Skip to content

GitLab

Unverified Commit 5daca95d authored by Thomas Wolf's avatar Thomas Wolf Committed by GitHub
Browse files

Merge pull request #2268 from aaugustin/improve-repository-structure 

Improve repository structure
parents 54abc67a 00204f2b
Hide whitespace changes
Inline Side-by-side
Showing with 4 additions and 1829 deletions
transformers/tests/tokenization_utils_test.py → tests/test_tokenization_utils.py
View file @ 5daca95d
......@@ -44,7 +44,3 @@ class TokenizerUtilsTest(unittest.TestCase):
@slow
def test_pretrained_tokenizers(self):
self.check_tokenizer_from_pretrained(GPT2Tokenizer)
if __name__ == "__main__":
unittest.main()
transformers/tests/tokenization_xlm_test.py → tests/test_tokenization_xlm.py
View file @ 5daca95d
......@@ -20,11 +20,11 @@ import unittest
from transformers.tokenization_xlm import VOCAB_FILES_NAMES, XLMTokenizer
from .tokenization_tests_commons import CommonTestCases
from .test_tokenization_common import TokenizerTesterMixin
from .utils import slow
class XLMTokenizationTest(CommonTestCases.CommonTokenizerTester):
class XLMTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = XLMTokenizer
......@@ -98,7 +98,3 @@ class XLMTokenizationTest(CommonTestCases.CommonTokenizerTester):
assert encoded_sentence == [1] + text + [1]
assert encoded_pair == [1] + text + [1] + text_2 + [1]
if __name__ == "__main__":
unittest.main()
transformers/tests/tokenization_xlnet_test.py → tests/test_tokenization_xlnet.py
View file @ 5daca95d
......@@ -19,14 +19,14 @@ import unittest
from transformers.tokenization_xlnet import SPIECE_UNDERLINE, XLNetTokenizer
from .tokenization_tests_commons import CommonTestCases
from .test_tokenization_common import TokenizerTesterMixin
from .utils import slow
SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "fixtures/test_sentencepiece.model")
class XLNetTokenizationTest(CommonTestCases.CommonTokenizerTester):
class XLNetTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
tokenizer_class = XLNetTokenizer
......@@ -183,7 +183,3 @@ class XLNetTokenizationTest(CommonTestCases.CommonTokenizerTester):
assert encoded_sentence == text + [4, 3]
assert encoded_pair == text + [4] + text_2 + [4, 3]
if __name__ == "__main__":
unittest.main()
transformers/tests/modeling_common_test.py deleted 100644 → 0
View file @ 54abc67a
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# 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.
from __future__ import absolute_import, division, print_function
import copy
import json
import logging
import os.path
import random
import shutil
import sys
import tempfile
import unittest
import uuid
from transformers import is_torch_available
from .utils import CACHE_DIR, require_torch, slow, torch_device
if is_torch_available():
import torch
import numpy as np
from transformers import (
AdaptiveEmbedding,
PretrainedConfig,
PreTrainedModel,
BertModel,
BertConfig,
BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
)
if sys.version_info[0] == 2:
class TemporaryDirectory(object):
"""Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
def __enter__(self):
self.name = tempfile.mkdtemp()
return self.name
def __exit__(self, exc_type, exc_value, traceback):
shutil.rmtree(self.name)
else:
TemporaryDirectory = tempfile.TemporaryDirectory
unicode = str
def _config_zero_init(config):
configs_no_init = copy.deepcopy(config)
for key in configs_no_init.__dict__.keys():
if "_range" in key or "_std" in key or "initializer_factor" in key:
setattr(configs_no_init, key, 0.0)
return configs_no_init
class CommonTestCases:
@require_torch
class CommonModelTester(unittest.TestCase):
model_tester = None
all_model_classes = ()
test_torchscript = True
test_pruning = True
test_resize_embeddings = True
test_head_masking = True
is_encoder_decoder = False
def test_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
out_2 = outputs[0].numpy()
out_2[np.isnan(out_2)] = 0
with TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
model.to(torch_device)
with torch.no_grad():
after_outputs = model(**inputs_dict)
# Make sure we don't have nans
out_1 = after_outputs[0].cpu().numpy()
out_1[np.isnan(out_1)] = 0
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad:
self.assertIn(
param.data.mean().item(),
[0.0, 1.0],
msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
)
def test_determinism(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
first = model(**inputs_dict)[0]
second = model(**inputs_dict)[0]
out_1 = first.cpu().numpy()
out_2 = second.cpu().numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
decoder_seq_length = (
self.model_tester.decoder_seq_length
if hasattr(self.model_tester, "decoder_seq_length")
else self.model_tester.seq_length
)
encoder_seq_length = (
self.model_tester.encoder_seq_length
if hasattr(self.model_tester, "encoder_seq_length")
else self.model_tester.seq_length
)
decoder_key_length = (
self.model_tester.key_length if hasattr(self.model_tester, "key_length") else decoder_seq_length
)
encoder_key_length = (
self.model_tester.key_length if hasattr(self.model_tester, "key_length") else encoder_seq_length
)
for model_class in self.all_model_classes:
config.output_attentions = True
config.output_hidden_states = False
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
if self.is_encoder_decoder:
self.assertEqual(out_len % 2, 0)
decoder_attentions = outputs[(out_len // 2) - 1]
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
# Check attention is always last and order is fine
config.output_attentions = True
config.output_hidden_states = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, True)
self_attentions = outputs[-1]
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, encoder_seq_length, encoder_key_length],
)
def test_torchscript(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
self._create_and_check_torchscript(config, inputs_dict)
def test_torchscript_output_attentions(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_attentions = True
self._create_and_check_torchscript(config, inputs_dict)
def test_torchscript_output_hidden_state(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
self._create_and_check_torchscript(config, inputs_dict)
def _create_and_check_torchscript(self, config, inputs_dict):
if not self.test_torchscript:
return
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
configs_no_init.torchscript = True
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
inputs = inputs_dict["input_ids"] # Let's keep only input_ids
try:
traced_gpt2 = torch.jit.trace(model, inputs)
except RuntimeError:
self.fail("Couldn't trace module.")
