Sarthak Pati
|
9a44061ff1
|
added conda recipe
|
2022-02-01 17:31:29 +00:00 |
Miltos
|
5d3ef87287
|
Update oom.py
|
2022-01-05 09:09:57 +00:00 |
mir-am
|
5e173e9fbf
|
Fix invalid version number for torch-scatter in GH workflow
|
2021-12-19 07:39:39 +00:00 |
mir-am
|
f45dc7b062
|
Specify version of torch-scatter in GH workflow to fix the compilation error
|
2021-12-19 07:39:39 +00:00 |
mir-am
|
ff5034398e
|
Fix the required PyTorch version in dependencies and docs
|
2021-12-19 07:39:39 +00:00 |
Miltos Allamanis
|
53af8080f1
|
Use trick shown here https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html#use-parameter-grad-none-instead-of-model-zero-grad-or-optimizer-zero-grad
|
2021-10-21 20:10:52 +01:00 |
Miltos Allamanis
|
ef13a9fec0
|
Option to catch occasional CUDA OOMs to allow for more robust training.
|
2021-09-14 09:23:10 +01:00 |
Miltos
|
bf72e7f18a
|
Improve einsum notation
|
2021-09-03 11:30:18 +01:00 |
Miltos Allamanis
|
8da3c1d91e
|
Remove redundant `np.array` invocation.
|
2021-07-28 13:38:46 +01:00 |
Miltos Allamanis
|
6c55894406
|
Fix reported training loss when using AMP
|
2021-07-25 20:09:12 +01:00 |
Miltos Allamanis
|
f62a6165dd
|
Fixes.
|
2021-07-22 09:29:11 +01:00 |
Miltos Allamanis
|
03ab80499b
|
Explicitly set the starting epoch index to the restored one.
|
2021-07-22 09:29:11 +01:00 |
Miltos Allamanis
|
764a7038ff
|
Fix bugs in distributed saving of optimizer state.
|
2021-07-22 09:29:11 +01:00 |
Miltos Allamanis
|
3bd5966e88
|
Allow restoring optimizer/scheduler state.
|
2021-07-22 09:29:11 +01:00 |
Miltos Allamanis
|
5020bb7f6f
|
Improve some code comments about message-passing layers.
|
2021-07-21 09:42:34 +01:00 |
Miltos
|
27b3102184
|
Merge pull request #13 from microsoft/dev/miallama/distributedtrainer
A multi-GPU distributed trainer
|
2021-07-14 14:43:11 +01:00 |
Miltos Allamanis
|
ce0d8f6814
|
Fix AMP which was inconsistently implemented: https://pytorch.org/docs/stable/notes/amp_examples.html
|
2021-07-14 12:40:59 +01:00 |
Miltos Allamanis
|
6edb8aeeba
|
Reduce gpu->cpu transfers when progress bar is disabled.
|
2021-07-14 11:36:08 +00:00 |
Miltos Allamanis
|
c724767872
|
Remove unneeded gpu->cpu transfers during metrics update.
|
2021-07-14 12:04:03 +01:00 |
Miltos Allamanis
|
cd02705dab
|
Fix how validation-improved hooks work.
|
2021-07-09 12:08:15 +01:00 |
Miltos Allamanis
|
66f0046517
|
Improved hooks/AML support fixes.
|
2021-07-09 12:00:11 +01:00 |
Miltos Allamanis
|
80e8879469
|
Reduce insanely long buffering.
|
2021-07-08 19:16:04 +01:00 |
Miltos Allamanis
|
ed8aac4bb2
|
Merge branch 'dev/miallama/distributedtrainer' of github.com:microsoft/ptgnn into dev/miallama/distributedtrainer
|
2021-07-08 17:23:23 +01:00 |
Miltos Allamanis
|
95727b9802
|
Minor AML fixes.
|
2021-07-08 17:22:10 +01:00 |
Miltos Allamanis
|
b22b91fca7
|
Restore models in-cpu.
