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

Add LUKE (#11223)

Browse Source 
* Rebase with master

* Minor bug fix in docs

* Copy files from adding_luke_v2 and improve docs

* change the default value of use_entity_aware_attention to True

* remove word_hidden_states

* fix head models

* fix tests

* fix the conversion script

* add integration tests for the pretrained large model

* improve docstring

* Improve docs, make style

* fix _init_weights for pytorch 1.8

* improve docs

* fix tokenizer to construct entity sequence with [MASK] entity when entities=None

* Make fix-copies

* Make style & quality

* Bug fixes

* Add LukeTokenizer to init

* Address most comments by @patil-suraj and @LysandreJik

* rename _compute_extended_attention_mask to get_extended_attention_mask

* add comments to LukeSelfAttention

* fix the documentation of the tokenizer

* address comments by @patil-suraj, @LysandreJik, and @sgugger

* improve docs

* Make style, quality and fix-copies

* Improve docs

* fix docs

* add "entity_span_classification" task

* update example code for LukeForEntitySpanClassification

* improve docs

* improve docs

* improve the code example in luke.rst

* rename the classification layer in LukeForEntityClassification from typing to classifier

* add bias to the classifier in LukeForEntitySpanClassification

* update docs to use fine-tuned hub models in code examples of the head models

* update the example sentences

* Make style & quality

* Add require_torch to tokenizer tests

* Add require_torch to tokenizer tests

* Address comments by @sgugger and add community notebooks

* Make fix-copies

Co-authored-by: Ikuya Yamada <ikuya@ikuya.net>
This commit is contained in:
NielsRogge
2021-05-03 15:07:29 +02:00
committed by GitHub
parent 6a11e4c2ad
commit f3cf8ae7b3
19 changed files with 4707 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

