4e5741a | tomhennigan | 21 June 2018, 12:23:37 UTC | Temporarily remove eager/graph annotation until the class level one is released. PiperOrigin-RevId: 201509633 | 17 July 2018, 15:55:38 UTC |
6926503 | Deepmind | 13 June 2018, 07:31:24 UTC | Add a bidirectional recurrent core to sonnet. Based off encoder component from: https://arxiv.org/pdf/1409.0473.pdf I avoided tf.while_loop as it complicates the implementation and we want explicit access to states and outputs from all steps of the sequence. PiperOrigin-RevId: 200346579 | 17 July 2018, 15:55:32 UTC |
4f34f4f | Deepmind | 11 June 2018, 14:01:31 UTC | Support for tf.bfloat16 in Sonnet. PiperOrigin-RevId: 200045112 | 17 July 2018, 15:55:26 UTC |
e61955b | tomhennigan | 08 June 2018, 16:03:33 UTC | Add eager mode tests for sonnet nets. PiperOrigin-RevId: 199800734 | 17 July 2018, 15:55:20 UTC |
ef9146d | mareynolds | 07 June 2018, 15:26:12 UTC | Fix dependencies in examples/BUILD - SciPy was missing. PiperOrigin-RevId: 199638219 | 17 July 2018, 15:55:01 UTC |
1eac5c6 | mareynolds | 07 June 2018, 15:00:26 UTC | Make dataset_nth_farthest python3 compatible. PiperOrigin-RevId: 199635116 | 17 July 2018, 15:54:55 UTC |
5332938 | mareynolds | 07 June 2018, 14:42:56 UTC | Make rmc_nth_farthest python3 compatible PiperOrigin-RevId: 199633253 | 17 July 2018, 15:54:49 UTC |
220832e | mareynolds | 07 June 2018, 14:05:08 UTC | Fix rmc_nth_farthest.ipynb PiperOrigin-RevId: 199629407 | 17 July 2018, 15:54:43 UTC |
9fe0a35 | Deepmind | 07 June 2018, 12:52:07 UTC | Adds demo for Relational Memory Core for "nth farthest" task from paper. PiperOrigin-RevId: 199622123 | 17 July 2018, 15:54:30 UTC |
31bc79c | diegolascasas | 05 June 2018, 16:52:26 UTC | Sonnet version update produced on Tuesday, 05. June 2018 PiperOrigin-RevId: 199312224 | 05 June 2018, 16:53:02 UTC |
0c9ed73 | mareynolds | 05 June 2018, 16:44:04 UTC | Updated changelog for version 1.23 PiperOrigin-RevId: 199311026 | 05 June 2018, 16:45:51 UTC |
05e2109 | mareynolds | 04 June 2018, 14:01:06 UTC | Fix error message in DeepRNN. Because we enumerate through core_sizes[1:] we need to add 1 to i when printing out which core we are referring to in the case of an error. Also fixed an issue where one of the shapes printed had the first dimension removed, and the other one didn't. PiperOrigin-RevId: 199127178 | 05 June 2018, 16:45:43 UTC |
3dd3d33 | mareynolds | 04 June 2018, 13:34:19 UTC | Make relational_memory python3 compatible. PiperOrigin-RevId: 199124668 | 05 June 2018, 16:45:35 UTC |
3ab2bea | Deepmind | 04 June 2018, 13:00:31 UTC | Change test assertion function. PiperOrigin-RevId: 199121362 | 05 June 2018, 16:45:26 UTC |
5b314d6 | Deepmind | 04 June 2018, 12:00:00 UTC | Add relational memory module to sonnet. Adding relational memory implementation from "Relational Recurrent Neural Networks", Santoro et al., 2018. PiperOrigin-RevId: 199116117 | 05 June 2018, 16:45:17 UTC |
b66dadc | diegolascasas | 01 June 2018, 14:09:19 UTC | Sonnet version update produced on Friday, 01. June 2018 PiperOrigin-RevId: 198867845 | 01 June 2018, 14:20:59 UTC |
