v0.18.3

Bug fix:

• as v0.18.2, but fixes a regression in reduction

v0.18.2

Changes:

• tf.split enabled
• following examples from Aymeric Damien's Tensorflow-Examples run now:
  - recurrent_network.py
  - bidirectional_rnn.py
  - dynamic_rnn.py

v0.17.3

Bug fix release:

• tf.random_uniform and tf.random_normal should give equal results to the cpu version, on both Mac and Ubuntu
• tf.random_normal should no longer give all zeros results on Ubuntu, ie should fix #35
• the Mac wheel should have RPATH set correctly, ie hopefully should not give error messages about unable to load libclew.dylib or similar, ie should fix #39

v0.17.2

Changes:

• tf.random_normal and tf.random_uniform work now (Nuance: tf.random_normal works on Radeon, but fails on NVIDIA, currently)
• a few other operations enabled, such as as slicing, aggregation, concat, gather enabled
• Adam works now (needed random implemented in order to work)
• softmax works
• runs on both Ubuntu 16.04 (tested on NVIDIA) and on Mac Sierra (tested on Radeon Pro 450)
• multilayer_perceptron.py runs ok now :-)

v0.16.0

Mac build running similarly to the v0.14.0 ubuntu build now. Using Mac Sierra and Radeon Pro 450 (selected using export CL_GPUOFFSET=1):

test	Mac Sierra, using Radeon Pro 450 GPU
unit tests (py.test -v)	pass
linear_regression.py	slow, but works
logistic_regression.py	ok
nearest_neighbor.py	ok (accuracy 0.92)
multilayer_perceptron.py	missing adam
recurrent_network.py	missing adam
autoencoder.py	missing rmsprop

To use the wheel below (Mac only):

• download the tar
• tar -xf the tar
• pip install the untarred wheel

v0.15.0

• runs on Mac :-) . Tested on Mac Sierra, using Radeon Pro 450 GPU (courtesy of my employer ASAPP
• wheel for Mac Sierra below, (built using python 3.6)
• should/might work on other Mac/python versions, though you'll need to build-from-source
• No Ubuntu 16.04 for now, though the older v0.14.0 wheel has pretty much identical functionality from Ubuntu point of view
• screenshot of running tests on Mac:

Note that to test this on Radeon Pro 450 ,since there are two gpus, the index 0 one being the Intel HD530, which fails for now, I ran the tests like this:

CL_GPUOFFSET=1 py.test -svx

... this then chooses to run the tests on gpu index 1, ie the Radeon

important: note that the wheel attached is ONLY for Mac, Sierra, python 3.6. It wont work on anything else.

v0.14.0

• radical overhaul under the covers
• argmin / argmax work now
• softmax works now

v0.13.0

• beignet test results fairly solidly match K520 results now
• fixed the regression on not_equal operator
• removed the spam from memory copy

v0.12.1

FINALLY Fix critical bug on beignet. Unary and binary operators now work on beignet :-) (Tested on HD5500)

v0.11.0

• fixes critical bug in v0.10.0 release, where the number of devices was hard-coded to be 0 :-P
• Aymeric Damien's 2_BasicModels all run now, on NVIDIA K520. Seem broken on Intel HD5500 for now
• bunch of fixes underneath to get 2_BasicModels working ok on K520