Paul O'Leary McCann
2e8f0e9168
Rename coref params
2022-05-16 16:50:10 +09:00
kadarakos
7cf6bcca0e
merge misery
2022-05-10 17:19:16 +00:00
Paul O'Leary McCann
33f4f90ff0
Formatting
2022-05-10 19:09:52 +09:00
Paul O'Leary McCann
f852c5cea4
Split span predictor component into its own file
...
This runs. The imports in both of the split files could probably use a
close check to remove extras.
2022-05-10 18:53:45 +09:00
Paul O'Leary McCann
afd255c0ed
Undo multiply by 100
...
This was mistaken, not sure why my score seemed to be off before.
2022-04-14 18:42:09 +09:00
Paul O'Leary McCann
08729e0fbd
Remove end adjustment
...
The difference in environments was due to a change in Thinc, the code
here is fine.
2022-04-14 18:31:30 +09:00
Paul O'Leary McCann
8181d4570c
Multiply accuracy by 100
...
This seems to match with the scorer expectations better
2022-04-14 15:56:38 +09:00
Paul O'Leary McCann
e8af02700f
Remove all coref scoring exept LEA
...
This is necessary because one of the three old methods relied on scipy
for some complex problem solving. LEA is generally better for
evaluations.
The downside is that this means evaluations aren't comparable with many
papers, but canonical scoring can be supported using external eval
scripts or other methods.
2022-04-13 21:02:18 +09:00
Paul O'Leary McCann
2300f4df3d
Fix span score logging
2022-04-13 20:37:06 +09:00
Paul O'Leary McCann
d470fa03c1
Adjust end indices
...
It's not clear if this is technically correct or not but it won't run
without it for me.
2022-04-13 20:19:21 +09:00
kadarakos
b53113e3b8
Preparing span predictor for predicting from gold ( #10547 )
...
Note this is squashed because rebasing had conflicts.
* remove unnecessary .device
* span predictor debug start
* gearing up SpanPredictor for gold-heads
* merge SpanPredictor attributes
* remove useless extra prefix and device from spanpredictor
* make sure predicted and reference keeps aligned
* handle empty head_ids
* handle empty clusters
* addressing suggestions by @polm
* nicer restore
* fix score overwriting bug
* prepare for aligned heads-spans training
* span accuracy score
* update with eg.predited as other components
* add backprop callback to spanpredictor
* report start- and end-accuracies separately
* fixing scorer
Co-authored-by: Kádár Ákos <akos@onyx.uvt.nl>
2022-04-13 19:42:49 +09:00
Kádár Ákos
6aedd98d02
fixing scorer
2022-04-11 16:10:14 +02:00
Kádár Ákos
7a239f2ec7
report start- and end-accuracies separately
2022-04-08 14:57:19 +02:00
Kádár Ákos
3ba913109d
update with eg.predited as other components
2022-04-07 13:20:12 +02:00
Kádár Ákos
ef141ad399
span accuracy score
2022-04-04 18:10:09 +02:00
Kádár Ákos
a1d0219903
prepare for aligned heads-spans training
2022-04-04 15:26:15 +02:00
Kádár Ákos
63a41ba50a
fix score overwriting bug
2022-03-30 17:28:20 +02:00
Kádár Ákos
7ff99a3acc
nicer restore
2022-03-28 18:16:41 +02:00
Kádár Ákos
06d680b269
addressing suggestions by @polm
2022-03-28 14:31:51 +02:00
Kádár Ákos
e4b4b67ef6
handle empty clusters
2022-03-28 11:29:00 +02:00
Kádár Ákos
7304604edd
make sure predicted and reference keeps aligned
2022-03-25 18:29:33 +01:00
Kádár Ákos
83ac0477c8
remove useless extra prefix and device from spanpredictor
2022-03-24 16:44:50 +01:00
Kádár Ákos
706b2e6f25
gearing up SpanPredictor for gold-heads
2022-03-24 16:06:20 +01:00
Kádár Ákos
1eaf8fb0cf
span predictor debug start
2022-03-23 11:24:27 +01:00
Paul O'Leary McCann
2190cbc0e6
Add progress on SpanPredictor component
...
This isn't working. There is a CUDA error in the torch code during
initialization and it's not clear why.
