d0849e8f3f
Implement MultiGPU training
...
Using nn.DataParallel (single process, multi GPU with cuda pipelining)
2020-01-14 10:52:39 -08:00
61b8db12bf
Split MQAN model in encoder and decoder
...
This will make it easier to replace the encoder without touching
the decoder.
2020-01-14 10:09:49 -08:00
c4a9c49d48
Remove max-margin loss
...
It doesn't work
2020-01-14 09:58:12 -08:00
1b891f38c1
Remove MultiLingualTranslationModel
...
It was never performing well, it was a lot of complex code, and it
was likely broken by the refactor but we don't care enough to fix it.
2020-01-14 09:45:31 -08:00
61e371def1
Remove distributed training
...
We want multi-gpu (with nn.DataParallel), and we don't care about
distributed. Also it was unfinished and buggy, and was adding complexity
to the code.
2020-01-14 09:43:09 -08:00
d569037078
Make tests pass
...
Redo how predict works, fix reversing
2020-01-14 09:26:33 -08:00
ef2b47e5be
wip: make tests pass
2020-01-13 18:15:30 -08:00
d04fe60e28
Update to Python 3.7
...
Python 3.6 is obsolete
2020-01-13 17:15:30 -08:00
428cc49012
Remove some usages of Field
2019-12-21 08:07:20 -08:00
ccd1fb6c80
Move vocabulary building to Vocab class
...
Leaving nothing in the Field class
2019-12-21 07:59:38 -08:00
cde004b3e4
Remove unused portions of Field
...
Everything related to preprocessing, tokenization, numericalization
has been removed and is implemented elsewhere.
2019-12-21 07:43:49 -08:00
ea83b75089
Remove all usage of Field from tasks and datasets
2019-12-21 07:36:28 -08:00
aa628629bf
Remove torchtest.Iterator and torchtext.Batch
...
Now unused
2019-12-21 06:56:23 -08:00
50b9c05248
Switch to torch.utils.data.DataLoader to load the datasets
...
Introduce new classes for numericalization and padding, in preparation
for moving away from torchtext.Field and torchtext.Batch. The
new classes are based on namedtuples, which are more compatible with
pytorch.
Move the useful parts of torchtext.Iterator (batching and bucketing
by example length) into a new Iterator class.
2019-12-21 06:52:14 -08:00
f6a2e4608d
Remove BPTTIterator and Batch.fromvars
...
Unused.
2019-12-21 02:33:51 -08:00
65891f62cb
Replace Example class with a namedtuple-based class
...
torch's data loader works with tuples or namedtuples, not with
custom classes. Also namedtuple are faster, and the code is easier
to read as we don't do as much meta-programming.
2019-12-21 02:31:52 -08:00
f5467c69cb
Remove TabularDataset class
...
Unused.
2019-12-21 01:54:03 -08:00
6059c11ac7
Remove context_question example field
...
Unused.
2019-12-21 01:52:45 -08:00
652657acb5
Remove all torchtext datasets
...
Move the remaining code in the CQA-based classes
2019-12-21 01:49:02 -08:00
c1e4cf2283
Fix refactoring fallout
2019-12-10 18:04:06 -08:00
27b8dca79f
Remove a whole lot of unused Almond code
...
Remove the "thingpedia_as_context" option, and the different ThingTalk
grammars. They were unused, and also unlikely to be helpful, after
several experiments.
2019-12-10 18:03:01 -08:00
58355e0ccc
Remove code related to differentiable BLEU loss
...
This code is unused, has questionable license, was never updated
to Torch 1.0, and seems unlikely to work.
2019-12-10 17:49:12 -08:00
61b152cd50
Collapse torchtext dependency
...
torchtext never really took off as the official PyTorch library
for text, and we've been using an embedded fork for a while.
Acknowledge that, merge the code, and stop treating it like
a separate project.
2019-12-10 17:40:10 -08:00
28a2b4d014
Remove obsolete, unused script
...
Related to WikiSQL
2019-12-10 16:42:27 -08:00
d927b33458
Remove old, unused ablation models
...