609
tests/test_modeling_luke.py Normal file
View File

@@ -0,0 +1,609 @@
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Testing suite for the PyTorch LUKE model. """
import unittest
from transformers import is_torch_available
from transformers.testing_utils import require_torch, slow, torch_device
from .test_configuration_common import ConfigTester
from .test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
if is_torch_available():
import torch
from transformers import (
LukeConfig,
LukeForEntityClassification,
LukeForEntityPairClassification,
LukeForEntitySpanClassification,
LukeModel,
LukeTokenizer,
)
from transformers.models.luke.modeling_luke import LUKE_PRETRAINED_MODEL_ARCHIVE_LIST
class LukeModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
entity_length=3,
mention_length=5,
use_attention_mask=True,
use_token_type_ids=True,
use_entity_ids=True,
use_entity_attention_mask=True,
use_entity_token_type_ids=True,
use_entity_position_ids=True,
use_labels=True,
vocab_size=99,
entity_vocab_size=10,
entity_emb_size=6,
hidden_size=32,
num_hidden_layers=5,
num_attention_heads=4,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_entity_classification_labels=9,
num_entity_pair_classification_labels=6,
num_entity_span_classification_labels=4,
use_entity_aware_attention=True,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.entity_length = entity_length
self.mention_length = mention_length
self.use_attention_mask = use_attention_mask
self.use_token_type_ids = use_token_type_ids
self.use_entity_ids = use_entity_ids
self.use_entity_attention_mask = use_entity_attention_mask
self.use_entity_token_type_ids = use_entity_token_type_ids
self.use_entity_position_ids = use_entity_position_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.entity_vocab_size = entity_vocab_size
self.entity_emb_size = entity_emb_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_entity_classification_labels = num_entity_classification_labels
self.num_entity_pair_classification_labels = num_entity_pair_classification_labels
self.num_entity_span_classification_labels = num_entity_span_classification_labels
self.scope = scope
self.use_entity_aware_attention = use_entity_aware_attention
self.encoder_seq_length = seq_length
self.key_length = seq_length
self.num_hidden_states_types = 2 # hidden_states and entity_hidden_states
def prepare_config_and_inputs(self):
# prepare words
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
attention_mask = None
if self.use_attention_mask:
attention_mask = random_attention_mask([self.batch_size, self.seq_length])
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
# prepare entities
entity_ids = ids_tensor([self.batch_size, self.entity_length], self.entity_vocab_size)
entity_attention_mask = None
if self.use_entity_attention_mask:
entity_attention_mask = random_attention_mask([self.batch_size, self.entity_length])
entity_token_type_ids = None
if self.use_token_type_ids:
entity_token_type_ids = ids_tensor([self.batch_size, self.entity_length], self.type_vocab_size)
entity_position_ids = None
if self.use_entity_position_ids:
entity_position_ids = ids_tensor(
[self.batch_size, self.entity_length, self.mention_length], self.mention_length
)
sequence_labels = None
entity_classification_labels = None
entity_pair_classification_labels = None
entity_span_classification_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
entity_classification_labels = ids_tensor([self.batch_size], self.num_entity_classification_labels)
entity_pair_classification_labels = ids_tensor(
[self.batch_size], self.num_entity_pair_classification_labels
)
entity_span_classification_labels = ids_tensor(
[self.batch_size, self.entity_length], self.num_entity_span_classification_labels
)
config = LukeConfig(
vocab_size=self.vocab_size,
entity_vocab_size=self.entity_vocab_size,
entity_emb_size=self.entity_emb_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
is_decoder=False,
initializer_range=self.initializer_range,
use_entity_aware_attention=self.use_entity_aware_attention,
)
return (
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
)
def create_and_check_model(
self,
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
):
model = LukeModel(config=config)
model.to(torch_device)
model.eval()
# test with words + entities
result = model(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
entity_ids=entity_ids,
entity_attention_mask=entity_attention_mask,
entity_token_type_ids=entity_token_type_ids,
entity_position_ids=entity_position_ids,
)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
self.parent.assertEqual(
result.entity_last_hidden_state.shape, (self.batch_size, self.entity_length, self.hidden_size)
)
# test with words only
result = model(input_ids, token_type_ids=token_type_ids)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_entity_classification(
self,
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
):
config.num_labels = self.num_entity_classification_labels
model = LukeForEntityClassification(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
entity_ids=entity_ids,
entity_attention_mask=entity_attention_mask,
entity_token_type_ids=entity_token_type_ids,
entity_position_ids=entity_position_ids,
labels=entity_classification_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_entity_classification_labels))
def create_and_check_for_entity_pair_classification(
self,
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
):
config.num_labels = self.num_entity_pair_classification_labels
model = LukeForEntityClassification(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
entity_ids=entity_ids,
entity_attention_mask=entity_attention_mask,
entity_token_type_ids=entity_token_type_ids,
entity_position_ids=entity_position_ids,
labels=entity_pair_classification_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_entity_pair_classification_labels))
def create_and_check_for_entity_span_classification(
self,
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
):
config.num_labels = self.num_entity_span_classification_labels
model = LukeForEntitySpanClassification(config)
model.to(torch_device)
model.eval()
entity_start_positions = ids_tensor([self.batch_size, self.entity_length], self.seq_length)
entity_end_positions = ids_tensor([self.batch_size, self.entity_length], self.seq_length)
result = model(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
entity_ids=entity_ids,
entity_attention_mask=entity_attention_mask,
entity_token_type_ids=entity_token_type_ids,
entity_position_ids=entity_position_ids,
entity_start_positions=entity_start_positions,
entity_end_positions=entity_end_positions,
