Skip to content

Latest commit

 

History

History
63 lines (49 loc) · 3.13 KB

README-GPU.md

File metadata and controls

63 lines (49 loc) · 3.13 KB

CaImAn and Tensorflow

Some parts of CaImAn already use Tensorflow, and those parts can benefit from using a hardware-accelerated version of Tensorflow. These versions can be installed using Conda (presently only on the Linux and Windows platforms). Note that this is experimental and our ability to support this environment is limited. To do this:

  • First install the CaImAn environment as normal. Ideally test it.
  • Next, install tensorflow-gpu as follows: conda install tensorflow-gpu
  • Finally, reinstall opencv using the latest openblas build. conda update opencv

The last step should get you a current openblas-based build, although in some cases you will get an mkl build of opencv (which may not have graphical bindings). If this happens, you may be able to search for the most recent opencv with openblas and switch to it using existing conda tooling.

To do this, check after switching package versions that KERAS_BACKEND is set to tensorflow. If not, you can set it in your terminal by typing KERAS_BACKEND=tensorflow. On some platforms it defaults to theano, which will not be hardware-accelerated. This can affect the computational of the CaImAn online algorithm.

CaImAn and CUDA

CaImAn has experimental support for computing FFTs on your GPU, using the CUDA libraries instead of OpenCV or numpy. This can be used during motion correction.

Installation

We assume you have CaImAn generally working first. If you do not, follow the general instructions first and verify you have a working environment.

Next, you will want to install the CUDA libraries. One way to do this is to go to the nVIDIA site here: https://developer.nvidia.com/gpu-accelerated-libraries

Follow the docs there (this can be involved), install the samples, and make sure your videocard is found using the deviceQuery tool. If it does not show up, do not go further; debug your setup until this works.

When cuda is done being set up, enter your conda environment and do the following:

conda install -c lukepfister pycuda
pip install scikit-cuda

The above installs pycuda from what (as of this writing) is the best available distribution of pycuda in a conda repository, and it installs skcuda using pip (as it must be compiled against the CUDA on your system). Better options may become available for both of these as conda-forge evolves.

Use

The CUDA codepaths will only be active if the needed libraries are installed on your system. Otherwise non-CUDA codepaths will be active (even if CUDA is requested in our code).

The following functions have been extended with a use_cuda flag (defaults to false) that if set true will use CUDA for FFT computation (if the needed libraries are involved):

  • The MotionCorrect class - this is most likely the interface you want
  • MotionCorrect::motion_correct_rigid() - used internally by MotionCorrect class
  • register_translation()
  • tile_and_correct()
  • motion_correct_batch_rigid()
  • motion_correct_batch_pwrigid()
  • motion_correction_piecewise()

If you have your environment set up correctly, you can add use_cuda=True to the arguments of MotionCorrect() in your code and it will use your GPU for motion correction.