71c16c1 | Deepmind | 01 June 2018, 12:01:17 UTC | Switch to using private member variables when possible. There can be subtle differences to the public and private versions of member variables in the conv module (eg, Convolution's `stride` is the conv size + 2 but `_stride` is the conv size). The current module mixes and matches these. To reduce cognitive load a bit, we switch to using the private member variable when possible. PiperOrigin-RevId: 198857690 | 01 June 2018, 14:20:52 UTC |
3ec0d90 | diegolascasas | 31 May 2018, 16:50:31 UTC | Additional changes to VQVAE notebook to make it python3 compatible. PiperOrigin-RevId: 198735118 | 01 June 2018, 10:48:37 UTC |
ce869c1 | mareynolds | 31 May 2018, 16:06:34 UTC | Make brnn_ptb_test write checkpoints to temp directory PiperOrigin-RevId: 198729764 | 01 June 2018, 10:48:31 UTC |
f6a930b | Deepmind | 31 May 2018, 11:09:15 UTC | MergeDims: also handle dimensions of size zero. PiperOrigin-RevId: 198699387 | 01 June 2018, 10:48:25 UTC |
6307466 | diegolascasas | 31 May 2018, 10:13:49 UTC | Use six to load cPickle in VQVAE notebook PiperOrigin-RevId: 198694898 | 01 June 2018, 10:48:18 UTC |
35778ef | mareynolds | 30 May 2018, 14:56:05 UTC | Updated changelog for Sonnet version 1.21 PiperOrigin-RevId: 198559887 | 30 May 2018, 14:59:18 UTC |
d6a02aa | mareynolds | 30 May 2018, 14:54:14 UTC | internal change PiperOrigin-RevId: 198559691 | 30 May 2018, 14:59:11 UTC |
45d0f76 | mareynolds | 30 May 2018, 14:29:25 UTC | Jupyter Notebook demonstrating VQ-VAE training on CIFAR-10. PiperOrigin-RevId: 198556907 | 30 May 2018, 14:59:05 UTC |
024f263 | mareynolds | 24 May 2018, 13:42:30 UTC | Add control dependencie to VectorQuantizerEMA to avoid potential non-deterministic results. PiperOrigin-RevId: 197884089 | 30 May 2018, 14:43:34 UTC |
f72c4a5 | Deepmind | 24 May 2018, 09:51:21 UTC | Improve Sonnet's MergeDim behaviour on partially defined shapes. PiperOrigin-RevId: 197863174 | 30 May 2018, 14:43:27 UTC |
c3d3bc2 | sracaniere | 22 May 2018, 12:16:54 UTC | Add support for custom getters in ConvNet2D PiperOrigin-RevId: 197546042 | 30 May 2018, 14:43:20 UTC |
8a7ca0a | Deepmind | 21 May 2018, 16:50:47 UTC | Add SeparableConv1D class to sonnet. PiperOrigin-RevId: 197408513 | 30 May 2018, 14:43:14 UTC |
85492de | Deepmind | 21 May 2018, 09:41:27 UTC | skip_connection deprecation warning is printed only once PiperOrigin-RevId: 197370137 | 30 May 2018, 14:43:07 UTC |
0277e53 | adriap | 08 May 2018, 15:52:31 UTC | Sonnet version update produced on Tuesday, 08. May 2018 PiperOrigin-RevId: 195825275 | 30 May 2018, 14:42:59 UTC |
c4ac2b1 | mareynolds | 04 May 2018, 11:56:10 UTC | Add copyright header PiperOrigin-RevId: 195399161 | 30 May 2018, 14:42:46 UTC |
c595084 | mareynolds | 04 May 2018, 11:46:13 UTC | Add VQ-VAE plus EMA variant to Sonnet. PiperOrigin-RevId: 195398448 | 30 May 2018, 14:42:36 UTC |
f34755c | mareynolds | 30 April 2018, 13:54:41 UTC | Add snt.summarize_variables to Sonnet. This prints a summary of #variables, #scalars and memory usage for each datatype. PiperOrigin-RevId: 194779974 | 08 May 2018, 15:49:30 UTC |
df245c0 | Deepmind | 25 April 2018, 16:50:27 UTC | Make an error more explicit when the Sonnet module name is not a string. PiperOrigin-RevId: 194254264 | 08 May 2018, 15:49:14 UTC |
