SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[T5] allow config.decoder_layers to control decoder size (#7409)

Browse Source 
* Working assymmetrical T5

* rename decoder_layers -> num_decoder_layers

* Fix docstring

* Allow creation of asymmetric t5 students
This commit is contained in:
Sam Shleifer
2020-09-28 03:08:04 -04:00
committed by GitHub
parent 7296fea1d6
commit 748425d47d
5 changed files with 58 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
examples/seq2seq/test_make_student.py
View File

@@ -21,10 +21,8 @@ class MakeStudentTester(unittest.TestCase):
student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=1)
self.assertEqual(student.config.num_hidden_layers, 1)
def test_invalid_t5(self):
# T5 students must have the same e==d because there is only one config property
with self.assertRaises(AssertionError):
student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=None)
def test_asymmetric_t5(self):
student, *_ = create_student_by_copying_alternating_layers(TINY_T5, tempfile.mkdtemp(), e=1, d=None)
def test_same_decoder_small_encoder(self):
student, *_ = create_student_by_copying_alternating_layers(TINY_BART, tempfile.mkdtemp(), e=1, d=None)