with TemporaryDirectory() as tmp_dir_name:
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
try:
torch.jit.save(traced_gpt2, pt_file_name)
except Exception:
self.fail("Couldn't save module.")
try:
loaded_model = torch.jit.load(pt_file_name)
except Exception:
self.fail("Couldn't load module.")
model.to(torch_device)
model.eval()
loaded_model.to(torch_device)
loaded_model.eval()
model_params = model.parameters()
loaded_model_params = loaded_model.parameters()
models_equal = True
for p1, p2 in zip(model_params, loaded_model_params):
if p1.data.ne(p2.data).sum() > 0:
models_equal = False
self.assertTrue(models_equal)
def test_headmasking(self):
if not self.test_head_masking:
return
global_rng.seed(42)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
global_rng.seed()
config.output_attentions = True
config.output_hidden_states = True
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
# Prepare head_mask
# Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
head_mask = torch.ones(
self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads, device=torch_device
)
head_mask[0, 0] = 0
head_mask[-1, :-1] = 0
head_mask.requires_grad_(requires_grad=True)
inputs = inputs_dict.copy()
inputs["head_mask"] = head_mask
outputs = model(**inputs)
# Test that we can get a gradient back for importance score computation
output = sum(t.sum() for t in outputs[0])
output = output.sum()
output.backward()
multihead_outputs = head_mask.grad
attentions = outputs[-1]
hidden_states = outputs[-2]
# Remove Nan
for t in attentions:
self.assertLess(
torch.sum(torch.isnan(t)), t.numel() / 4
) # Check we don't have more than 25% nans (arbitrary)
attentions = [
t.masked_fill(torch.isnan(t), 0.0) for t in attentions
] # remove them (the test is less complete)
self.assertIsNotNone(multihead_outputs)
self.assertEqual(len(multihead_outputs), self.model_tester.num_hidden_layers)
self.assertAlmostEqual(attentions[0][..., 0, :, :].flatten().sum().item(), 0.0)
self.assertNotEqual(attentions[0][..., -1, :, :].flatten().sum().item(), 0.0)
self.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)
self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
def test_head_pruning(self):
if not self.test_pruning:
return
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if "head_mask" in inputs_dict:
del inputs_dict["head_mask"]
config.output_attentions = True
config.output_hidden_states = False
model = model_class(config=config)
model.to(torch_device)
model.eval()
heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
model.prune_heads(heads_to_prune)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def test_head_pruning_save_load_from_pretrained(self):
if not self.test_pruning:
return
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if "head_mask" in inputs_dict:
del inputs_dict["head_mask"]
config.output_attentions = True
config.output_hidden_states = False
model = model_class(config=config)
model.to(torch_device)
model.eval()
heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
model.prune_heads(heads_to_prune)
with TemporaryDirectory() as temp_dir_name:
model.save_pretrained(temp_dir_name)
model = model_class.from_pretrained(temp_dir_name)
model.to(torch_device)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def test_head_pruning_save_load_from_config_init(self):
if not self.test_pruning:
return
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if "head_mask" in inputs_dict:
del inputs_dict["head_mask"]
config.output_attentions = True
config.output_hidden_states = False
heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
config.pruned_heads = heads_to_prune
model = model_class(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def test_head_pruning_integration(self):
if not self.test_pruning:
return
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if "head_mask" in inputs_dict:
del inputs_dict["head_mask"]
config.output_attentions = True
config.output_hidden_states = False
heads_to_prune = {0: [0], 1: [1, 2]}
config.pruned_heads = heads_to_prune
model = model_class(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
with TemporaryDirectory() as temp_dir_name:
model.save_pretrained(temp_dir_name)
model = model_class.from_pretrained(temp_dir_name)
model.to(torch_device)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
heads_to_prune = {0: [0], 2: [1, 2]}
model.prune_heads(heads_to_prune)
with torch.no_grad():
outputs = model(**inputs_dict)
attentions = outputs[-1]
self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
config.output_hidden_states = True
config.output_attentions = False
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**inputs_dict)
hidden_states = outputs[-1]
self.assertEqual(model.config.output_attentions, False)
self.assertEqual(model.config.output_hidden_states, True)
self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[
self.model_tester.encoder_seq_length
if hasattr(self.model_tester, "encoder_seq_length")
else self.model_tester.seq_length,
self.model_tester.hidden_size,
],
)
def test_resize_tokens_embeddings(self):
original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.test_resize_embeddings:
return
for model_class in self.all_model_classes:
config = copy.deepcopy(original_config)
model = model_class(config)
model_vocab_size = config.vocab_size
# Retrieve the embeddings and clone theme
model_embed = model.resize_token_embeddings(model_vocab_size)
cloned_embeddings = model_embed.weight.clone()
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
model_embed = model.resize_token_embeddings(model_vocab_size + 10)
self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
# Check that it actually resizes the embeddings matrix
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
model_embed = 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
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
# Check that adding and removing tokens has not modified the first part of the embedding matrix.
models_equal = True
for p1, p2 in zip(cloned_embeddings, model_embed.weight):
if p1.data.ne(p2.data).sum() > 0:
models_equal = False
self.assertTrue(models_equal)
def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
self.assertIsInstance(model.get_input_embeddings(), (torch.nn.Embedding, AdaptiveEmbedding))
model.set_input_embeddings(torch.nn.Embedding(10, 10))
x = model.get_output_embeddings()
self.assertTrue(x is None or isinstance(x, torch.nn.Linear))
def test_tie_model_weights(self):
if not self.test_torchscript:
return
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def check_same_values(layer_1, layer_2):
equal = True
for p1, p2 in zip(layer_1.weight, layer_2.weight):
if p1.data.ne(p2.data).sum() > 0:
equal = False
return equal
for model_class in self.all_model_classes:
config.torchscript = True
model_not_tied = model_class(config)
if model_not_tied.get_output_embeddings() is None:
continue
params_not_tied = list(model_not_tied.parameters())
config_tied = copy.deepcopy(config)
config_tied.torchscript = False
model_tied = model_class(config_tied)
params_tied = list(model_tied.parameters())
# Check that the embedding layer and decoding layer are the same in size and in value
self.assertGreater(len(params_not_tied), len(params_tied))