c838ebf | adriap | 24 April 2018, 16:26:28 UTC | Sonnet version update produced on Tuesday, 24. April 2018 PiperOrigin-RevId: 194098244 | 24 April 2018, 16:28:00 UTC |
6db202c | mareynolds | 24 April 2018, 15:36:24 UTC | Make brnn_ptb and ptb_reader python3 compatible. Fixes https://github.com/deepmind/sonnet/issues/79 PiperOrigin-RevId: 194091717 | 24 April 2018, 15:51:37 UTC |
50b41df | mareynolds | 24 April 2018, 14:40:05 UTC | Update installation instructions PiperOrigin-RevId: 194085445 | 24 April 2018, 15:51:30 UTC |
a30da96 | mareynolds | 24 April 2018, 13:45:10 UTC | Add a unit test for brnn_ptb. Runs a small model with fake data for 1 epoch. PiperOrigin-RevId: 194079346 | 24 April 2018, 15:51:20 UTC |
5c7e355 | Deepmind | 16 April 2018, 09:31:21 UTC | Merge SeparableConv2D into _ConvND Merge the final convolution class into _ConvND. We also add a property field for `channel_multiplier` in `DepthwiseConv2D` as well as `SeparableConv2D`. PiperOrigin-RevId: 193007472 | 24 April 2018, 15:51:11 UTC |
1a313f6 | diegolascasas | 10 April 2018, 16:00:00 UTC | Sonnet version update produced on Tuesday, 10. April 2018 PiperOrigin-RevId: 192293232 | 10 April 2018, 16:01:18 UTC |
a86f044 | fviola | 09 April 2018, 15:35:19 UTC | Remove skip_connnection option from ConvLSTM. PiperOrigin-RevId: 192131304 | 10 April 2018, 14:10:07 UTC |
600268f | fviola | 09 April 2018, 15:17:07 UTC | Update test PiperOrigin-RevId: 192129236 | 10 April 2018, 14:10:01 UTC |
5980cf9 | Deepmind | 09 April 2018, 13:13:47 UTC | Move CausalConv related functionality into subclass. There are a few bits of functionality that are only related to CausalConv1D and don't need to be in _ConvND. We're moving them into the subclass and calling _ConvND machinery for the rest. PiperOrigin-RevId: 192117666 | 10 April 2018, 14:09:56 UTC |
1ca6866 | Deepmind | 09 April 2018, 13:03:43 UTC | Refactor DepthwiseConv2D to use _ConvND. Rewrite DepthwiseConv2D so that it shares as much functionality with _ConvND as possible. PiperOrigin-RevId: 192116657 | 10 April 2018, 14:09:50 UTC |
23cb8d9 | Deepmind | 09 April 2018, 11:22:14 UTC | Refactor InPlaneConv2D to use _ConvND This refactor removes almost all of the code in InPlaneConv2D and uses _ConvND instead. We also fix some tests that were failing when run on the GPU in conv_gpu_test due to numpy & tensorflow version upgrades. PiperOrigin-RevId: 192109587 | 10 April 2018, 14:09:45 UTC |
9802714 | Deepmind | 06 April 2018, 15:49:01 UTC | Add dilation rate tests for ConvNet2D. Adding testing for the `rates` argument in ConvNet2D. PiperOrigin-RevId: 191893167 | 10 April 2018, 14:09:39 UTC |
159adc5 | sracaniere | 26 March 2018, 12:21:05 UTC | Remove spurious spaces in doc. PiperOrigin-RevId: 190449915 | 10 April 2018, 14:09:34 UTC |
e660931 | adriap | 13 March 2018, 14:47:10 UTC | Use tensorflow nest for nest operations. PiperOrigin-RevId: 188871826 | 10 April 2018, 14:09:28 UTC |
1bee593 | sracaniere | 13 March 2018, 13:11:16 UTC | Use nest from pytflib to get rid of deprecation messages. PiperOrigin-RevId: 188862864 | 10 April 2018, 14:09:21 UTC |
4596fd0 | diegolascasas | 13 March 2018, 12:49:01 UTC | Sonnet version update produced on Monday, 12. March 2018 PiperOrigin-RevId: 188861040 | 13 March 2018, 12:50:56 UTC |
