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

🚨🚨🚨 TF: Remove TFWrappedEmbeddings (breaking: TF embedding initialization updated for encoder-decoder models) (#19263)

Browse Source 
* added test

* correct embedding init

* some changes in blenderbot (incomplete)

* update blenderbot (diff to be used as reference)

* update blenderbot_small

* update LED

* update marian

* update T5 and remove TFWrappedEmbeddings

* nullcontext() -> ContextManagers()

* fix embedding init
This commit is contained in:
Joao Gante
2022-10-11 16:48:03 +01:00
committed by GitHub
parent 8e4ee28e34
commit 462cd641d9
18 changed files with 515 additions and 1032 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
tests/models/led/test_modeling_tf_led.py
View File

@@ -228,69 +228,6 @@ class TFLEDModelTest(TFModelTesterMixin, unittest.TestCase):
name = model.get_bias()
assert name is None
def test_resize_token_embeddings(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def _get_word_embedding_weight(model, embedding_layer):
if hasattr(embedding_layer, "weight"):
return embedding_layer.weight
else:
# Here we build the word embeddings weights if not exists.
# And then we retry to get the attribute once built.
model(model.dummy_inputs)
if hasattr(embedding_layer, "weight"):
return embedding_layer.weight
else:
return None
for model_class in self.all_model_classes:
for size in [config.vocab_size - 10, config.vocab_size + 10, None]:
# build the embeddings
model = model_class(config=config)
old_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
old_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
old_final_logits_bias = model.get_bias()
# reshape the embeddings
model.resize_token_embeddings(size)
new_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
new_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
new_final_logits_bias = model.get_bias()
# check that the resized embeddings size matches the desired size.
assert_size = size if size is not None else config.vocab_size
self.assertEqual(new_input_embeddings.shape[0], assert_size)
# check that weights remain the same after resizing
models_equal = True
for p1, p2 in zip(old_input_embeddings.value(), new_input_embeddings.value()):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
if old_output_embeddings is not None and new_output_embeddings is not None:
self.assertEqual(new_output_embeddings.shape[0], assert_size)
models_equal = True
for p1, p2 in zip(old_output_embeddings.value(), new_output_embeddings.value()):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
if old_final_logits_bias is not None and new_final_logits_bias is not None:
old_final_logits_bias = old_final_logits_bias["final_logits_bias"]
new_final_logits_bias = new_final_logits_bias["final_logits_bias"]
self.assertEqual(new_final_logits_bias.shape[0], 1)
self.assertEqual(new_final_logits_bias.shape[1], assert_size)
models_equal = True
for old, new in zip(old_final_logits_bias.value(), new_final_logits_bias.value()):
for p1, p2 in zip(old, new):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
inputs_dict["global_attention_mask"] = tf.zeros_like(inputs_dict["attention_mask"])
@@ -374,20 +311,6 @@ class TFLEDModelTest(TFModelTesterMixin, unittest.TestCase):
pass
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
if a is None and b is None:
return True
try:
if tf.debugging.assert_near(a, b, atol=atol):
return True
raise
except Exception:
if len(prefix) > 0:
prefix = f"{prefix}: "
raise AssertionError(f"{prefix}{a} != {b}")
def _long_tensor(tok_lst):
return tf.constant(tok_lst, dtype=tf.int32)