# self.assertTrue(check_same_values(embeddings, decoding))
# # Check that after modification, they remain the same.
# embeddings.weight.data.div_(2)
# # Check that the embedding layer and decoding layer are the same in size and in value
# self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
# self.assertTrue(check_same_values(embeddings, decoding))
# # Check that after modification, they remain the same.
# decoding.weight.data.div_(4)
# # Check that the embedding layer and decoding layer are the same in size and in value
# self.assertTrue(embeddings.weight.shape, decoding.weight.shape)
# self.assertTrue(check_same_values(embeddings, decoding))
# Check that after resize they remain tied.
model_tied.resize_token_embeddings(config.vocab_size + 10)
params_tied_2 = list(model_tied.parameters())
self.assertGreater(len(params_not_tied), len(params_tied))
self.assertEqual(len(params_tied_2), len(params_tied))
# decoding.weight.data.mul_(20)
# # Check that the embedding layer and decoding layer are the same in size and in value
# self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
# self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.is_encoder_decoder:
input_ids = inputs_dict["input_ids"]
del inputs_dict["input_ids"]
else:
encoder_input_ids = inputs_dict["encoder_input_ids"]
decoder_input_ids = inputs_dict["decoder_input_ids"]
del inputs_dict["encoder_input_ids"]
del inputs_dict["decoder_input_ids"]
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
inputs_dict["inputs_embeds"] = wte(input_ids)
else:
inputs_dict["encoder_inputs_embeds"] = wte(encoder_input_ids)
inputs_dict["decoder_inputs_embeds"] = wte(decoder_input_ids)
with torch.no_grad():
outputs = model(**inputs_dict)
class GPTModelTester(CommonModelTester):
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_position_ids=True,
use_token_type_ids=True,
use_labels=True,
vocab_size=99,
n_positions=33,
hidden_size=32,
num_hidden_layers=5,
num_attention_heads=4,
n_choices=3,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
scope=None,
config_class=None,
base_model_class=None,
lm_head_model_class=None,
double_head_model_class=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_position_ids = use_position_ids
self.use_token_type_ids = use_token_type_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.n_positions = n_positions
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.n_choices = n_choices
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.scope = scope
self.config_class = config_class
self.base_model_class = base_model_class
self.lm_head_model_class = lm_head_model_class
self.double_head_model_class = double_head_model_class
self.all_model_classes = (base_model_class, lm_head_model_class, double_head_model_class)
def prepare_config_and_inputs(self):
total_num_tokens = self.vocab_size
input_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_num_tokens)
position_ids = None
if self.use_position_ids:
position_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.n_positions)
token_type_ids = None
if self.use_token_type_ids:
total_voc = self.vocab_size
token_type_ids = ids_tensor([self.batch_size, self.n_choices, self.seq_length], total_voc)
mc_labels = None
lm_labels = None
mc_token_ids = None
if self.use_labels:
mc_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
lm_labels = ids_tensor([self.batch_size, self.n_choices, self.seq_length], self.num_labels)
mc_token_ids = ids_tensor([self.batch_size, self.n_choices], self.seq_length)
config = self.config_class(
vocab_size=self.vocab_size,
n_positions=self.n_positions,
n_embd=self.hidden_size,
n_layer=self.num_hidden_layers,
n_head=self.num_attention_heads,
initializer_range=self.initializer_range,
)
return (config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids)
def create_and_check_base_model(
self, config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids
):
model = self.base_model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids, position_ids, token_type_ids)
outputs = model(input_ids, position_ids)
outputs = model(input_ids)
hidden_state = outputs[0]
self.parent.assertListEqual(
list(hidden_state.size()), [self.batch_size, self.n_choices, self.seq_length, self.hidden_size]
)
def create_and_check_lm_head(
self, config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids
):
model = self.lm_head_model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
loss, lm_logits = outputs[:2]
total_voc = self.vocab_size
self.parent.assertListEqual(
list(lm_logits.size()), [self.batch_size, self.n_choices, self.seq_length, total_voc]
)
self.parent.assertListEqual(list(loss.size()), [])
def create_and_check_presents(
self, config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids
):
for model_class in self.all_model_classes:
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(input_ids)
presents = outputs[-1]
self.parent.assertEqual(self.num_hidden_layers, len(presents))
self.parent.assertListEqual(
list(presents[0].size()),
[
2,
self.batch_size * self.n_choices,
self.num_attention_heads,
self.seq_length,
self.hidden_size // self.num_attention_heads,
],
)
def create_and_check_double_heads(
self, config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids
):
model = self.double_head_model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(
input_ids,
mc_token_ids,
lm_labels=lm_labels,
mc_labels=mc_labels,
token_type_ids=token_type_ids,
position_ids=position_ids,
)
lm_loss, mc_loss, lm_logits, mc_logits = outputs[:4]
loss = [lm_loss, mc_loss]
total_voc = self.vocab_size
self.parent.assertListEqual(
list(lm_logits.size()), [self.batch_size, self.n_choices, self.seq_length, total_voc]
)
self.parent.assertListEqual(list(mc_logits.size()), [self.batch_size, self.n_choices])
self.parent.assertListEqual([list(l.size()) for l in loss], [[], []])
def create_and_check_model_from_pretrained(self):
for model_name in list(self.base_model_class.pretrained_model_archive_map.keys())[:1]:
model = self.base_model_class.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.parent.assertIsNotNone(model)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(config, input_ids, token_type_ids, position_ids, mc_labels, lm_labels, mc_token_ids) = config_and_inputs
inputs_dict = {"input_ids": input_ids}
return config, inputs_dict
def run_common_tests(self, test_presents=False):
config_and_inputs = self.prepare_config_and_inputs()
self.create_and_check_base_model(*config_and_inputs)