These are the models that were used to ablate MQAN. They are worse
than MQAN, and by removing them we can facilitate refactorings.
2019-12-10 16:40:38 -08:00
4cf9a1941f
Remove ELMo embeddings
...
Like Cove, ELMo embeddings are also obsolete, and the removal
allows us to remove the large allennlp library.
2019-12-10 16:38:23 -08:00
e7481840f7
Remove cove
...
Cove is an obsolete form of pretraining, of dubious utility. It
depends on an obsolete, unmaintained, library, and it causes
hacks in the code. Clean up.
2019-12-10 16:34:35 -08:00
6911a5092d
Revert "update FastText embedding"
...
This reverts commit 1339767da3
2019-11-06 08:22:36 -08:00
2507270c4b
Revert "fix"
...
This reverts commit 3addf7da94
2019-11-06 08:22:36 -08:00
f5627d6f2e
Revert "fix"
...
This reverts commit a5b885ecd4
2019-11-06 08:22:34 -08:00
a5b885ecd4
fix
2019-11-06 01:47:28 -08:00
3addf7da94
fix
2019-11-06 01:18:08 -08:00
1339767da3
update FastText embedding
2019-11-06 01:09:55 -08:00
4e10b5240b
Fix downloading fasttext vectors
2019-11-05 21:54:06 -08:00
06eca199d7
Merge pull request #19 from stanford-oval/wip/dockerupdate
...
docker: Switch to ubi8 (RHEL 8)
2019-11-04 16:22:35 -08:00
9cdb85abbb
Merge pull request #20 from stanford-oval/wip/i18n
...
Support word embeddings for arbitrary languages
2019-11-03 22:10:16 -08:00
2973165de0
util: load "locale" parameter from config.json
...
And if not present, set it to "en"
2019-11-03 21:25:57 -08:00
ba4c963176
fasttext: use common-crawl not wiki vectors
...
They are trained on a larger corpus and thus better
2019-11-01 17:38:38 -07:00
d5aacba674
embeddings: allow passing full locale tags as --locale
...
This way, we don't need to do anything too special in Genie,
and we can call decanlp with --locale zh-tw or --locale zh-cn
if needed to distinguish
2019-11-01 17:36:36 -07:00
d46353d352
Add fast-text support to cache-embeddings command
2019-11-01 17:30:59 -07:00
026e481df9
Add support for arbitrary word embeddings
...
FastText has all the word embeddings we care about anyway
2019-11-01 17:27:35 -07:00
668cc6f917
docker: Switch to ubi8 (RHEL 8)
...
ubi7 was updated to rhel 7.7, which should contain python36 natively,
but python3 is not part of the ubi repos. Luckily, it is part of the
ubi8 repos instead.
2019-11-01 17:08:06 -07:00
a546c7b62c
Fix cuda image name
2019-08-31 02:22:04 +02:00
28e8296032
Merge pull request #17 from stanford-oval/wip/docker
...
Remove obsolete Dockerfiles, and replace with a new one
2019-08-31 00:36:44 +02:00
dcb4f020ae
docker: fix hooks
...
They are executed from the directory containing the Dockerfile
2019-08-30 23:13:28 +02:00
86a51de115
MQAN: fix decoding with out of vocabulary words
...
If the vocabulary was limited due to max_generative_words and we
encounter an OOV word, we need to map it back to the full index
in GloVe, using the limited_idx_to_full_idx map set by the server
dynamically.
2019-08-30 21:39:17 +02:00
1c0a71aa0b
Add debugging to docker hooks
...
To see exactly how they get called...
2019-08-30 21:37:42 +02:00
fb72c84f73
docker: move embeddings to a shared directory
...
This way, the image can be used as a base image by almond-cloud
(which runs as a different user)
2019-08-30 21:25:54 +02:00
ff311bd35b
Add docker hub hooks to build both CPU and GPU images
2019-08-30 21:01:47 +02:00
a9264a699d
Remove obsolete Dockerfiles, and replace with a new one
...
A new one that wraps the new decanlp command and installs all
the dependencies correctly.
2019-08-30 18:55:28 +02:00
Giovanni Campagna