80d4784 | Deepmind | 09 March 2018, 13:53:39 UTC | Fix typo in DeepRNN docs. PiperOrigin-RevId: 188474053 | 12 March 2018, 14:10:13 UTC |
e331172 | sracaniere | 08 March 2018, 17:15:11 UTC | Support the root scope (an empty string) in get_variable_scope. PiperOrigin-RevId: 188341067 | 12 March 2018, 14:10:08 UTC |
7d65b8a | gabrielbm | 07 March 2018, 13:30:37 UTC | Improve the way reuse_variables handles name scopes This fixes a bug where outer name scopes were ignored by reuse_variables. PiperOrigin-RevId: 188163153 | 12 March 2018, 14:10:03 UTC |
b1191ad | mareynolds | 05 March 2018, 17:55:14 UTC | Docstring clarification PiperOrigin-RevId: 187880593 | 12 March 2018, 14:09:58 UTC |
e62f660 | Deepmind | 01 March 2018, 11:49:03 UTC | Add mention to get_all_variables in snt.Sequential warning. PiperOrigin-RevId: 187456723 | 12 March 2018, 14:09:53 UTC |
553159b | Deepmind | 23 February 2018, 17:35:53 UTC | Added optional dilation rates argument for ConvNet2D. PiperOrigin-RevId: 186779967 | 12 March 2018, 14:09:48 UTC |
c855672 | gabrielbm | 23 February 2018, 12:53:10 UTC | Sort variables returned by get_all_variables() by name. The variables returned by get_all_variables() are stored in a set, so they must be sorted to determine an ordering over them. This change sorts the variables by name before returning them. PiperOrigin-RevId: 186752578 | 12 March 2018, 14:09:42 UTC |
a1d1a37 | Deepmind | 22 February 2018, 15:16:07 UTC | Add missing argument to example code. PiperOrigin-RevId: 186613534 | 12 March 2018, 14:09:36 UTC |
108fce7 | mareynolds | 21 February 2018, 16:32:06 UTC | Bump version to 1.17 PiperOrigin-RevId: 186463459 | 21 February 2018, 17:32:51 UTC |
c522ae6 | Aditya Paliwal | 14 February 2018, 11:58:02 UTC | Updated README - TensorFlow v1.5 required GIT_ORIGIN_REV_ID=264bc35ad75f1dad605407295e857846237917be PiperOrigin-RevId: 185666441 | 21 February 2018, 16:10:30 UTC |
62f4399 | gabrielbm | 14 February 2018, 11:38:14 UTC | Implementation of get_all_variables() for sonnet modules. We introduce a module call stack, which tracks the order in which modules are called. When a module enters __call__ (or _enter_variable_scope) it adds itself to the top of the stack. Variables created inside of the custom_getter are added to a collection specific to the module on the top of the stack. Before exiting __call__ (or _enter_variable_scope) the module moves all variables added to this graph collection into `_all_variables`, removes itself from the top of the stack, and adds all of the variables from `self._all_variables` to collection for the module that is currently at the top of the module stack. PiperOrigin-RevId: 185664981 | 21 February 2018, 16:10:23 UTC |
06b8e4f | Deepmind | 13 February 2018, 14:50:57 UTC | Replace keep_dims with keepdims in call to tf.reduce_prod() PiperOrigin-RevId: 185524194 | 21 February 2018, 16:10:17 UTC |
f9e3859 | Deepmind | 13 February 2018, 11:50:15 UTC | Fix an error message which was incorrect. PiperOrigin-RevId: 185509698 | 21 February 2018, 16:10:10 UTC |
c75bca4 | Deepmind | 13 February 2018, 10:04:36 UTC | Fix comment. PiperOrigin-RevId: 185501790 | 21 February 2018, 16:10:03 UTC |
