Proposition: After the OpenCL branch maturing for over 3 years, maybe it's time to give Caffe a bump#6427
Conversation
Mention SKX support
The logic for setting the library RPATH checks whether or not
${CMAKE_INSTALL_PREFIX}/lib is a system directory, and if not
adds it to the library RPATH. However, caffe does not install
to ${CMAKE_INSTALL_PREFIX}/lib, it installs to
${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR} (from
GNUInstallDirs). CMAKE_INSTALL_LIBDIR may be something like
"lib/x86_64-linux-gnu"
Fix division operator for Compatibility of python 3 in classifier.py
The pointers could be used by CUDA wrapper libraries in Python such as PyCUDA, gnumpy, Theano etc.
This line is needed for Ubuntu 16.04:
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
For reference:
* https://github.com/BVLC/caffe/wiki/Ubuntu-16.04-or-15.10-Installation-Guide
* https://youtu.be/DnIs4DRjNL4
Expose GPU pointers to Python
fix bilinear filler (and make constant filler more strict, as it should be)
Signed-off-by: Finnian Anderson <get@finnian.io>
…op_k, no need for fancy priority_queue etc. (2) GPU implementation
…ch-1 [docs] packages needed by Ubuntu 16.04, not just Ubuntu 14.04
Add absolute tolerance to test_net.py to prevent random Travis fails
fixes "NameError: name 'int_tp' is not defined" when running plot_training_log.py
fixes "NameError: name 'int_tp' is not defined" when running plot_training_log.py
Update link to google style guide.
Fix default mode warning in io.resize_image
Fixed bug where make distribute duplicates python files
fix web demo install instruction link
This commit is to fix issue #5718. * reference: 1. https://groups.google.com/forum/#!topic/caffe-users/nBpWJCcJoCU 2. https://stackoverflow.com/questions/28692209/using-gpu-despite-setting-cpu-only-yielding-unexpected-keyword-argument Signed-off-by: Geunsik Lim <geunsik.lim@samsung.com>
Fix: mean shape incompatible with input shape
…default parameters for VC4C and Mali.
Opencl: Remove 2d input check
|
@naibaf7 Thank you for dedicating so much effort and engineering to OpenCL Caffe! This branch has only improved since the old times of #2610. Throughout the discussion and comments you have always contributed code, and that takes a lot of time and thought. I have clearly not had the bandwidth for Caffe development that I once did. I still have my sights on a 1.1 to include module layers (and by extension Halide layers), because these are excellent directions that I was always interested in—among many other good directions, of course—but it has been a long while now. While I still hope to rally the time to take care of last bits of grooming and merge those, I do not expect to be able to contribute as necessary to make a difference in the ~100k line patch that you are hammering on here. That said, do not let that be an obstacle to your hacking! The only constraint I would advise is that existing capabilities not go missing, like multi-GPU training, HDF5 IO, and so forth. Whether this line of development is a branch, fork, or merged in master can be up to you and the others who keep pushing, pulling, reviewing, and of course brewing ☕ |
|
"Thanks for all the hard work" - Lets it rot as a dangling PR forever. |
|
@naibaf7 : what happened? whats the ultimate decision ? |
|
I have to say that right now, there's compilers issues. Msvc rejects the function instantiation macros being used in some layers. |
|
@TechnikEmpire |
|
Thanks for your contribution @naibaf7. It's unfortunate but Nvidia has a strangle hold on deep learning, and now Intel has joined the ranks, even destroying opencv and making it depend on proprietary acceleration that doesn't even work on amd cpus. People like you with your contributions are opening these technologies to the general public while others hold back progress to line their corporate pockets. Will watch your personal repos. Thanks again. |
|
@naibaf7 This was indeed a welcome commit. Can you please check the PR That I made? If a few issues related to "double" is solved then the code will be fully functional. However, there is a HSA version of caffe being developed at ROCm! I am waiting for that. The older version of the code is also there in my personal repo and it should work without any hitch. |
|
Thanks for your support. Stay tuned for future developments in deep learning by me, it just may take a while longer. :) |
This comment was marked as off-topic.
This comment was marked as off-topic.
20 participants
The OpenCL branch has been maturing for 3 years. In this time, a lot of features have been accumulated that are absent in the main branch. Notably, there are many features (and many more to come) that would differentiate Caffe from other frameworks such as TensorFlow, Caffe2 and Torch in a positive way. Caffe has been lacking such features lately.
I'm also standing on a crossroad now, where adding more features to my branch would ultimately make it too difficult to merge upstream changes from the master branch into the opencl branch.
A selection of core features:
Upcoming features:
Currently broken features:
Missing/incomplete features:
So I hope we can get some maintainers to help me hunt down the remaining few bugs in this branch and make it ready for a merge with the master branch (maybe until December?).
I hope this finds consideration among other Caffe maintainers. I will follow up with all changes and additions made to this branch once core maintainers will agree to go down this path for the future of Caffe.