labels=entity_span_classification_labels,
)
self.parent.assertEqual(
result.logits.shape, (self.batch_size, self.entity_length, self.num_entity_span_classification_labels)
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
attention_mask,
token_type_ids,
entity_ids,
entity_attention_mask,
entity_token_type_ids,
entity_position_ids,
sequence_labels,
entity_classification_labels,
entity_pair_classification_labels,
entity_span_classification_labels,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"token_type_ids": token_type_ids,
"attention_mask": attention_mask,
"entity_ids": entity_ids,
"entity_token_type_ids": entity_token_type_ids,
"entity_attention_mask": entity_attention_mask,
"entity_position_ids": entity_position_ids,
}
return config, inputs_dict
@require_torch
class LukeModelTest(ModelTesterMixin, unittest.TestCase):
all_model_classes = (
(
LukeModel,
LukeForEntityClassification,
LukeForEntityPairClassification,
LukeForEntitySpanClassification,
)
if is_torch_available()
else ()
)
test_pruning = False
test_torchscript = False
test_resize_embeddings = True
test_head_masking = True
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
if model_class == LukeForEntitySpanClassification:
inputs_dict["entity_start_positions"] = torch.zeros(
(self.model_tester.batch_size, self.model_tester.entity_length), dtype=torch.long, device=torch_device
)
inputs_dict["entity_end_positions"] = torch.ones(
(self.model_tester.batch_size, self.model_tester.entity_length), dtype=torch.long, device=torch_device
)
if return_labels:
if model_class in (LukeForEntityClassification, LukeForEntityPairClassification):
inputs_dict["labels"] = torch.zeros(
self.model_tester.batch_size, dtype=torch.long, device=torch_device
)
elif model_class == LukeForEntitySpanClassification:
inputs_dict["labels"] = torch.zeros(
(self.model_tester.batch_size, self.model_tester.entity_length),
dtype=torch.long,
device=torch_device,
)
return inputs_dict
def setUp(self):
self.model_tester = LukeModelTester(self)
self.config_tester = ConfigTester(self, config_class=LukeConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
@slow
def test_model_from_pretrained(self):
for model_name in LUKE_PRETRAINED_MODEL_ARCHIVE_LIST:
model = LukeModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_for_entity_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_entity_classification(*config_and_inputs)
def test_for_entity_pair_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_entity_pair_classification(*config_and_inputs)
def test_for_entity_span_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_entity_span_classification(*config_and_inputs)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
seq_length = self.model_tester.seq_length
entity_length = self.model_tester.entity_length
key_length = seq_length + entity_length
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length + entity_length, key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
added_hidden_states = self.model_tester.num_hidden_states_types
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, seq_length + entity_length, key_length],
)
def test_entity_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
entity_hidden_states = outputs.entity_hidden_states
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
self.assertEqual(len(entity_hidden_states), expected_num_layers)
entity_length = self.model_tester.entity_length
self.assertListEqual(
list(entity_hidden_states[0].shape[-2:]),
[entity_length, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
def test_retain_grad_entity_hidden_states(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
config.output_attentions = True
# no need to test all models as different heads yield the same functionality
model_class = self.all_model_classes[0]
model = model_class(config)
model.to(torch_device)
inputs = self._prepare_for_class(inputs_dict, model_class)
outputs = model(**inputs)
output = outputs[0]
entity_hidden_states = outputs.entity_hidden_states[0]
entity_hidden_states.retain_grad()
output.flatten()[0].backward(retain_graph=True)
self.assertIsNotNone(entity_hidden_states.grad)
@require_torch
class LukeModelIntegrationTests(unittest.TestCase):
@slow
def test_inference_base_model(self):
model = LukeModel.from_pretrained("studio-ousia/luke-base").eval()
model.to(torch_device)
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", task="entity_classification")
text = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck as a fortuitous netcord helped the new world number one avoid a humiliating second- round exit at Wimbledon ."
span = (39, 42)
encoding = tokenizer(text, entity_spans=[span], add_prefix_space=True, return_tensors="pt")
# move all values to device
for key, value in encoding.items():
encoding[key] = encoding[key].to(torch_device)
outputs = model(**encoding)
# Verify word hidden states
expected_shape = torch.Size((1, 42, 768))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor(
[[0.0037, 0.1368, -0.0091], [0.1099, 0.3329, -0.1095], [0.0765, 0.5335, 0.1179]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))
# Verify entity hidden states
expected_shape = torch.Size((1, 1, 768))
self.assertEqual(outputs.entity_last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor([[0.1457, 0.1044, 0.0174]])
self.assertTrue(torch.allclose(outputs.entity_last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))
@slow
def test_inference_large_model(self):
model = LukeModel.from_pretrained("studio-ousia/luke-large").eval()
model.to(torch_device)
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large", task="entity_classification")
text = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck as a fortuitous netcord helped the new world number one avoid a humiliating second- round exit at Wimbledon ."
span = (39, 42)
encoding = tokenizer(text, entity_spans=[span], add_prefix_space=True, return_tensors="pt")
# move all values to device
for key, value in encoding.items():
encoding[key] = encoding[key].to(torch_device)
outputs = model(**encoding)
# Verify word hidden states
expected_shape = torch.Size((1, 42, 1024))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor(
[[0.0133, 0.0865, 0.0095], [0.3093, -0.2576, -0.7418], [-0.1720, -0.2117, -0.2869]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))
# Verify entity hidden states
expected_shape = torch.Size((1, 1, 1024))
self.assertEqual(outputs.entity_last_hidden_state.shape, expected_shape)
expected_slice = torch.tensor([[0.0466, -0.0106, -0.0179]])
self.assertTrue(torch.allclose(outputs.entity_last_hidden_state[0, :3, :3], expected_slice, atol=1e-4))