|
2021-07-08 15:45:47 +00:00 |
Miltos Allamanis
|
060765d6d6
|
Remove uncessary `enumerate`
|
2021-07-08 10:29:10 +01:00 |
Miltos Allamanis
|
b37f3830c0
|
Fix typos
|
2021-07-08 10:28:34 +01:00 |
Miltos Allamanis
|
948ba35646
|
Bypass deadlock by throwing on join()
|
2021-07-07 22:22:39 +00:00 |
Miltos Allamanis
|
a3a841676a
|
Repair bug.
|
2021-07-07 13:55:33 +00:00 |
Miltos Allamanis
|
25e9b29e28
|
Polish distributed trainer.
|
2021-07-07 08:03:23 +00:00 |
Miltos Allamanis
|
12496e4ee4
|
WIP implementation of distributed data parallel training.
|
2021-07-06 22:26:59 +00:00 |
Miltos
|
8df6d85368
|
Merge pull request #12 from microsoft/dev/miallam/ecg-implementation
An implementation of ECG
|
2021-04-16 08:05:45 +01:00 |
Miltos Allamanis
|
5e7e146cc0
|
An implementation of ECG
|
2021-04-12 13:54:48 +01:00 |
Miltos
|
d30cb0d585
|
Merge pull request #10 from microsoft/dev/miallama/better-names
Add more informative aliases for the GnnOutput
|
2021-03-16 14:24:30 +00:00 |
Miltos Allamanis
|
b36fd52f37
|
Add more informative aliases for the GnnOutput
|
2021-03-16 13:46:33 +00:00 |
Miltos Allamanis
|
9f0c0f0019
|
Fix AMP issue in PNA.
|
2021-03-12 21:13:57 +00:00 |
Miltos
|
207074d246
|
Merge pull request #8 from microsoft/dev/miallama/pna-layer
Add PNA layer.
|
2021-03-12 07:55:42 +00:00 |
Miltos Allamanis
|
6a00b4a65c
|
Simplify PNA aggregation/layers.
|
2021-03-11 17:12:06 +00:00 |
Miltos Allamanis
|
4d45e10a4f
|
Merge branch 'master' into dev/miallama/pna-layer
|
2021-03-11 16:21:22 +00:00 |
Miltos Allamanis
|
35b45d5f8b
|
Update AML files.
|
2021-03-11 16:20:50 +00:00 |
Miltos Allamanis
|
e627ca507d
|
Update AML files.
|
2021-03-11 16:19:29 +00:00 |
Miltos
|
66cf9eae29
|
Update setup.py
|
2021-02-23 12:47:01 +00:00 |
Miltos Allamanis
|
7f59fb0a84
|
Merge branch 'master' into dev/miallama/pna-layer
|
2021-02-22 10:15:47 +00:00 |
Miltos
|
7a23ce4910
|
Merge pull request #9 from microsoft/dev/miallama/edge-features
Allow using arbitrary edge features in GNNs
|
2021-02-22 10:12:23 +00:00 |
Miltos Allamanis
|
8dca26f8cc
|
Address Marc's feedback.
|
2021-02-22 09:47:35 +00:00 |
Miltos Allamanis
|
b42df7765d
|
Bug fix on no edge features.
|
2021-02-15 12:57:02 +00:00 |
Miltos Allamanis
|
e5e1aff090
|
Fixes to the edge-attrs.
|
2021-02-15 12:41:49 +00:00 |
Miltos Allamanis
|
d5702cda39
|
Allow adding features to edges.
|
2021-02-15 12:16:01 +00:00 |
Miltos Allamanis
|
bc5323d2fd
|
Minor AML updates.
|
2021-01-25 08:32:49 +00:00 |
Miltos Allamanis
|
b4525bec3d
|
Add PNA layer.
|
2021-01-12 14:59:34 +00:00 |