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

[Pegasus] Refactor Tokenizer (#8731)

Browse Source 
* refactor

* further refactor

* fix the rest tomorrow

* save intermediate

* finish slow tokenizer

* make more tests pass

* finish refactor

* fix comment

* clean further

* fix name

* fix naming

* Update src/transformers/models/reformer/tokenization_reformer.py

* Apply suggestions from code review

* Apply suggestions from code review

* refactor

* fix init tokenizers

* refactor

* improve convert

* refactor

* correct convert slow tokenizer

* final fix for Pegasus Tok

* remove ipdb

* improve links
This commit is contained in:
Patrick von Platen
2020-11-29 16:57:43 +01:00
committed by GitHub
parent 36b60ce9e8
commit 5ced23dc84
11 changed files with 327 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
tests/test_tokenization_pegasus.py
View File

@@ -26,21 +26,34 @@ class PegasusTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer.save_pretrained(self.tmpdirname)
@cached_property
def pegasus_large_tokenizer(self):
def _large_tokenizer(self):
return PegasusTokenizer.from_pretrained("google/pegasus-large")
@unittest.skip("add_tokens does not work yet")
def test_swap_special_token(self):
pass
def get_tokenizer(self, **kwargs) -> PegasusTokenizer:
return PegasusTokenizer.from_pretrained(self.tmpdirname, **kwargs)
def get_input_output_texts(self, tokenizer):
return ("This is a test", "This is a test")
def test_pegasus_large_tokenizer_settings(self):
tokenizer = self.pegasus_large_tokenizer
def test_mask_tokens_rust_pegasus(self):
rust_tokenizer = self.rust_tokenizer_class.from_pretrained(self.tmpdirname)
py_tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname)
raw_input_str = "Let's see which <unk> is the better <unk_token_11> one <mask_1> It seems like this <mask_2> was important </s> <pad> <pad> <pad>"
rust_ids = rust_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
py_ids = py_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
# TODO: (Thom, Patrick) - this fails because the rust tokenizer does not know about the <mask_1>, <mask_2>, and those <unk_token_x> yet
self.assertListEqual(py_ids, rust_ids)
def test_large_mask_tokens(self):
tokenizer = self._large_tokenizer
# <mask_1> masks whole sentence while <mask_2> masks single word
raw_input_str = "<mask_1> To ensure a <mask_2> flow of bank resolutions."
desired_result = [2, 413, 615, 114, 3, 1971, 113, 1679, 10710, 107, 1]
ids = tokenizer([raw_input_str], return_tensors=None).input_ids[0]
self.assertListEqual(desired_result, ids)
def test_large_tokenizer_settings(self):
tokenizer = self._large_tokenizer
# The tracebacks for the following asserts are **better** without messages or self.assertEqual
assert tokenizer.vocab_size == 96103
assert tokenizer.pad_token_id == 0
@@ -48,20 +61,18 @@ class PegasusTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
assert tokenizer.offset == 103
assert tokenizer.unk_token_id == tokenizer.offset + 2 == 105
assert tokenizer.unk_token == "<unk>"
assert tokenizer.mask_token is None
assert tokenizer.mask_token_id is None
assert tokenizer.model_max_length == 1024
raw_input_str = "To ensure a smooth flow of bank resolutions."
desired_result = [413, 615, 114, 2291, 1971, 113, 1679, 10710, 107, 1]
ids = tokenizer([raw_input_str], return_tensors=None).input_ids[0]
self.assertListEqual(desired_result, ids)
assert tokenizer.convert_ids_to_tokens([0, 1, 2]) == ["<pad>", "</s>", "unk_2"]
assert tokenizer.convert_ids_to_tokens([0, 1, 2, 3]) == ["<pad>", "</s>", "<mask_1>", "<mask_2>"]
@require_torch
def test_pegasus_large_seq2seq_truncation(self):
def test_large_seq2seq_truncation(self):
src_texts = ["This is going to be way too long." * 150, "short example"]
tgt_texts = ["not super long but more than 5 tokens", "tiny"]
batch = self.pegasus_large_tokenizer.prepare_seq2seq_batch(
batch = self._large_tokenizer.prepare_seq2seq_batch(
src_texts, tgt_texts=tgt_texts, max_target_length=5, return_tensors="pt"
)
assert batch.input_ids.shape == (2, 1024)