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

[PyTorch] Refactor Resize Token Embeddings (#8880)

Browse Source 
* fix resize tokens

* correct mobile_bert

* move embedding fix into modeling_utils.py

* refactor

* fix lm head resize

* refactor

* break lines to make sylvain happy

* add news tests

* fix typo

* improve test

* skip bart-like for now

* check if base_model = get(...) is necessary

* clean files

* improve test

* fix tests

* revert style templates

* Update templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_{{cookiecutter.lowercase_modelname}}.py
This commit is contained in:
Patrick von Platen
2020-12-02 19:19:50 +01:00
committed by GitHub
parent e52f9c0ade
commit 443f67e887
30 changed files with 273 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
tests/test_modeling_common.py
View File

@@ -815,6 +815,10 @@ class ModelTesterMixin:
# Check that the model can still do a forward pass successfully (every parameter should be resized)
# Input ids should be clamped to the maximum size of the vocabulary
inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
# make sure that decoder_input_ids are resized as well
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
model(**self._prepare_for_class(inputs_dict, model_class))
# Check that adding and removing tokens has not modified the first part of the embedding matrix.
@@ -825,6 +829,57 @@ class ModelTesterMixin:
self.assertTrue(models_equal)
def test_resize_embeddings_untied(self):
(
original_config,
inputs_dict,
) = self.model_tester.prepare_config_and_inputs_for_common()
if not self.test_resize_embeddings:
return
original_config.tie_word_embeddings = False
# if model cannot untied embeddings -> leave test
if original_config.tie_word_embeddings:
return
for model_class in self.all_model_classes:
config = copy.deepcopy(original_config)
model = model_class(config).to(torch_device)
# if no output embeddings -> leave test
if model.get_output_embeddings() is None:
continue
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
model_vocab_size = config.vocab_size
model.resize_token_embeddings(model_vocab_size + 10)
self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
output_embeds = model.get_output_embeddings()
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
# Check bias if present
if output_embeds.bias is not None:
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
model(**self._prepare_for_class(inputs_dict, model_class))
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
model.resize_token_embeddings(model_vocab_size - 15)
self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
# Check that it actually resizes the embeddings matrix
output_embeds = model.get_output_embeddings()
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
# Check bias if present
if output_embeds.bias is not None:
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
# Input ids should be clamped to the maximum size of the vocabulary
inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
model(**self._prepare_for_class(inputs_dict, model_class))
def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()