f56de38bba
Reduce test code in tutorials and fixes
2017-07-07 13:44:58 -07:00
5ecc834722
updating links to old wiki - referencing now the doc site
2017-06-07 15:55:34 +02:00
d6b55e28af
Bump version numbers
2017-06-01 15:12:14 +02:00
8dc56e2d7c
Bump version numbers
2017-05-24 14:01:23 +02:00
9b48ea7648
Updated notebooks and added missing tests
2017-05-12 10:28:14 -07:00
e56535ff2a
input -> input_variable complete
2017-05-11 17:00:07 -07:00
95c755ec04
Fixed py2.7 bug
2017-04-26 19:50:33 -07:00
244aa602bd
Addressed CR comments for notebook update
2017-04-26 17:01:58 -07:00
d08da86eca
Updating the CNTK_202 notebook and test
2017-04-26 11:14:46 -07:00
a1e10ff9e7
Bump version numbers
2017-04-21 15:58:00 +02:00
861bfbd714
Bump versions
2017-04-03 18:16:31 +02:00
8a9020f78e
CNTK v2 library: Migrate past_value and future_value to sequence
2017-03-30 12:11:20 -07:00
d251977f38
Merge with master.
2017-03-29 15:05:51 -07:00
9adade3311
Adapting examples; set MinibatchData.data to be a property
2017-03-29 17:09:13 +02:00
90287b6db7
Fix Notebook...
2017-03-29 02:07:15 -07:00
503983b0d1
Add an ability to limit the number of sweeps for training
...
* Add MinibatchSourceConfig struct that makes construction of a reader
a little more straight-forward (esp. on the c++ side)
* Refactor python implementation of MinibatchSource.
* Add support for specifying the maximum number of sweeps over the input
dataset when instating a minibatch source.
2017-03-28 13:50:17 +02:00
8da12bcd49
Merge with master.
2017-03-17 10:36:40 -07:00
124f07a0cc
Another attempt at streamlining device selection
...
Update device selection API as agreed:
* drop BestDevice() and DefaultDevice()
* add an option to specify excluded devices that should not be used as
globally default (neither auto-selected nor explicitly specified)
* add a mechanism to query device properties
* add a method to determine if a device is locked or not (in real-time)
* in SetDefaultDevice, offer an option to exclusively lock the specified
device.
2017-03-17 12:34:54 +01:00
e25dd44c70
Merge with master.
2017-03-15 22:10:17 -07:00
d86c231d1f
Move all examples and notebook to use the new namespace.
2017-03-15 22:07:51 -07:00
97e674c934
Bump versions
2017-03-15 17:16:54 +01:00
1bdd3da4e4
reverted depth_first_search distance/depth in stack variable to prior;
...
deleted some unused left-over code;
fixed some import/namespace issues with new tests
2017-03-03 14:02:18 -08:00
6256c6e355
Updated examples to use the latest progress writer API
2017-03-02 17:57:57 -08:00
dc0dde69ca
Bump versions
...
Conflict resolved for:
bindings/python/doc/tutorials.rst
2017-02-23 17:10:19 +01:00
ad2dfb7180
Merge branch 'master' into nluehr/examplefixes
2017-02-15 14:29:27 -08:00
2f7a2f04c8
Cosmetic tweak in tutorials.
...
Make progress printer show correct epochs limit in most cases.
2017-02-10 13:19:39 -08:00
a435f5a159
Bump versions
2017-02-10 15:57:35 +01:00
adf0a9238a
changed signature of splice() from list to *args, to make it consistent with all other ops;
...
renamed cntk.ops.reduce_log_sum() to reduce_log_add_exp(), to make it consistent with log_add_exp();
associative binary ops like plus() now accept >2 arguments;
new ops element_max(), element_min();
softmax() now takes axis parameter;
constant() now takes dtype parameter argument
2017-02-09 17:06:09 -08:00
19d2e5f7ea
changed all imports of blocks and models to layers
2017-02-09 15:41:12 -08:00
99c25c0327
changed Trainer() signature from Trainer(model, loss, metric, ...) to Trainer(model, (loss, metric), ...) to prep for Layers lib where (loss, metric) is the criterion function;
...
fixed an existing bug in a doc example
2017-02-09 15:22:45 -08:00
6a603b0c79
Update images
2017-02-02 17:40:20 -08:00
d4b49b8208
Bump versions
2017-02-01 18:13:00 +01:00
0451daa7c3
Bump versions
2017-01-20 17:26:07 +01:00
9b6c6bde34
Add a default (true) for the unit-gain flag value
2017-01-20 09:27:36 +01:00
3dd737f500
Bump version
2017-01-16 20:15:06 +01:00
3ab246855b
Add unit-gain momentum flag
...
This commit contains a breaking change: it adds a flag to both
C++ and python Learner API that indicates if the momentum should be applied in
the regular fashion or as a unit-gain filter. This flag is a required
parameter for learners that use momentum (momentum_sgd, nesterov_sgd and
adam_sgd).
2017-01-16 09:39:39 +01:00
0fc7cdc3ae
fix an issue on CNTK 202 introduced with new handling of sequence inputs
2017-01-12 18:15:30 -08:00
4665583810
Make Jupyter Notebook tests device-aware
2017-01-12 13:59:56 +01:00
6c73818bf7
[CNTK Core] Change BatchNormalization default to CuDnn in both BrainScript and V2 C++/Python, and add fallback to cntk engine when not supported by cudnn
...
Fix epsilon in CuDnnBatchNormalization to be consistent in forward/backward
Fix BatchNormalization parameter gradient accumulation
Add message on CuDnnBatchNormalization forward infer batch size limitation, and adjust CNTK_202_Language_Understanding eval minibatch size to avoid CUDA failure
2017-01-11 01:00:09 -08:00
fe24966f7e
CNTK v2 library: Change BlockFunction implementation to store the mapping of block function arguments and outputs in the variables themselves instead of maps on teh side
2017-01-05 10:38:57 -08:00
b83484401a
CNTK v2 library: BlockFunction feature to create heirarchical function graphs
2016-12-24 00:28:22 -08:00
89ffdf4ee4
Bump version
2016-12-22 10:52:53 +01:00
8f4cad8f40
Enable Python 2.7
2016-12-19 13:49:05 +01:00
80f70b638d
Bump version
2016-12-12 18:59:15 +01:00
f7d0150221
Add test for CNTK_202_Language_Understanding.ipynb
2016-12-09 13:20:12 +01:00
e2b7fee2d9
Incorporated CR feedback
2016-12-01 11:24:52 -08:00
e6c96f8dd0
Added couple of clarifying comments without any test
2016-12-01 11:24:51 -08:00
b052cece03
Tutorials: more fixes
2016-11-18 19:15:15 +01:00
e442dd4273
Tutorials: more fixes
2016-11-18 14:47:17 +01:00
e8c9865977
Fix some typos, tune
2016-11-17 20:39:10 +01:00
REDMOND\sayanpa
Wolfgang Manousek
Mark Hillebrand
Nikos Karampatziakis
Amit Agarwal
Emad Barsoum
Willi Richert
Alexey Reznichenko
Frank Seide
NORTHAMERICA\vistepan
Chris Basoglu
Nathan Luehr
KeDengMS