0737482 | Deepmind | 12 February 2018, 17:36:54 UTC | Internal change. PiperOrigin-RevId: 185389700 | 21 February 2018, 16:09:55 UTC |
551d652 | diegolascasas | 30 January 2018, 15:55:29 UTC | Fix Python3 incompatibilities in new tests and methods. PiperOrigin-RevId: 183831798 | 30 January 2018, 16:10:36 UTC |
b23489a | mareynolds | 29 January 2018, 17:58:15 UTC | Update changelog PiperOrigin-RevId: 183680938 | 29 January 2018, 17:59:14 UTC |
2136d31 | adriap | 29 January 2018, 17:55:18 UTC | Add backwards compatibility with tests for ones and zeros name scopes. PiperOrigin-RevId: 183680464 | 29 January 2018, 17:59:00 UTC |
1d2d99e | diegolascasas | 29 January 2018, 11:57:14 UTC | Remove fixed seed dependency in gated_rnn_test.LSTMTest.testRecurrentDropout PiperOrigin-RevId: 183644341 | 29 January 2018, 17:03:51 UTC |
132924d | Deepmind | 25 January 2018, 17:25:35 UTC | Refactor out bias construction and application. Refactor out the construction of the bias variable and its application to the output of all Convolution modules. PiperOrigin-RevId: 183249262 | 29 January 2018, 17:03:46 UTC |
84bab04 | mareynolds | 25 January 2018, 16:44:55 UTC | Enable custom_getter for TrainableVariable. PiperOrigin-RevId: 183243768 | 29 January 2018, 17:03:40 UTC |
438b320 | Deepmind | 25 January 2018, 14:33:20 UTC | Refactor out inputs verification. Remove duplicate code that is performing the same checks on the `inputs` tensor in every Convolution class. PiperOrigin-RevId: 183230711 | 29 January 2018, 17:03:35 UTC |
2a8e6e9 | tfgg | 24 January 2018, 13:09:04 UTC | Fix typo. PiperOrigin-RevId: 183071198 | 29 January 2018, 17:03:29 UTC |
de769df | Deepmind | 22 January 2018, 14:09:01 UTC | 1. Start supporting TensorShapes for the output_shape input for Conv*DTranspose classes. 2. Support the use of tf.Dimension objects instead of integers for dimensions. 3. Add tests to make sure output_shape inference works correctly. PiperOrigin-RevId: 182768071 | 29 January 2018, 17:03:24 UTC |
3aead32 | adriap | 22 January 2018, 11:46:16 UTC | Correct typo in warning. PiperOrigin-RevId: 182755077 | 29 January 2018, 17:03:18 UTC |
109d606 | Deepmind | 19 January 2018, 12:11:15 UTC | Make masked Conv2D usable with ResourceVariables by avoiding use of *=. PiperOrigin-RevId: 182516172 | 29 January 2018, 17:03:11 UTC |
7393cfb | mareynolds | 16 January 2018, 14:13:07 UTC | Explicitly enter the scope of the connected Graph in AbstractModule.get_variables() This fixeds the following inconsistency: ``` with tf.Graph().as_default() as graph: input = tf.constant(np.random.randn(16, 784)) lin = snt.Linear(output_size=256) output = lin(input) print(lin.get_variables()) # prints tuple of 2 variables print(lin.get_variables()) # prints empty tuple, as technically we are in a different Graph. ``` Also added a .graph readonly property to AbstractModule. PiperOrigin-RevId: 182044744 | 29 January 2018, 16:56:19 UTC |
7bdd40c | Deepmind | 15 January 2018, 20:38:19 UTC | Provide better examples for snt.reuse_variables. PiperOrigin-RevId: 181984575 | 29 January 2018, 16:56:13 UTC |
3943452 | Deepmind | 12 January 2018, 15:51:09 UTC | Push BBB library and Bayesian RNN example to open source sonnet. PiperOrigin-RevId: 181743627 | 29 January 2018, 16:56:08 UTC |