config_and_inputs = self.prepare_config_and_inputs()
self.create_and_check_lm_head(*config_and_inputs)
config_and_inputs = self.prepare_config_and_inputs()
self.create_and_check_double_heads(*config_and_inputs)
if test_presents:
config_and_inputs = self.prepare_config_and_inputs()
self.create_and_check_presents(*config_and_inputs)
@slow
def run_slow_tests(self):
self.create_and_check_model_from_pretrained()
class ConfigTester(object):
def __init__(self, parent, config_class=None, **kwargs):
self.parent = parent
self.config_class = config_class
self.inputs_dict = kwargs
def create_and_test_config_common_properties(self):
config = self.config_class(**self.inputs_dict)
self.parent.assertTrue(hasattr(config, "vocab_size"))
self.parent.assertTrue(hasattr(config, "hidden_size"))
self.parent.assertTrue(hasattr(config, "num_attention_heads"))
self.parent.assertTrue(hasattr(config, "num_hidden_layers"))
def create_and_test_config_to_json_string(self):
config = self.config_class(**self.inputs_dict)
obj = json.loads(config.to_json_string())
for key, value in self.inputs_dict.items():
self.parent.assertEqual(obj[key], value)
def create_and_test_config_to_json_file(self):
config_first = self.config_class(**self.inputs_dict)
json_file_path = os.path.join(os.getcwd(), "config_" + str(uuid.uuid4()) + ".json")
config_first.to_json_file(json_file_path)
config_second = self.config_class.from_json_file(json_file_path)
os.remove(json_file_path)
self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
def run_common_tests(self):
self.create_and_test_config_common_properties()
self.create_and_test_config_to_json_string()
self.create_and_test_config_to_json_file()
global_rng = random.Random()
def ids_tensor(shape, vocab_size, rng=None, name=None):
"""Creates a random int32 tensor of the shape within the vocab size."""
if rng is None:
rng = global_rng
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(rng.randint(0, vocab_size - 1))
return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
def floats_tensor(shape, scale=1.0, rng=None, name=None):
"""Creates a random float32 tensor of the shape within the vocab size."""
if rng is None:
rng = global_rng
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(rng.random() * scale)
return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
@require_torch
class ModelUtilsTest(unittest.TestCase):
@slow
def test_model_from_pretrained(self):
logging.basicConfig(level=logging.INFO)
for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
config = BertConfig.from_pretrained(model_name)
self.assertIsNotNone(config)
self.assertIsInstance(config, PretrainedConfig)
model = BertModel.from_pretrained(model_name)
model, loading_info = BertModel.from_pretrained(model_name, output_loading_info=True)
self.assertIsNotNone(model)
self.assertIsInstance(model, PreTrainedModel)
for value in loading_info.values():
self.assertEqual(len(value), 0)
config = BertConfig.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
model = BertModel.from_pretrained(model_name, output_attentions=True, output_hidden_states=True)
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, True)
self.assertEqual(model.config, config)
if __name__ == "__main__":
unittest.main()
transformers/tests/modeling_tf_common_test.py deleted 100644 → 0
View file @ 54abc67a
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# 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.
from __future__ import absolute_import, division, print_function
import copy
import os
import random
import shutil
import sys
import tempfile
import unittest
from transformers import is_tf_available, is_torch_available
from .utils import require_tf
if is_tf_available():
import tensorflow as tf
import numpy as np
# from transformers.modeling_bert import BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
if sys.version_info[0] == 2:
class TemporaryDirectory(object):
"""Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
def __enter__(self):
self.name = tempfile.mkdtemp()
return self.name
def __exit__(self, exc_type, exc_value, traceback):
shutil.rmtree(self.name)
else:
TemporaryDirectory = tempfile.TemporaryDirectory
unicode = str
def _config_zero_init(config):
configs_no_init = copy.deepcopy(config)
for key in configs_no_init.__dict__.keys():
if "_range" in key or "_std" in key:
setattr(configs_no_init, key, 0.0)
return configs_no_init
class TFCommonTestCases:
@require_tf
class TFCommonModelTester(unittest.TestCase):
model_tester = None
all_model_classes = ()
test_torchscript = True
test_pruning = True
test_resize_embeddings = True
is_encoder_decoder = False
def test_initialization(self):
pass
# config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# configs_no_init = _config_zero_init(config)
# for model_class in self.all_model_classes:
# model = model_class(config=configs_no_init)
# for name, param in model.named_parameters():
# if param.requires_grad:
# self.assertIn(param.data.mean().item(), [0.0, 1.0],
# msg="Parameter {} of model {} seems not properly initialized".format(name, model_class))
def test_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
outputs = model(inputs_dict)
with TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
after_outputs = model(inputs_dict)
# Make sure we don't have nans
out_1 = after_outputs[0].numpy()
out_2 = outputs[0].numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_pt_tf_model_equivalence(self):
if not is_torch_available():
return
import torch
import transformers
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
pt_model_class_name = model_class.__name__[2:] # Skip the "TF" at the beggining
pt_model_class = getattr(transformers, pt_model_class_name)
config.output_hidden_states = True
tf_model = model_class(config)
pt_model = pt_model_class(config)
# Check we can load pt model in tf and vice-versa with model => model functions
tf_model = transformers.load_pytorch_model_in_tf2_model(tf_model, pt_model, tf_inputs=inputs_dict)
pt_model = transformers.load_tf2_model_in_pytorch_model(pt_model, tf_model)
# Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
pt_model.eval()
pt_inputs_dict = dict(
(name, torch.from_numpy(key.numpy()).to(torch.long)) for name, key in inputs_dict.items()
)
with torch.no_grad():
pto = pt_model(**pt_inputs_dict)
tfo = tf_model(inputs_dict, training=False)
tf_hidden_states = tfo[0].numpy()
pt_hidden_states = pto[0].numpy()
tf_hidden_states[np.isnan(tf_hidden_states)] = 0
pt_hidden_states[np.isnan(pt_hidden_states)] = 0
