8265 Commits

Author SHA1 Message Date
Loïc Estève
6192516dfd
Use scikit-learn rather than sklearn in docs/requirements.txt 2021-06-21 12:03:44 +02:00
jax authors
50f48e8a48 Merge pull request #7027 from jakevdp:transpose-validation
PiperOrigin-RevId: 380364559
2021-06-19 09:38:21 -07:00
Jake VanderPlas
7d2057b5c7 BUG: fix validation of permutation in transpose 2021-06-18 21:54:28 -07:00
jax authors
9450b8f8f9 Merge pull request #6954 from lukepfister:complex_convolve
PiperOrigin-RevId: 380260457
2021-06-18 14:05:58 -07:00
Luke Pfister
c33388b136 Support complex numbers in jax.scipy.signal.convolve/correlate 2021-06-18 13:07:00 -06:00
jax authors
605389025c Merge pull request #6914 from njunge94:custom_root_aux
PiperOrigin-RevId: 380227301
2021-06-18 11:25:30 -07:00
jax authors
8e0ec1124e Merge pull request #6996 from jakevdp:sparse-grad
PiperOrigin-RevId: 380201282
2021-06-18 09:26:59 -07:00
jax authors
48c046c170 Merge pull request #7012 from jakevdp:fix-random-bits
PiperOrigin-RevId: 380194114
2021-06-18 08:47:50 -07:00
Jake VanderPlas
a385d975b9 [sparse] add sparse-aware wrapper for grad & value_and_grad 2021-06-18 08:18:56 -07:00
Peter Hawkins
4406c77086 [XLA] Improve the accuracy of complex log1p close to 0.
https://github.com/google/jax/issues/7004

PiperOrigin-RevId: 380188368
2021-06-18 08:12:23 -07:00
George Necula
a4e74a269c Disable testing dot_general with preferred_element_type on GPU.
Due to a XLA bug we get non-deterministic NaN on GPU.
This fixes flakiness in lax_test.py

PiperOrigin-RevId: 380179946
2021-06-18 07:12:54 -07:00
jax authors
9f8fb243f9 Merge pull request #7014 from gnecula:master_main
PiperOrigin-RevId: 380144198
2021-06-18 02:27:50 -07:00
George Necula
6a48c60a72 Rename master to main in embedded links.
Tried to avoid the change on external links to repos that
have not yet renamed master.
2021-06-18 10:00:01 +03:00
George Necula
4e58106fcb [jax2tf] Add back the limitation for dot_general with preferred_element_type on GPU
That case is still broken, it fails non-deterministically on GPU

