[T5] allow config.decoder_layers to control decoder size #7409
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
sshleifer
commented
Sep 26, 2020
| @@ -57,6 +57,8 @@ class T5Config(PretrainedConfig): | |||
| Size of the intermediate feed forward layer in each :obj:`T5Block`. | |||
| num_layers (:obj:`int`, `optional`, defaults to 6): | |||
| Number of hidden layers in the Transformer encoder. | |||
There was a problem hiding this comment.
this was previously documented incorrectly. Now it is correct!
Codecov Report
@@ Coverage Diff @@
## master #7409 +/- ##
==========================================
- Coverage 77.58% 76.85% -0.74%
==========================================
Files 181 181
Lines 35725 35784 +59
==========================================
- Hits 27719 27501 -218
- Misses 8006 8283 +277
Continue to review full report at Codecov.
|
sgugger
approved these changes
Sep 27, 2020
src/transformers/configuration_t5.py
Outdated
| @@ -57,6 +57,8 @@ class T5Config(PretrainedConfig): | |||
| Size of the intermediate feed forward layer in each :obj:`T5Block`. | |||
| num_layers (:obj:`int`, `optional`, defaults to 6): | |||
| Number of hidden layers in the Transformer encoder. | |||
| decoder_layers (:obj:`int`, `optional`, defaults to num_layers): | |||
There was a problem hiding this comment.
Documented defaults should match the signature. It will eventually default to num_layers and this should be said, but in the line after, suggestion:
decoder_layers (:obj:`int`, `optional`):
Number of hidden layers in the Transformer decoder. Will use the same value as :obj:`num_layers` if not set.
patrickvonplaten
approved these changes
Sep 27, 2020
|
Renamed it to |
thevasudevgupta
pushed a commit
to thevasudevgupta/transformers
that referenced
this pull request
Sep 29, 2020
…#7409) * Working assymmetrical T5 * rename decoder_layers -> num_decoder_layers * Fix docstring * Allow creation of asymmetric t5 students
fabiocapsouza
pushed a commit
to fabiocapsouza/transformers
that referenced
this pull request
Nov 15, 2020
…#7409) * Working assymmetrical T5 * rename decoder_layers -> num_decoder_layers * Fix docstring * Allow creation of asymmetric t5 students
fabiocapsouza
added a commit
to fabiocapsouza/transformers
that referenced
this pull request
Nov 15, 2020
…gingface#7409)" This reverts commit c352b87.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Problem
arxiv.org/abs/2006.10369, among others, shows that models with fewer decoder layers than encoder layers can perform well and run generation much faster. Right now it is difficult to do distillation on t5 because there is only
T5Config.num_layerswhich controls encoder layers and decoder layers.Solution
config.decoder_layersto control decoder num layersconfig.decoder_layers = num_layersTesting