2022-03-19 19:39:49 +09:00
Paul O'Leary McCann
a098849112
Add fake batching
...
The way fake batching works is that the pipeline component calls the
model repeatedly in a loop internally. It feels like this should break
something, but it worked in testing.
Another issue is that this changes the signature of some of the pipeline
functions, though I don't think that's an issue.
Tested with batch size of 2, so more testing is needed, but this is a
start.
2022-03-18 19:46:58 +09:00
Paul O'Leary McCann
1a79d18796
Formatting
2022-03-16 20:10:47 +09:00
Paul O'Leary McCann
6855df0e66
Skeleton for span predictor component
...
This should be moved into its own file, but for now just stubbing out
the methods.
2022-03-16 20:09:33 +09:00
Paul O'Leary McCann
7811a1194b
Change architecture
2022-03-16 14:57:15 +09:00
Paul O'Leary McCann
55039a66ad
Remove old default config
2022-03-15 19:53:09 +09:00
Paul O'Leary McCann
17d017a177
Remove span2head
...
This doesn't work as a component because it needs to modify gold data,
so instead it's a conversion script (in another repo).
2022-03-15 19:52:20 +09:00
Paul O'Leary McCann
0522a43116
Make span2head component
2022-03-15 19:19:15 +09:00
Paul O'Leary McCann
dfec6993d6
Training works now
2022-03-14 19:27:23 +09:00
Paul O'Leary McCann
8eadf3781b
Training runs now
...
Evaluation needs fixing, and code still needs cleanup.
2022-03-14 19:02:17 +09:00
Paul O'Leary McCann
d22a002641
Forward/backward pass works
...
Evaluate does not work - predict hasn't been updated
2022-03-14 17:26:27 +09:00
Paul O'Leary McCann
230698dc83
Fix bug in scorer
...
Scoring code was just using one metric, not all three of interest.
2021-08-12 18:22:08 +09:00
Paul O'Leary McCann
8bd0474730
Run black
2021-07-18 20:20:22 +09:00
Paul O'Leary McCann
bc081c24fa
Add full traditional scoring
...
This calculates scores as an average of three metrics. As noted in the
code, these metrics all have issues, but we want to use them to match up
with prior work.
This should be replaced with some simpler default scoring and the scorer
here should be moved to an external project to be passed in just for
generating the traditional scores.
2021-07-18 20:13:10 +09:00
Paul O'Leary McCann
80a17071d3
Remove unused code
2021-07-11 18:46:39 +09:00
Paul O'Leary McCann
447c7070e3
Fix loss
...
Accidentally deleted it
2021-07-10 22:45:25 +09:00
Paul O'Leary McCann
e00bd422d9
Fix span embeds
...
Some of the lengths and backprop weren't right.
Also various cleanup.
2021-07-10 21:38:53 +09:00
Paul O'Leary McCann
8f66176b2d
Fix loss?
...
This rewrites the loss to not use the Thinc crossentropy code at all.
The main difference here is that the negative predictions are being
masked out (= marginalized over), but negative gradient is still being
reflected.
I'm still not sure this is exactly right but models seem to train
reliably now.
2021-07-05 18:17:10 +09:00
Paul O'Leary McCann
2d3c559dc4
On initialize, use just two samples
...
Coref docs are kind of long, and using 10 samples on a smallish GPU can
cause OOMs.
2021-07-03 18:43:03 +09:00
Paul O'Leary McCann
f2e0e9dc28
Move placeholder handling into model code
2021-07-03 18:38:48 +09:00
Paul O'Leary McCann
a62121e3b4
Expose more hyperparameters
2021-06-17 21:21:46 +09:00
Paul O'Leary McCann
67d9ebc922
Transpose before calculating loss
2021-06-04 17:56:08 +09:00
svlandeg
04b55bf054
removing unused imports
2021-05-27 16:31:38 +02:00
svlandeg
910026582d
set versions to v1 instead of v0
2021-05-27 16:17:20 +02:00
Paul O'Leary McCann
a484245f35
Remove references to coref_er
2021-05-24 19:08:45 +09:00
Paul O'Leary McCann
d6389b133d
Don't use a generator for no reason
2021-05-24 19:06:15 +09:00