575
tests/test_tokenization_luke.py Normal file
View File

@@ -0,0 +1,575 @@
# coding=utf-8
# Copyright 2021 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
from transformers import AddedToken, LukeTokenizer
from transformers.testing_utils import require_torch, slow
from .test_tokenization_common import TokenizerTesterMixin
class Luke(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = LukeTokenizer
from_pretrained_kwargs = {"cls_token": "<s>"}
def setUp(self):
super().setUp()
self.special_tokens_map = {"entity_token_1": "<ent>", "entity_token_2": "<ent2>"}
def get_tokenizer(self, task=None, **kwargs):
kwargs.update(self.special_tokens_map)
return self.tokenizer_class.from_pretrained("studio-ousia/luke-base", task=task, **kwargs)
def get_input_output_texts(self, tokenizer):
input_text = "lower newer"
output_text = "lower newer"
return input_text, output_text
def test_full_tokenizer(self):
tokenizer = self.tokenizer_class.from_pretrained("studio-ousia/luke-base")
text = "lower newer"
bpe_tokens = ["lower", "\u0120newer"]
tokens = tokenizer.tokenize(text) # , add_prefix_space=True)
self.assertListEqual(tokens, bpe_tokens)
input_tokens = tokens + [tokenizer.unk_token]
input_bpe_tokens = [29668, 13964, 3]
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
def luke_dict_integration_testing(self):
tokenizer = self.get_tokenizer()
self.assertListEqual(tokenizer.encode("Hello world!", add_special_tokens=False), [0, 31414, 232, 328, 2])
self.assertListEqual(
tokenizer.encode("Hello world! cécé herlolip 418", add_special_tokens=False),
[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2],
)
@slow
def test_sequence_builders(self):
tokenizer = self.tokenizer_class.from_pretrained("studio-ousia/luke-large")
text = tokenizer.encode("sequence builders", add_special_tokens=False)
text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
encoded_text_from_decode = tokenizer.encode(
"sequence builders", add_special_tokens=True, add_prefix_space=False
)
encoded_pair_from_decode = tokenizer.encode(
"sequence builders", "multi-sequence build", add_special_tokens=True, add_prefix_space=False
)
encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
assert encoded_sentence == encoded_text_from_decode
assert encoded_pair == encoded_pair_from_decode
def test_space_encoding(self):
tokenizer = self.get_tokenizer()
sequence = "Encode this sequence."
space_encoding = tokenizer.byte_encoder[" ".encode("utf-8")[0]]
# Testing encoder arguments
encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=False)
first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0]
self.assertNotEqual(first_char, space_encoding)
encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=True)
first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0]
self.assertEqual(first_char, space_encoding)
tokenizer.add_special_tokens({"bos_token": "<s>"})
encoded = tokenizer.encode(sequence, add_special_tokens=True)
first_char = tokenizer.convert_ids_to_tokens(encoded[1])[0]
self.assertNotEqual(first_char, space_encoding)
# Testing spaces after special tokens
mask = "<mask>"
tokenizer.add_special_tokens(
{"mask_token": AddedToken(mask, lstrip=True, rstrip=False)}
) # mask token has a left space
mask_ind = tokenizer.convert_tokens_to_ids(mask)
sequence = "Encode <mask> sequence"
sequence_nospace = "Encode <mask>sequence"
encoded = tokenizer.encode(sequence)
mask_loc = encoded.index(mask_ind)
first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0]
self.assertEqual(first_char, space_encoding)
encoded = tokenizer.encode(sequence_nospace)
mask_loc = encoded.index(mask_ind)
first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0]
self.assertNotEqual(first_char, space_encoding)
def test_pretokenized_inputs(self):
pass
def test_embeded_special_tokens(self):
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest("{} ({})".format(tokenizer.__class__.__name__, pretrained_name)):
tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
sentence = "A, <mask> AllenNLP sentence."
tokens_r = tokenizer_r.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True)
tokens_p = tokenizer_p.encode_plus(sentence, add_special_tokens=True, return_token_type_ids=True)
# token_type_ids should put 0 everywhere
self.assertEqual(sum(tokens_r["token_type_ids"]), sum(tokens_p["token_type_ids"]))
# token_type_ids should put 0 everywhere
self.assertEqual(sum(tokens_r["token_type_ids"]), sum(tokens_p["token_type_ids"]))
# attention_mask should put 1 everywhere, so sum over length should be 1
self.assertEqual(