max_diff = np.amax(np.abs(tf_hidden_states - pt_hidden_states))
self.assertLessEqual(max_diff, 2e-2)
# Check we can load pt model in tf and vice-versa with checkpoint => model functions
with TemporaryDirectory() as tmpdirname:
pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
torch.save(pt_model.state_dict(), pt_checkpoint_path)
tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(tf_model, pt_checkpoint_path)
tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
tf_model.save_weights(tf_checkpoint_path)
pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(pt_model, tf_checkpoint_path)
# Check predictions on first output (logits/hidden-states) are close enought given low-level computational differences
pt_model.eval()
pt_inputs_dict = dict(
(name, torch.from_numpy(key.numpy()).to(torch.long)) for name, key in inputs_dict.items()
)
with torch.no_grad():
pto = pt_model(**pt_inputs_dict)
tfo = tf_model(inputs_dict)
tfo = tfo[0].numpy()
pto = pto[0].numpy()
tfo[np.isnan(tfo)] = 0
pto[np.isnan(pto)] = 0
max_diff = np.amax(np.abs(tfo - pto))
self.assertLessEqual(max_diff, 2e-2)
def test_compile_tf_model(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if self.is_encoder_decoder:
input_ids = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"
),
"encoder_input_ids": tf.keras.Input(
batch_shape=(2, 2000), name="encoder_input_ids", dtype="int32"
),
}
else:
input_ids = tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32")
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
for model_class in self.all_model_classes:
# Prepare our model
model = model_class(config)
# Let's load it from the disk to be sure we can use pretrained weights
with TemporaryDirectory() as tmpdirname:
outputs = model(inputs_dict) # build the model
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname)
outputs_dict = model(input_ids)
hidden_states = outputs_dict[0]
# Add a dense layer on top to test intetgration with other keras modules
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
# Compile extended model
extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
def test_keyword_and_dict_args(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
outputs_dict = model(inputs_dict)
inputs_keywords = copy.deepcopy(inputs_dict)
input_ids = inputs_keywords.pop(
"input_ids" if not self.is_encoder_decoder else "decoder_input_ids", None
)
outputs_keywords = model(input_ids, **inputs_keywords)
output_dict = outputs_dict[0].numpy()
output_keywords = outputs_keywords[0].numpy()
self.assertLess(np.sum(np.abs(output_dict - output_keywords)), 1e-6)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
decoder_seq_length = (
self.model_tester.decoder_seq_length
if hasattr(self.model_tester, "decoder_seq_length")
else self.model_tester.seq_length
)
encoder_seq_length = (
self.model_tester.encoder_seq_length
if hasattr(self.model_tester, "encoder_seq_length")
else self.model_tester.seq_length
)
decoder_key_length = (
self.model_tester.key_length if hasattr(self.model_tester, "key_length") else decoder_seq_length
)
encoder_key_length = (
self.model_tester.key_length if hasattr(self.model_tester, "key_length") else encoder_seq_length
)
for model_class in self.all_model_classes:
config.output_attentions = True
config.output_hidden_states = False
model = model_class(config)
outputs = model(inputs_dict)
attentions = [t.numpy() for t in outputs[-1]]
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
if self.is_encoder_decoder:
self.assertEqual(out_len % 2, 0)
decoder_attentions = outputs[(out_len // 2) - 1]
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, False)
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
# Check attention is always last and order is fine
config.output_attentions = True
config.output_hidden_states = True
model = model_class(config)
outputs = model(inputs_dict)
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
self.assertEqual(model.config.output_attentions, True)
self.assertEqual(model.config.output_hidden_states, True)
attentions = [t.numpy() for t in outputs[-1]]
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
config.output_hidden_states = True
config.output_attentions = False
model = model_class(config)
outputs = model(inputs_dict)
hidden_states = [t.numpy() for t in outputs[-1]]
self.assertEqual(model.config.output_attentions, False)
self.assertEqual(model.config.output_hidden_states, True)
self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
self.assertListEqual(
list(hidden_states[0].shape[-2:]), [self.model_tester.seq_length, self.model_tester.hidden_size]
)
def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
x = model.get_output_embeddings()
assert x is None or isinstance(x, tf.keras.layers.Layer)
def test_determinism(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
first, second = model(inputs_dict, training=False)[0], model(inputs_dict, training=False)[0]
out_1 = first.numpy()
out_2 = second.numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def _get_embeds(self, wte, input_ids):
# ^^ In our TF models, the input_embeddings can take slightly different forms,
# so we try a few of them.
# We used to fall back to just synthetically creating a dummy tensor of ones:
try:
x = wte(input_ids, mode="embedding")
except Exception:
try:
x = wte([input_ids], mode="embedding")
except Exception:
try:
x = wte([input_ids, None, None, None], mode="embedding")
except Exception:
if hasattr(self.model_tester, "embedding_size"):
x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32)
else:
x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
return x
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.is_encoder_decoder:
input_ids = inputs_dict["input_ids"]
del inputs_dict["input_ids"]
else:
encoder_input_ids = inputs_dict["encoder_input_ids"]
decoder_input_ids = inputs_dict["decoder_input_ids"]
del inputs_dict["encoder_input_ids"]
del inputs_dict["decoder_input_ids"]
for model_class in self.all_model_classes:
model = model_class(config)
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
inputs_dict["inputs_embeds"] = self._get_embeds(wte, input_ids)
else:
inputs_dict["encoder_inputs_embeds"] = self._get_embeds(wte, encoder_input_ids)
inputs_dict["decoder_inputs_embeds"] = self._get_embeds(wte, decoder_input_ids)
outputs = model(inputs_dict)
def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
"""Creates a random int32 tensor of the shape within the vocab size."""
