v8.3.1

Compared to v8.1.3:

• added ScaledTanh on python
• added ELU
• removed some build warnings on Mac
• various fixes and tweaks (see https://github.com/hughperkins/DeepCL/commits/v8.3.1 )

Native library downloads for linux and Windows:

http://deepcl.hughperkins.com/Downloads/
usage instructions in the readme, https://github.com/hughperkins/DeepCL

Python packages on pypi: https://pypi.python.org/pypi/DeepCL/8.3.1

v8.1.3

Code changes:

• removed the norbloader unit test, which systematically fails (since requires norb dataset to be downloaded)

Release changes:

• Edward Geist has very kindly created a Mac OS X Homebrew formula for this release, DeepCL v8.1.3, which you can find at: https://gist.github.com/GOFAI/53a25ff22f31ec144608

Native library downloads for linux and Windows:

• http://deepcl.hughperkins.com/Downloads/
• usage instructions in the readme, https://github.com/hughperkins/DeepCL

v8.1.2

Bug fixes mostly (compared to v8.0.0):

• use int64_t, to ensure 64-bit ints are 64-bit on all platforms
• add guards to some kernels to prevent crashing on certain geometries
• python build works on OS X now
• fixed some OpenCL warnings, and some failing tests
• python build works now when opencl headers not present

Also, one minor feature-ette:

• in python, can call 'getLabels' on a SoftMax layer

Native library downloads (for linux and Windows; Mac you need to build your own for now...):

• http://deepcl.hughperkins.com/Downloads/
• usage instructions in the readme, https://github.com/hughperkins/DeepCL

v8.0.0

(these notes are kind of a bit retroactive, but better late than never :-) I've just copied them from the 8.0.0rc9 notes actually...)

Native library downloads:

• http://deepcl.hughperkins.com/Downloads/
• usage instructions in the readme, https://github.com/hughperkins/DeepCL

Python wrappers:

• first activate the native libraries, as per the native library readme above
• now do pip install --pre DeepCL
• do python -c "import PyDeepCL; cl = PyDeepCL.DeepCL()" to check all is working

Changes since 5.x:

• added clblas
• added im2col
• python wrappers now a thin layer around the standard native libraries, rather than having separate native library build process for python
• lua wrappers removed (please use https://github.com/hughperkins/cltorch and https://github.com/hughperkins/clnn if you want to train convolutional neural networks using lua)
• installation process has changed, so please follow the instructions in the readme
• the commandline utilities are now called deepcl_train and deepcl_predict, in place of deepclrun, deepclexec, etc

(Note that 6.x only had a single release, then the clblas / build overhaul was immediately started, producing 8.x. I skipped 7.x by accident :-P )

v8.0.0rc9

Probalby the final rc release for 8.0.0, before the first non-rc release.

Native library downloads:

• http://deepcl.hughperkins.com/Downloads/
• usage instructions in the readme, https://github.com/hughperkins/DeepCL

Python wrappers:

• first activate the native libraries, as per the native library readme above
• now do pip install --pre DeepCL
• do python -c "import PyDeepCL; cl = PyDeepCL.DeepCL()" to check all is working

Changes since 5.x:

• added clblas
• added im2col
• python wrappers now a thin layer around the standard native libraries, rather than having separate native library build process for python
• lua wrappers removed (please use https://github.com/hughperkins/cltorch and https://github.com/hughperkins/clnn if you want to train convolutional neural networks using lua)
• installation process has changed, so please follow the instructions in the readme
• the commandline utilities are now called deepcl_train and deepcl_predict, in place of deepclrun, deepclexec, etc

(Note that 6.x only had a single release, then the clblas / build overhaul was immediately started, producing 8.x. I skipped 7.x by accident :-P )

v6.0.0

This release doesnt so much add any new features, as deprecate a couple of ones, and hopefully fix a build issue in 5.8 on Python.

Changes this release:

• lua wrappers removed. lua is now supported via cltorch and clnn, which are still under development, but fairly far advanced
• deepclrun and train commands are both deprecated in favor of deepcl_train
• predict becomes deepcl_predict
• under the covers, easycl has been upgraded to use lua kernel templates

Please download the binaries from http://deepcl.hughperkins.com/Downloads/ For python, you can install via pypi

v5.10.2

New:

• new commandline option to predict, outputlayer=, which selects which layer you want to record the output from (ie, doesnt have to be the last layer)

Bug fixes:

• each example in output in text format of predict is now on a separate line, rather than all examples from each batch on one single line
• in predict, every other input example is no longer skipped
• gpuinfo works on hd5500 integrated graphics now, without causing display corruption

Commandline tools and c++ library: http://deepcl.hughperkins.com/Downloads
Python wrapper: https://pypi.python.org/pypi/DeepCL/5.10.2

v5.9.0

New:

• deepclrun becomes train: handles training, and validation, using labelled data
• deepclexec becomes predict: handles creating predictions from unlabelled data
• predict can take input from a file, using GenericLoader, same formats as training, or from stdin
• predict can output to a file, or to stdout, in text or binary format

Please download commandline version and library from http://deepcl.hughperkins.com/Downloads
Python wrappers available on pypi at https://pypi.python.org/pypi/DeepCL/5.9.0 and https://pypi.python.org/pypi/DeepCL/v5.9.0 (apparently I need to fix the version so linux and Windows go to the same pypi version)

v5.8.3

Bug fixes:

• clean up some memory leaks in unit tests

Download from http://deepcl.hughperkins.com/Downloads/

v5.8.2

Bug fixes:

• remove memory leaks in various clMath objects, associated with not deleting kernels