sum(tokens_p["attention_mask"]) / len(tokens_p["attention_mask"]),
)
tokens_p_str = tokenizer_p.convert_ids_to_tokens(tokens_p["input_ids"])
# Rust correctly handles the space before the mask while python doesnt
self.assertSequenceEqual(tokens_p["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2])
self.assertSequenceEqual(
tokens_p_str, ["<s>", "A", ",", "<mask>", "ĠAllen", "N", "LP", "Ġsentence", ".", "</s>"]
)
@require_torch
class LukeTokenizerIntegrationTests(unittest.TestCase):
tokenizer_class = LukeTokenizer
from_pretrained_kwargs = {"cls_token": "<s>"}
def setUp(self):
super().setUp()
def test_single_text_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
entities = ["Ana Ivanovic", "Thursday", "Dummy Entity"]
spans = [(9, 21), (30, 38), (39, 42)]
encoding = tokenizer(sentence, entities=entities, entity_spans=spans, return_token_type_ids=True)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday she could hardly believe her luck.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][3:6], spaces_between_special_tokens=False), " Ana Ivanovic"
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][8:9], spaces_between_special_tokens=False), " Thursday"
)
self.assertEqual(tokenizer.decode(encoding["input_ids"][9:10], spaces_between_special_tokens=False), " she")
self.assertEqual(
encoding["entity_ids"],
[
tokenizer.entity_vocab["Ana Ivanovic"],
tokenizer.entity_vocab["Thursday"],
tokenizer.entity_vocab["[UNK]"],
],
)
self.assertEqual(encoding["entity_attention_mask"], [1, 1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[3, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[8, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[9, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
]
)
# fmt: on
def test_single_text_only_entity_spans_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
spans = [(9, 21), (30, 38), (39, 42)]
encoding = tokenizer(sentence, entity_spans=spans, return_token_type_ids=True)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday she could hardly believe her luck.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][3:6], spaces_between_special_tokens=False), " Ana Ivanovic"
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][8:9], spaces_between_special_tokens=False), " Thursday"
)
self.assertEqual(tokenizer.decode(encoding["input_ids"][9:10], spaces_between_special_tokens=False), " she")
mask_id = tokenizer.entity_vocab["[MASK]"]
self.assertEqual(encoding["entity_ids"], [mask_id, mask_id, mask_id])
self.assertEqual(encoding["entity_attention_mask"], [1, 1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[3, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[8, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, ],
[9, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, ]
]
)
# fmt: on
def test_single_text_padding_pytorch_tensors(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
entities = ["Ana Ivanovic", "Thursday", "Dummy Entity"]
spans = [(9, 21), (30, 38), (39, 42)]
encoding = tokenizer(
sentence,
entities=entities,
entity_spans=spans,
return_token_type_ids=True,
padding="max_length",
max_length=30,
max_entity_length=16,
return_tensors="pt",
)
# test words
self.assertEqual(encoding["input_ids"].shape, (1, 30))
self.assertEqual(encoding["attention_mask"].shape, (1, 30))
self.assertEqual(encoding["token_type_ids"].shape, (1, 30))
# test entities
self.assertEqual(encoding["entity_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_attention_mask"].shape, (1, 16))
self.assertEqual(encoding["entity_token_type_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_position_ids"].shape, (1, 16, tokenizer.max_mention_length))
def test_text_pair_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday"
sentence_pair = "She could hardly believe her luck."
entities = ["Ana Ivanovic", "Thursday"]
entities_pair = ["Dummy Entity"]
spans = [(9, 21), (30, 38)]
spans_pair = [(0, 3)]
encoding = tokenizer(
sentence,
sentence_pair,
entities=entities,
entities_pair=entities_pair,
entity_spans=spans,
entity_spans_pair=spans_pair,
return_token_type_ids=True,
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday</s></s>She could hardly believe her luck.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][3:6], spaces_between_special_tokens=False), " Ana Ivanovic"
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][8:9], spaces_between_special_tokens=False), " Thursday"
)
self.assertEqual(tokenizer.decode(encoding["input_ids"][11:12], spaces_between_special_tokens=False), "She")
self.assertEqual(
encoding["entity_ids"],
[
tokenizer.entity_vocab["Ana Ivanovic"],