315719e | Deepmind | 12 January 2018, 11:35:07 UTC | Select the correct input channel and stride values within Conv{1,2,3}DTranspose.transpose() when the data_format is NC*. Add more tests for Conv*Transpose transpose functionality. PiperOrigin-RevId: 181724912 | 29 January 2018, 16:56:02 UTC |
7865f86 | Deepmind | 11 January 2018, 15:08:07 UTC | Improve an error message. This is supposing `data_format` is a string, but if one is incorrectly passing something else, the error message will fail e.g. with ValueError: Unknown format code 's' for object of type 'int' With this change the error becomes: ValueError: Invalid data_format 256. Allowed formats set(['NCHW', 'NHWC']) PiperOrigin-RevId: 181604070 | 29 January 2018, 16:55:57 UTC |
3a21b8d | diegolascasas | 11 January 2018, 11:56:09 UTC | Remove dependency on fixed seed for testZoneout. This means it won't be possible to test for expected state values. PiperOrigin-RevId: 181587515 | 29 January 2018, 16:55:48 UTC |
1d2171d | diegolascasas | 10 January 2018, 15:51:51 UTC | Sonnet version update produced on Monday, 8. January 2018 PiperOrigin-RevId: 181464426 | 10 January 2018, 17:19:34 UTC |
f11b5af | Deepmind | 09 January 2018, 11:04:00 UTC | Add Con[1-3]DTranspose classes to conv_gpu_tests. Expanding testing coverage of the transpose classes. PiperOrigin-RevId: 181293937 | 10 January 2018, 16:53:39 UTC |
3c18209 | Deepmind | 05 January 2018, 19:42:45 UTC | Add Causal1DConv testing to conv_gpu_test. PiperOrigin-RevId: 180955048 | 08 January 2018, 12:15:39 UTC |
a1b1f34 | Deepmind | 05 January 2018, 18:04:43 UTC | Stop using constant initializers in conv_gpu_test.py. Unit tests were using constant initializers to check for correctness in data_format permutation operations. This is problematic; you want the weights to be very random so that they don't inadvertently mask an issue in their functionality. PiperOrigin-RevId: 180941577 | 08 January 2018, 12:15:33 UTC |
fe91080 | Deepmind | 05 January 2018, 13:24:42 UTC | Rename SUPPORTED_DATA_FORMATS to SUPPORTED_2D_DATA_FORMATS. As we support far more than just 2D conv now, this variable needs to change. PiperOrigin-RevId: 180915317 | 08 January 2018, 12:15:28 UTC |
865f4aa | Deepmind | 05 January 2018, 12:41:21 UTC | Refactor CausalConv1D to use _ConvND. Tests for clone functionality in all _ConvND subclasses. CausalConv1D can exist using the new _ConvND superclass. Add support within _ConvND to accommodate padded input. Refactored _ConvND a bit to get smaller methods as _build() was getting unwieldy. With CausalConv1D inheriting from _ConvND, it is now cloneable by default. I noticed that most of the classes lacked testing of their clone functionality, so I added those. Finally, in supporting multiple data_formats, I noticed there aren't enough tests that check to see an exception is thrown when an invalid data_format is used. So I added those. PiperOrigin-RevId: 180912740 | 08 January 2018, 12:15:22 UTC |
699f8cb | Deepmind | 05 January 2018, 11:12:33 UTC | Refactor Conv{1,2,3}Transpose classes into one. Add tests for transpose functionality. Rearchitect the Transpose classes into one base class that is then instantiated for each dimension convolution we want. Add tests for Conv{1,3}Transpose; we make sure to test N*C and NC* data formats. PiperOrigin-RevId: 180907942 | 08 January 2018, 12:15:16 UTC |