PiperOrigin-RevId: 380122429
2021-06-17 23:23:13 -07:00
Jake VanderPlas
d20ecf5db8 random: generate large sequences via vmap rather than Python loops 2021-06-17 16:10:30 -07:00
jax authors
1d0c461ea6 Merge pull request #6995 from jakevdp:sparse-submodule
PiperOrigin-RevId: 380040999
2021-06-17 14:08:28 -07:00
jax authors
27ee4b64b0 Merge pull request #7008 from njunge94:index-concretizations
PiperOrigin-RevId: 380024740
2021-06-17 12:55:25 -07:00
Nicholas Junge
cba5b13a39 Improve concretization errors for jnp indexing routines 2021-06-17 20:26:16 +02:00
Nicholas Junge
ccc8bb7f19 Add auxiliary data support to lax.custom_root 2021-06-17 19:25:46 +02:00
Jake VanderPlas
623d3c7ad3 [sparse] create experimental sparse submodule 2021-06-17 08:09:31 -07:00
jax authors
6d805cae6e Merge pull request #7007 from gnecula:jax_change_6
PiperOrigin-RevId: 379947590
2021-06-17 06:51:49 -07:00
George Necula
8b320eca82 Test 6 for rename master to main 2021-06-17 16:33:06 +03:00
George Necula
26499a38e6 Merge pull request #7006 from gnecula:jax_test_5
PiperOrigin-RevId: 379942050
2021-06-17 15:25:38 +02:00
George Necula
4fa1506825 One more change, to tickle copybara 2021-06-17 16:08:54 +03:00
George Necula
268a242c65 Test 5 for rename master to main 2021-06-17 15:58:51 +03:00
jax authors
97a5719fcb Merge pull request #7002 from gnecula:jax_test_3
PiperOrigin-RevId: 379932113
2021-06-17 04:56:03 -07:00
George Necula
0ec53eec72 Test PR merged to main, before making main default
Want to see the GitHub actions working, and copybara import
2021-06-17 14:48:21 +03:00
jax authors
daef97b1a1 Merge pull request #6999 from gnecula:rename_master
PiperOrigin-RevId: 379929701
2021-06-17 04:38:00 -07:00
George Necula
a6c16119c9 Renaming master to main: step 1.
Here we just enable the GitHub actions to run on both master and main.
2021-06-17 14:25:01 +03:00
George Necula
0d16f3cafd [jax2tf] Remove workaround for tf.math.atan2 not working on [b]float16.
PiperOrigin-RevId: 379919580
2021-06-17 03:21:51 -07:00
jax authors
409684f9c2 Merge pull request #6982 from jakevdp:bcoo-dot-general
PiperOrigin-RevId: 379810039
2021-06-16 14:21:24 -07:00
jax authors
cac09d44c9 Merge pull request #6990 from zhangqiaorjc:fix_typos
PiperOrigin-RevId: 379807021
2021-06-16 14:07:27 -07:00
Qiao Zhang
57234e7eaa Fix typos and indent. 2021-06-16 11:10:42 -07:00
jax authors
4bae894862 Merge pull request #6918 from colemanliyah:LRU_system
PiperOrigin-RevId: 379736780
2021-06-16 09:22:34 -07:00
jax authors
ea1be21ff1 Merge pull request #6927 from apaszke:xmap-loops
PiperOrigin-RevId: 379736111
2021-06-16 09:19:07 -07:00
Jake VanderPlas
6003f79063 [sparse] add bcoo_rdot_general 2021-06-16 08:54:30 -07:00
Liyah Coleman
0e9f7de6c8 Add LRU eviction policy to persisent compilation cache 2021-06-16 15:38:28 +00:00
jax authors
8f5a78404b Merge pull request #6964 from gnecula:tf_gather
PiperOrigin-RevId: 379693316
2021-06-16 04:42:51 -07:00
jax authors
ea711b31a2 Merge pull request #6986 from gnecula:call_tf_x64
PiperOrigin-RevId: 379693059
2021-06-16 04:39:10 -07:00
George Necula
d394f6917b Added clipping of OOB indices to match XLA behavior 2021-06-16 14:24:28 +03:00
George Necula
c956e83df0 [jax2tf] Add a best-effort conversion for gather with enable_xla=False
In an attempt to support lax.gather in absence of XLA, e.g., for
TFjs converter, we add an alternative conversion for the case
when enable_xla=False. Previously, the conversion would fail in this case.

The important parts of the code are lifeted from an old PR #3486, which was
then removed in #4030 (to avoid having to think about out-of-bounds accesses).

There is still a discrepancy between the JAX and TF behavior for out-of-bounds
accesses.
2021-06-16 14:24:28 +03:00
George Necula
e405a73663 [call_tf] Fixed the handling of x64 values in JAX_ENABLE_X64=False
TensorFlow interprets Python scalars in x64 mode, or it may contain
explicit x64 computation. When used with call_tf, we canonicalize
inputs and outputs using the JAX rules.
2021-06-16 13:25:56 +03:00
jax authors
ed96cac457 Merge pull request #6985 from gnecula:call_tf_grad
PiperOrigin-RevId: 379681210
2021-06-16 03:05:00 -07:00
George Necula
9831371e38 [jax2tf] Fix the custom gradients for call_tf
The previous implementation did not work correctly when
the call_tf function was embedded in a larger JAX function
for which we take the gradient.

Also fixed handling of non-float args/results.
2021-06-16 12:04:32 +03:00
jax authors
a82f0d8355 Merge pull request #6984 from jakevdp:bcoo-traced-repr
PiperOrigin-RevId: 379592043
2021-06-15 15:49:13 -07:00
Jake VanderPlas
591507a3eb [sparse] update sparse object repr to show tracing 2021-06-15 15:16:58 -07:00
jax authors
096f77dc70 Merge pull request #6983 from skye:test_fix
PiperOrigin-RevId: 379572178
2021-06-15 14:20:57 -07:00
Skye Wanderman-Milne
63355a9a25 Add jaxlib version check on newly-added NumpyLinalgTest.testEighZeroDiagonal test.
This is testing behavior recently added to jaxlib, so fails on older jaxlibs.
2021-06-15 20:44:52 +00:00
Adam Paszke
1e288b2f42 Add support for anonymous serial loops 2021-06-15 18:15:44 +00:00
jax authors
fec82c8aa1 Merge pull request #6980 from jakevdp:fix-faq
PiperOrigin-RevId: 379511290
2021-06-15 09:53:12 -07:00