tokenizer.entity_vocab["Thursday"],
tokenizer.entity_vocab["[UNK]"],
],
)
self.assertEqual(encoding["entity_attention_mask"], [1, 1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[3, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[8, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[11, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
]
)
# fmt: on
def test_text_pair_only_entity_spans_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday"
sentence_pair = "She could hardly believe her luck."
spans = [(9, 21), (30, 38)]
spans_pair = [(0, 3)]
encoding = tokenizer(
sentence,
sentence_pair,
entity_spans=spans,
entity_spans_pair=spans_pair,
return_token_type_ids=True,
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday</s></s>She could hardly believe her luck.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][3:6], spaces_between_special_tokens=False), " Ana Ivanovic"
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][8:9], spaces_between_special_tokens=False), " Thursday"
)
self.assertEqual(tokenizer.decode(encoding["input_ids"][11:12], spaces_between_special_tokens=False), "She")
mask_id = tokenizer.entity_vocab["[MASK]"]
self.assertEqual(encoding["entity_ids"], [mask_id, mask_id, mask_id])
self.assertEqual(encoding["entity_attention_mask"], [1, 1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[3, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[8, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[11, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
]
)
# fmt: on
def test_text_pair_padding_pytorch_tensors(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", return_token_type_ids=True)
sentence = "Top seed Ana Ivanovic said on Thursday"
sentence_pair = "She could hardly believe her luck."
entities = ["Ana Ivanovic", "Thursday"]
entities_pair = ["Dummy Entity"]
spans = [(9, 21), (30, 38)]
spans_pair = [(0, 3)]
encoding = tokenizer(
sentence,
sentence_pair,
entities=entities,
entities_pair=entities_pair,
entity_spans=spans,
entity_spans_pair=spans_pair,
return_token_type_ids=True,
padding="max_length",
max_length=30,
max_entity_length=16,
return_tensors="pt",
)
# test words
self.assertEqual(encoding["input_ids"].shape, (1, 30))
self.assertEqual(encoding["attention_mask"].shape, (1, 30))
self.assertEqual(encoding["token_type_ids"].shape, (1, 30))
# test entities
self.assertEqual(encoding["entity_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_attention_mask"].shape, (1, 16))
self.assertEqual(encoding["entity_token_type_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_position_ids"].shape, (1, 16, tokenizer.max_mention_length))
def test_entity_classification_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base", task="entity_classification")
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck as a fortuitous netcord helped the new world number one avoid a humiliating second- round exit at Wimbledon ."
span = (39, 42)
encoding = tokenizer(sentence, entity_spans=[span], return_token_type_ids=True)
# test words
self.assertEqual(len(encoding["input_ids"]), 42)
self.assertEqual(len(encoding["attention_mask"]), 42)
self.assertEqual(len(encoding["token_type_ids"]), 42)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday<ent> she<ent> could hardly believe her luck as a fortuitous netcord helped the new world number one avoid a humiliating second- round exit at Wimbledon.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][9:12], spaces_between_special_tokens=False), "<ent> she<ent>"
)
# test entities
self.assertEqual(encoding["entity_ids"], [2])
self.assertEqual(encoding["entity_attention_mask"], [1])
self.assertEqual(encoding["entity_token_type_ids"], [0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[9, 10, 11, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]
]
)
# fmt: on
def test_entity_classification_padding_pytorch_tensors(self):
tokenizer = LukeTokenizer.from_pretrained(
"studio-ousia/luke-base", task="entity_classification", return_token_type_ids=True
)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck as a fortuitous netcord helped the new world number one avoid a humiliating second- round exit at Wimbledon ."
# entity information
span = (39, 42)
encoding = tokenizer(
sentence, entity_spans=[span], return_token_type_ids=True, padding="max_length", return_tensors="pt"
)
# test words
self.assertEqual(encoding["input_ids"].shape, (1, 512))
self.assertEqual(encoding["attention_mask"].shape, (1, 512))
self.assertEqual(encoding["token_type_ids"].shape, (1, 512))