if rng is None:
rng = random.Random()
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(rng.randint(0, vocab_size - 1))
output = tf.constant(values, shape=shape, dtype=dtype if dtype is not None else tf.int32)
return output
if __name__ == "__main__":
unittest.main()
transformers/tests/tokenization_tests_commons.py deleted 100644 → 0
View file @ 54abc67a
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# 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.
from __future__ import absolute_import, division, print_function, unicode_literals
import os
import shutil
import sys
import tempfile
import unittest
from io import open
if sys.version_info[0] == 2:
import cPickle as pickle
class TemporaryDirectory(object):
"""Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
def __enter__(self):
self.name = tempfile.mkdtemp()
return self.name
def __exit__(self, exc_type, exc_value, traceback):
shutil.rmtree(self.name)
else:
import pickle
TemporaryDirectory = tempfile.TemporaryDirectory
unicode = str
class CommonTestCases:
class CommonTokenizerTester(unittest.TestCase):
tokenizer_class = None
def setUp(self):
self.tmpdirname = tempfile.mkdtemp()
def tearDown(self):
shutil.rmtree(self.tmpdirname)
def get_tokenizer(self, **kwargs):
raise NotImplementedError
def get_input_output_texts(self):
raise NotImplementedError
def test_tokenizers_common_properties(self):
tokenizer = self.get_tokenizer()
attributes_list = [
"bos_token",
"eos_token",
"unk_token",
"sep_token",
"pad_token",
"cls_token",
"mask_token",
]
for attr in attributes_list:
self.assertTrue(hasattr(tokenizer, attr))
self.assertTrue(hasattr(tokenizer, attr + "_id"))
self.assertTrue(hasattr(tokenizer, "additional_special_tokens"))
self.assertTrue(hasattr(tokenizer, "additional_special_tokens_ids"))
attributes_list = ["max_len", "init_inputs", "init_kwargs", "added_tokens_encoder", "added_tokens_decoder"]
for attr in attributes_list:
self.assertTrue(hasattr(tokenizer, attr))
def test_save_and_load_tokenizer(self):
# safety check on max_len default value so we are sure the test works
tokenizer = self.get_tokenizer()
self.assertNotEqual(tokenizer.max_len, 42)
# Now let's start the test
tokenizer = self.get_tokenizer(max_len=42)
before_tokens = tokenizer.encode("He is very happy, UNwant\u00E9d,running", add_special_tokens=False)
with TemporaryDirectory() as tmpdirname:
tokenizer.save_pretrained(tmpdirname)
tokenizer = self.tokenizer_class.from_pretrained(tmpdirname)
after_tokens = tokenizer.encode("He is very happy, UNwant\u00E9d,running", add_special_tokens=False)
self.assertListEqual(before_tokens, after_tokens)
self.assertEqual(tokenizer.max_len, 42)
tokenizer = self.tokenizer_class.from_pretrained(tmpdirname, max_len=43)
self.assertEqual(tokenizer.max_len, 43)
def test_pickle_tokenizer(self):
tokenizer = self.get_tokenizer()
self.assertIsNotNone(tokenizer)
text = "Munich and Berlin are nice cities"
subwords = tokenizer.tokenize(text)
with TemporaryDirectory() as tmpdirname:
filename = os.path.join(tmpdirname, "tokenizer.bin")
with open(filename, "wb") as handle:
pickle.dump(tokenizer, handle)
with open(filename, "rb") as handle:
tokenizer_new = pickle.load(handle)
subwords_loaded = tokenizer_new.tokenize(text)
self.assertListEqual(subwords, subwords_loaded)
def test_added_tokens_do_lower_case(self):
tokenizer = self.get_tokenizer(do_lower_case=True)
special_token = tokenizer.all_special_tokens[0]
text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
toks0 = tokenizer.tokenize(text) # toks before adding new_toks
new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", "AAAAA BBBBBB", "CCCCCCCCCDDDDDDDD"]
added = tokenizer.add_tokens(new_toks)
self.assertEqual(added, 2)
toks = tokenizer.tokenize(text)
toks2 = tokenizer.tokenize(text2)
self.assertEqual(len(toks), len(toks2))
self.assertNotEqual(len(toks), len(toks0)) # toks0 should be longer
self.assertListEqual(toks, toks2)
# Check that none of the special tokens are lowercased
sequence_with_special_tokens = "A " + " yEs ".join(tokenizer.all_special_tokens) + " B"
tokenized_sequence = tokenizer.tokenize(sequence_with_special_tokens)
for special_token in tokenizer.all_special_tokens:
self.assertTrue(special_token in tokenized_sequence)
tokenizer = self.get_tokenizer(do_lower_case=False)
added = tokenizer.add_tokens(new_toks)
self.assertEqual(added, 4)
toks = tokenizer.tokenize(text)
toks2 = tokenizer.tokenize(text2)
self.assertEqual(len(toks), len(toks2)) # Length should still be the same
self.assertNotEqual(len(toks), len(toks0))
self.assertNotEqual(toks[1], toks2[1]) # But at least the first non-special tokens should differ
def test_add_tokens_tokenizer(self):
tokenizer = self.get_tokenizer()
vocab_size = tokenizer.vocab_size
all_size = len(tokenizer)
self.assertNotEqual(vocab_size, 0)
self.assertEqual(vocab_size, all_size)
new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
added_toks = tokenizer.add_tokens(new_toks)
vocab_size_2 = tokenizer.vocab_size
all_size_2 = len(tokenizer)
