank you for taking the time to report the solution! We'd need #53 to solve this seamlessly. We're currently using `setuptools` to find...

I've attempted to simplify cudamat's `setup.py` in #79 -- if you have the time, please try that and report back! Thank you!

> How to verify the installation of cudamat master using pip command.... If by "verify" you mean running the tests, you will need to [download the test script](https://github.com/cudamat/cudamat/raw/master/test/test_cudamat.py) and run...

Just compiling everything for all architectures might work as well, of course. I guess the only way to detect the compute capability at configure time is by compiling and running...

> I'm using NVIDIA Geforce 940M Maxwell architevture, cuda 7.5, nvidia-352.39. How I get the right architecture? As per [a table maintained on Wikipedia](https://en.wikipedia.org/wiki/CUDA#Supported_GPUs), a 940M has compute capability 5.0....

> I guess the only way to detect the compute capability at configure time is by compiling and running a small program calling cudaGetDeviceProperties() for every device. My past self...

I think nvblas may not be what you want. It assumes your data is in host memory and estimates whether copying it to device memory, executing the BLAS call and...

> I think the training algorithm can be modified to use two cards [...] There are several possibilities. Some relevant papers: http://arxiv.org/abs/1312.5853 http://arxiv.org/abs/1404.5997 Implementing this would require some extensions to...