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

clean up all byte-level bpe tests

Browse Source
This commit is contained in:
thomwolf
2019-08-30 12:43:08 +02:00
parent ca1a00a302
commit 5dd7b677ad
3 changed files with 10 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
pytorch_transformers/tests/tokenization_gpt2_test.py
View File

@@ -57,12 +57,12 @@ class GPT2TokenizationTest(CommonTestCases.CommonTokenizerTester):
def test_full_tokenizer(self): def test_full_tokenizer(self):
tokenizer = GPT2Tokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map) tokenizer = GPT2Tokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
text = "lower newer" text = "lower newer"
bpe_tokens = ["\u0120low", "er", "\u0120newer"] bpe_tokens = ["\u0120low", "er", "\u0120", "n", "e", "w", "er"]
tokens = tokenizer.tokenize(text) tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, bpe_tokens) self.assertListEqual(tokens, bpe_tokens)
input_tokens = tokens + [tokenizer.unk_token] input_tokens = tokens + [tokenizer.unk_token]
input_bpe_tokens = [14, 15, 19] input_bpe_tokens = [14, 15, 10, 9, 3, 2, 15, 19]
self.assertListEqual( self.assertListEqual(
tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens) tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)

6
pytorch_transformers/tests/tokenization_roberta_test.py
View File

@@ -55,13 +55,13 @@ class RobertaTokenizationTest(CommonTestCases.CommonTokenizerTester):
def test_full_tokenizer(self): def test_full_tokenizer(self):
tokenizer = RobertaTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map) tokenizer = RobertaTokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
text = "lower" text = "lower newer"
bpe_tokens = ["\u0120low", "er"] bpe_tokens = ["\u0120low", "er", "\u0120", "n", "e", "w", "er"]
tokens = tokenizer.tokenize(text) tokens = tokenizer.tokenize(text)
self.assertListEqual(tokens, bpe_tokens) self.assertListEqual(tokens, bpe_tokens)
input_tokens = tokens + [tokenizer.unk_token] input_tokens = tokens + [tokenizer.unk_token]
input_bpe_tokens = [14, 15, 19] input_bpe_tokens = [14, 15, 10, 9, 3, 2, 15, 19]
self.assertListEqual( self.assertListEqual(
tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens) tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)

9
pytorch_transformers/tokenization_gpt2.py
View File

@@ -64,13 +64,14 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
@lru_cache() @lru_cache()
def bytes_to_unicode(): def bytes_to_unicode():
""" """
Returns list of utf-8 byte and a corresponding list of unicode strings. Returns list of utf-8 byte and a mapping to unicode strings.
We specifically avoids mapping to whitespace/control characters the bpe code barfs on.
The reversible bpe codes work on unicode strings. The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a signficant percentage of your normal, say, 32K bpe vocab. This is a signficant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings. To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
""" """
_chr = unichr if sys.version_info[0] == 2 else chr _chr = unichr if sys.version_info[0] == 2 else chr
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
@@ -176,9 +177,9 @@ class GPT2Tokenizer(PreTrainedTokenizer):
bpe_tokens = [] bpe_tokens = []
for token in re.findall(self.pat, text): for token in re.findall(self.pat, text):
if sys.version_info[0] == 2: if sys.version_info[0] == 2:
token = ''.join(self.byte_encoder[ord(b)] for b in token) token = ''.join(self.byte_encoder[ord(b)] for b in token) # Maps all our bytes to unicode strings, avoiding controle tokens of the BPE (spaces in our case)
else: else:
token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) # Maps all our bytes to unicode strings, avoiding controle tokens of the BPE (spaces in our case)
bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(' ')) bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(' '))
return bpe_tokens return bpe_tokens