self.assertNotEqual(vocab_size_2, 0)
self.assertEqual(vocab_size, vocab_size_2)
self.assertEqual(added_toks, len(new_toks))
self.assertEqual(all_size_2, all_size + len(new_toks))
tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False)
out_string = tokenizer.decode(tokens)
self.assertGreaterEqual(len(tokens), 4)
self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
vocab_size_3 = tokenizer.vocab_size
all_size_3 = len(tokenizer)
self.assertNotEqual(vocab_size_3, 0)
self.assertEqual(vocab_size, vocab_size_3)
self.assertEqual(added_toks_2, len(new_toks_2))
self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))
tokens = tokenizer.encode(
">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False
)
out_string = tokenizer.decode(tokens)
self.assertGreaterEqual(len(tokens), 6)
self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
self.assertGreater(tokens[0], tokens[1])
self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
self.assertGreater(tokens[-2], tokens[-3])
self.assertEqual(tokens[0], tokenizer.eos_token_id)
self.assertEqual(tokens[-2], tokenizer.pad_token_id)
def test_add_special_tokens(self):
tokenizer = self.get_tokenizer()
input_text, output_text = self.get_input_output_texts()
special_token = "[SPECIAL TOKEN]"
tokenizer.add_special_tokens({"cls_token": special_token})
encoded_special_token = tokenizer.encode(special_token, add_special_tokens=False)
assert len(encoded_special_token) == 1
text = " ".join([input_text, special_token, output_text])
encoded = tokenizer.encode(text, add_special_tokens=False)
input_encoded = tokenizer.encode(input_text, add_special_tokens=False)
output_encoded = tokenizer.encode(output_text, add_special_tokens=False)
special_token_id = tokenizer.encode(special_token, add_special_tokens=False)
assert encoded == input_encoded + special_token_id + output_encoded
decoded = tokenizer.decode(encoded, skip_special_tokens=True)
assert special_token not in decoded
def test_required_methods_tokenizer(self):
tokenizer = self.get_tokenizer()
input_text, output_text = self.get_input_output_texts()
tokens = tokenizer.tokenize(input_text)
ids = tokenizer.convert_tokens_to_ids(tokens)
ids_2 = tokenizer.encode(input_text, add_special_tokens=False)
self.assertListEqual(ids, ids_2)
tokens_2 = tokenizer.convert_ids_to_tokens(ids)
text_2 = tokenizer.decode(ids)
self.assertEqual(text_2, output_text)
self.assertNotEqual(len(tokens_2), 0)
self.assertIsInstance(text_2, (str, unicode))
def test_encode_decode_with_spaces(self):
tokenizer = self.get_tokenizer()
new_toks = ["[ABC]", "[DEF]", "GHI IHG"]
tokenizer.add_tokens(new_toks)
input = "[ABC] [DEF] [ABC] GHI IHG [DEF]"
encoded = tokenizer.encode(input, add_special_tokens=False)
decoded = tokenizer.decode(encoded)
self.assertEqual(decoded, input)
def test_pretrained_model_lists(self):
weights_list = list(self.tokenizer_class.max_model_input_sizes.keys())
weights_lists_2 = []
for file_id, map_list in self.tokenizer_class.pretrained_vocab_files_map.items():
weights_lists_2.append(list(map_list.keys()))
for weights_list_2 in weights_lists_2:
self.assertListEqual(weights_list, weights_list_2)
def test_mask_output(self):
if sys.version_info <= (3, 0):
return
tokenizer = self.get_tokenizer()
if tokenizer.build_inputs_with_special_tokens.__qualname__.split(".")[0] != "PreTrainedTokenizer":
seq_0 = "Test this method."
seq_1 = "With these inputs."
information = tokenizer.encode_plus(seq_0, seq_1, add_special_tokens=True)
sequences, mask = information["input_ids"], information["token_type_ids"]
self.assertEqual(len(sequences), len(mask))
def test_number_of_added_tokens(self):
tokenizer = self.get_tokenizer()
seq_0 = "Test this method."
seq_1 = "With these inputs."
sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=False)
attached_sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=True)
# Method is implemented (e.g. not GPT-2)
if len(attached_sequences) != 2:
self.assertEqual(tokenizer.num_added_tokens(pair=True), len(attached_sequences) - len(sequences))
def test_maximum_encoding_length_single_input(self):
tokenizer = self.get_tokenizer()
seq_0 = "This is a sentence to be encoded."
stride = 2
sequence = tokenizer.encode(seq_0, add_special_tokens=False)
num_added_tokens = tokenizer.num_added_tokens()
total_length = len(sequence) + num_added_tokens
information = tokenizer.encode_plus(
seq_0,
max_length=total_length - 2,
add_special_tokens=True,
stride=stride,
return_overflowing_tokens=True,
)
truncated_sequence = information["input_ids"]
overflowing_tokens = information["overflowing_tokens"]
self.assertEqual(len(overflowing_tokens), 2 + stride)
self.assertEqual(overflowing_tokens, sequence[-(2 + stride) :])
self.assertEqual(len(truncated_sequence), total_length - 2)
self.assertEqual(truncated_sequence, tokenizer.build_inputs_with_special_tokens(sequence[:-2]))
def test_maximum_encoding_length_pair_input(self):
tokenizer = self.get_tokenizer()
seq_0 = "This is a sentence to be encoded."
seq_1 = "This is another sentence to be encoded."