511f11b | Deepmind | 03 January 2018, 20:47:11 UTC | Add a Recurrent Highway Network cell. PiperOrigin-RevId: 180705745 | 08 January 2018, 12:15:06 UTC |
5fe263b | mareynolds | 03 January 2018, 16:40:55 UTC | Improve snt.Embed performance in distributed training. Adds a fix to avoid excess computation on parameter servers. PiperOrigin-RevId: 180674688 | 08 January 2018, 12:15:01 UTC |
6f9f46e | Deepmind | 02 January 2018, 11:14:13 UTC | Update AbstractModule documentation. PiperOrigin-RevId: 180531448 | 08 January 2018, 12:14:55 UTC |
ca13720 | Deepmind | 22 December 2017, 13:32:52 UTC | Creation of a ConvND class. More flexible masks. Lots more tests. Creating a ConvND class that is then subclassed for Conv{1,2,3}D. This removes a lot of duplicated boilerplate code. As part of this we, are also adding a `mask` argument to the Conv{1,3}D classes. Finally, we remove rank restrictions on the mask argument. We add testing for new masking functionality. PiperOrigin-RevId: 179918026 | 08 January 2018, 12:14:49 UTC |
d0f50a6 | Deepmind | 22 December 2017, 12:47:04 UTC | Extend batch_norm_v2 data_format default across dimensions. PiperOrigin-RevId: 179915706 | 08 January 2018, 12:14:44 UTC |
89c2091 | Deepmind | 22 December 2017, 11:34:10 UTC | Improve snt.Embed performance in distributed training. Adds a fix to avoid excess computation on parameter servers. PiperOrigin-RevId: 179912285 | 08 January 2018, 12:14:38 UTC |
9a547ed | Deepmind | 21 December 2017, 16:56:19 UTC | Implemented snt.BatchNormV2, which differs from snt.BatchNorm in the following ways: * Automatically computes updates to moving statistics by default (i.e. update_ops_collection=None). * Uses moving statistics by default when testing (i.e. test_local_stats=False). * Takes a data_format string (NC/NWC/NCW/NHWC/NCHW/NDHWC/NCDHW) rather than axes; reduces along all non-C axes. * Uses fused batch normalization by default. If the data_format isn't NHWC or NCHW, reshapes the batch internally. * Uses flat variables for the moving statistics, scale, and offset so that they can be shared between different data_formats. PiperOrigin-RevId: 179819339 | 08 January 2018, 12:14:32 UTC |
b920570 | Deepmind | 20 December 2017, 19:52:49 UTC | tf.float16 support for batch_norm PiperOrigin-RevId: 179714930 | 08 January 2018, 12:14:26 UTC |
e2ce831 | mareynolds | 19 December 2017, 17:49:09 UTC | Improve error message when partitioners/regularizers/initializers used wrong. PiperOrigin-RevId: 179566444 | 08 January 2018, 12:14:20 UTC |
3b9f636 | mareynolds | 19 December 2017, 17:48:00 UTC | When Sequential contains no layers, simply act as identity. Previously, a sequential with no layers would return any input wrapped in a 1 element tuple, which is a strange inconsistency. PiperOrigin-RevId: 179566305 | 08 January 2018, 12:14:14 UTC |
a7a557e | mareynolds | 19 December 2017, 17:10:36 UTC | Improve error message when partitioners/regularizers/initializers used wrong. PiperOrigin-RevId: 179561994 | 08 January 2018, 12:14:09 UTC |