# test entities
self.assertEqual(encoding["entity_ids"].shape, (1, 1))
self.assertEqual(encoding["entity_attention_mask"].shape, (1, 1))
self.assertEqual(encoding["entity_token_type_ids"].shape, (1, 1))
self.assertEqual(
encoding["entity_position_ids"].shape, (1, tokenizer.max_entity_length, tokenizer.max_mention_length)
)
def test_entity_pair_classification_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained(
"studio-ousia/luke-base", task="entity_pair_classification", return_token_type_ids=True
)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
# head and tail information
spans = [(9, 21), (39, 42)]
encoding = tokenizer(sentence, entity_spans=spans, return_token_type_ids=True)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed<ent> Ana Ivanovic<ent> said on Thursday<ent2> she<ent2> could hardly believe her luck.</s>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][3:8], spaces_between_special_tokens=False),
"<ent> Ana Ivanovic<ent>",
)
self.assertEqual(
tokenizer.decode(encoding["input_ids"][11:14], spaces_between_special_tokens=False), "<ent2> she<ent2>"
)
self.assertEqual(encoding["entity_ids"], [2, 3])
self.assertEqual(encoding["entity_attention_mask"], [1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[3, 4, 5, 6, 7, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[11, 12, 13, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
]
)
# fmt: on
def test_entity_pair_classification_padding_pytorch_tensors(self):
tokenizer = LukeTokenizer.from_pretrained(
"studio-ousia/luke-base", task="entity_pair_classification", return_token_type_ids=True
)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
# head and tail information
spans = [(9, 21), (39, 42)]
encoding = tokenizer(
sentence,
entity_spans=spans,
return_token_type_ids=True,
padding="max_length",
max_length=30,
return_tensors="pt",
)
# test words
self.assertEqual(encoding["input_ids"].shape, (1, 30))
self.assertEqual(encoding["attention_mask"].shape, (1, 30))
self.assertEqual(encoding["token_type_ids"].shape, (1, 30))
# test entities
self.assertEqual(encoding["entity_ids"].shape, (1, 2))
self.assertEqual(encoding["entity_attention_mask"].shape, (1, 2))
self.assertEqual(encoding["entity_token_type_ids"].shape, (1, 2))
self.assertEqual(
encoding["entity_position_ids"].shape, (1, tokenizer.max_entity_length, tokenizer.max_mention_length)
)
def test_entity_span_classification_no_padding_or_truncation(self):
tokenizer = LukeTokenizer.from_pretrained(
"studio-ousia/luke-base", task="entity_span_classification", return_token_type_ids=True
)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
spans = [(0, 8), (9, 21), (39, 42)]
encoding = tokenizer(sentence, entity_spans=spans, return_token_type_ids=True)
self.assertEqual(
tokenizer.decode(encoding["input_ids"], spaces_between_special_tokens=False),
"<s>Top seed Ana Ivanovic said on Thursday she could hardly believe her luck.</s>",
)
self.assertEqual(encoding["entity_ids"], [2, 2, 2])
self.assertEqual(encoding["entity_attention_mask"], [1, 1, 1])
self.assertEqual(encoding["entity_token_type_ids"], [0, 0, 0])
# fmt: off
self.assertEqual(
encoding["entity_position_ids"],
[
[1, 2, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[3, 4, 5, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
[9, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
]
)
# fmt: on
self.assertEqual(encoding["entity_start_positions"], [1, 3, 9])
self.assertEqual(encoding["entity_end_positions"], [2, 5, 9])
def test_entity_span_classification_padding_pytorch_tensors(self):
tokenizer = LukeTokenizer.from_pretrained(
"studio-ousia/luke-base", task="entity_span_classification", return_token_type_ids=True
)
sentence = "Top seed Ana Ivanovic said on Thursday she could hardly believe her luck."
spans = [(0, 8), (9, 21), (39, 42)]
encoding = tokenizer(
sentence,
entity_spans=spans,
return_token_type_ids=True,
padding="max_length",
max_length=30,
max_entity_length=16,
return_tensors="pt",
)
# test words
self.assertEqual(encoding["input_ids"].shape, (1, 30))
self.assertEqual(encoding["attention_mask"].shape, (1, 30))
self.assertEqual(encoding["token_type_ids"].shape, (1, 30))
# test entities
self.assertEqual(encoding["entity_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_attention_mask"].shape, (1, 16))
self.assertEqual(encoding["entity_token_type_ids"].shape, (1, 16))
self.assertEqual(encoding["entity_position_ids"].shape, (1, 16, tokenizer.max_mention_length))
self.assertEqual(encoding["entity_start_positions"].shape, (1, 16))
self.assertEqual(encoding["entity_end_positions"].shape, (1, 16))