History log of /external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
Revision Date Author Comments (<<< Hide modified files) (Show modified files >>>)
f5b30312013df5b7bd3a50555b2facadd4aed204 16-Feb-2018 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: Support for embedding layers, add FisherFactor.{multiply, multiply_inverse}.

PiperOrigin-RevId: 185920837
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
ca3ac2a464b92f4c0498dfde875f99102a0d410c 27-Jan-2018 A. Unique TensorFlower <gardener@tensorflow.org> Fixed bug: inconsistency with how damping normalization was applied to ConvDiagonalFB blocks.

PiperOrigin-RevId: 183472440
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
3252d0d41c15c1b26376c9c86c537aa275a1bb65 17-Jan-2018 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: Expose protected functions from fisher_blocks and fisher_factors and constant strings from layer_collection in their respective library modules. This allows consistent development of blocks and factors outside tensorflow.contrib.kfac.

PiperOrigin-RevId: 182197356
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
2139d204f038ed944cecf05873a565f7f7d424b2 08-Dec-2017 A. Unique TensorFlower <gardener@tensorflow.org> Finishing the migration of the RNN / shared fully-connected layer blocks, and associated changes to LayerCollection.

Moved check_registration function to LayerCollection which is a more natural home for it.

Fixed major bug: make_inverse_update_ops was misspelled in FullyConnectedMultiKF meaning that inverses weren't being computed for these factors.

Many other minor tweaks and fixes to code and documentation.

PiperOrigin-RevId: 178372885
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
48ce41e160c2dcd8fde82db5b4a2a44a5f36894f 01-Dec-2017 A. Unique TensorFlower <gardener@tensorflow.org> Moved compute_pi from utils.py to fisher_blocks.py and added global constant that controls the type of pi-based damping adjustment to use. Current options are "tracenorm", which is what we were already doing and is the default, and "off".

PiperOrigin-RevId: 177610677
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
370e521762f3cbd558a7e56992e3b062236b626f 01-Dec-2017 A. Unique TensorFlower <gardener@tensorflow.org> Adds a fisher block for fully connected recurrent layers.

`FullyConnectedSeriesFB` uses an approximation to the Fisher information
matrix designed for RNNs. This CL also adds support for dtypes
other than `float32` to `fisher_factors`.

PiperOrigin-RevId: 177558080
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
bd9d8064c8342a59db03301b4afd10238002e344 13-Nov-2017 A. Unique TensorFlower <gardener@tensorflow.org> Small change to how damping constants get transformed. Switched from using tf.square to **0.5. This will preserve them as floats if they are already, so the dictionaries work properly. See b/62785471

PiperOrigin-RevId: 175526788
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
453dd5848f5652f520eb0faf17a732f20779cdb1 31-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: Support for tf.AUTO_REUSE when re-using registrations. Multi-tower support for FullFB, NaiveDiagonalFB. Removal of LayerCollection.generic_registrations.

PiperOrigin-RevId: 174092003
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
b9337de5b354c7869c01f4d0cc1eb40209b6290c 30-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: Multi-tower support for ConvKFCBasicFB

PiperOrigin-RevId: 173932013
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
434695921de7cfd713b789533173e1e0c3fc7691 23-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: _check_registration() supports multiple towers.

PiperOrigin-RevId: 173115870
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
670dddf4ad81c67fc76b370bf7b9d77263824358 23-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> Multi-minibatch support for
tf.contrib.kfac.fisher_blocks.FullyConnectedKFACBasicFB.

PiperOrigin-RevId: 173109677
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
eea089bdb66597c9e66180d39b94eea2c17be93e 23-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> K-FAC: Multi-tower support for ConvDiagonalFB.

PiperOrigin-RevId: 173105412
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
76ee352ccdb6927ca961c100b21b31f4431134b0 19-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> Multi-minibatch support for tf.contrib.kfac.fisher_blocks.FullyConnectedDiagonalFB.

PiperOrigin-RevId: 172695625
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
531b66789bd25291266aba1fc7f7d33ece7089ce 16-Oct-2017 A. Unique TensorFlower <gardener@tensorflow.org> Added a cleaner mechanism to set the global constants in fisher_blocks.py and fisher_factors.py in the form of a function "set_global_constants".

The old way of just manually setting these constants by importing the specific modules and accessing them directly should still work, but this new method is preferred.

PiperOrigin-RevId: 172345996
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
b1728aa3c5d2d8545acea781f1e2d6ffeccf3f7a 28-Sep-2017 A. Unique TensorFlower <gardener@tensorflow.org> - adding new FisherBlock / FisherFactor supporting diagonal approximations for conv layers

- added some more documentation to fisher_factors.py

PiperOrigin-RevId: 170384291
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py
ac13836b7d6920a09ce25e834a7ac1e1a4230740 28-Sep-2017 A. Unique TensorFlower <gardener@tensorflow.org> Initial release of tf.contrib.kfac

PiperOrigin-RevId: 170286115
/external/tensorflow/tensorflow/contrib/kfac/python/ops/fisher_blocks.py