stride = 2
sequence_0_no_special_tokens = tokenizer.encode(seq_0, add_special_tokens=False)
sequence_1_no_special_tokens = tokenizer.encode(seq_1, add_special_tokens=False)
sequence = tokenizer.encode(seq_0, seq_1, add_special_tokens=True)
truncated_second_sequence = tokenizer.build_inputs_with_special_tokens(
tokenizer.encode(seq_0, add_special_tokens=False),
tokenizer.encode(seq_1, add_special_tokens=False)[:-2],
)
information = tokenizer.encode_plus(
seq_0,
seq_1,
max_length=len(sequence) - 2,
add_special_tokens=True,
stride=stride,
truncation_strategy="only_second",
return_overflowing_tokens=True,
)
information_first_truncated = tokenizer.encode_plus(
seq_0,
seq_1,
max_length=len(sequence) - 2,
add_special_tokens=True,
stride=stride,
truncation_strategy="only_first",
return_overflowing_tokens=True,
)
truncated_sequence = information["input_ids"]
overflowing_tokens = information["overflowing_tokens"]
overflowing_tokens_first_truncated = information_first_truncated["overflowing_tokens"]
self.assertEqual(len(overflowing_tokens), 2 + stride)
self.assertEqual(overflowing_tokens, sequence_1_no_special_tokens[-(2 + stride) :])
self.assertEqual(overflowing_tokens_first_truncated, sequence_0_no_special_tokens[-(2 + stride) :])
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
self.assertEqual(truncated_sequence, truncated_second_sequence)
def test_encode_input_type(self):
tokenizer = self.get_tokenizer()
sequence = "Let's encode this sequence"
tokens = tokenizer.tokenize(sequence)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
formatted_input = tokenizer.encode(sequence, add_special_tokens=True)
self.assertEqual(tokenizer.encode(tokens, add_special_tokens=True), formatted_input)
self.assertEqual(tokenizer.encode(input_ids, add_special_tokens=True), formatted_input)
def test_special_tokens_mask(self):
tokenizer = self.get_tokenizer()
sequence_0 = "Encode this."
sequence_1 = "This one too please."
# Testing single inputs
encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
encoded_sequence_dict = tokenizer.encode_plus(
sequence_0, add_special_tokens=True, return_special_tokens_mask=True
)
encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
filtered_sequence = [
(x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)
]
filtered_sequence = [x for x in filtered_sequence if x is not None]
self.assertEqual(encoded_sequence, filtered_sequence)
# Testing inputs pairs
encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False) + tokenizer.encode(
sequence_1, add_special_tokens=False
)
encoded_sequence_dict = tokenizer.encode_plus(
sequence_0, sequence_1, add_special_tokens=True, return_special_tokens_mask=True
)
encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
filtered_sequence = [
(x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)
]
filtered_sequence = [x for x in filtered_sequence if x is not None]
self.assertEqual(encoded_sequence, filtered_sequence)
# Testing with already existing special tokens
if tokenizer.cls_token_id == tokenizer.unk_token_id and tokenizer.cls_token_id == tokenizer.unk_token_id:
tokenizer.add_special_tokens({"cls_token": "</s>", "sep_token": "<s>"})
encoded_sequence_dict = tokenizer.encode_plus(
sequence_0, add_special_tokens=True, return_special_tokens_mask=True
)
encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
special_tokens_mask_orig = encoded_sequence_dict["special_tokens_mask"]
special_tokens_mask = tokenizer.get_special_tokens_mask(
encoded_sequence_w_special, already_has_special_tokens=True
)
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
self.assertEqual(special_tokens_mask_orig, special_tokens_mask)
def test_padding_to_max_length(self):
tokenizer = self.get_tokenizer()
sequence = "Sequence"
padding_size = 10
padding_idx = tokenizer.pad_token_id
# RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
tokenizer.padding_side = "right"
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
padded_sequence = tokenizer.encode(
sequence, max_length=sequence_length + padding_size, pad_to_max_length=True
)
padded_sequence_length = len(padded_sequence)
assert sequence_length + padding_size == padded_sequence_length
assert encoded_sequence + [padding_idx] * padding_size == padded_sequence
# LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
tokenizer.padding_side = "left"
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
padded_sequence = tokenizer.encode(
sequence, max_length=sequence_length + padding_size, pad_to_max_length=True
)
padded_sequence_length = len(padded_sequence)
assert sequence_length + padding_size == padded_sequence_length
assert [padding_idx] * padding_size + encoded_sequence == padded_sequence
# RIGHT & LEFT PADDING - Check that nothing is done when a maximum length is not specified
encoded_sequence = tokenizer.encode(sequence)
sequence_length = len(encoded_sequence)
tokenizer.padding_side = "right"
padded_sequence_right = tokenizer.encode(sequence, pad_to_max_length=True)
padded_sequence_right_length = len(padded_sequence_right)
tokenizer.padding_side = "left"
padded_sequence_left = tokenizer.encode(sequence, pad_to_max_length=True)
padded_sequence_left_length = len(padded_sequence_left)
assert sequence_length == padded_sequence_right_length
assert encoded_sequence == padded_sequence_right
assert sequence_length == padded_sequence_left_length
assert encoded_sequence == padded_sequence_left
def test_encode_plus_with_padding(self):
tokenizer = self.get_tokenizer()
sequence = "Sequence"
padding_size = 10
padding_idx = tokenizer.pad_token_id
token_type_padding_idx = tokenizer.pad_token_type_id
encoded_sequence = tokenizer.encode_plus(sequence, return_special_tokens_mask=True)
input_ids = encoded_sequence["input_ids"]
token_type_ids = encoded_sequence["token_type_ids"]
attention_mask = encoded_sequence["attention_mask"]
special_tokens_mask = encoded_sequence["special_tokens_mask"]
sequence_length = len(input_ids)
# Test right padding
tokenizer.padding_side = "right"
padded_sequence = tokenizer.encode_plus(
sequence,
max_length=sequence_length + padding_size,
pad_to_max_length=True,
return_special_tokens_mask=True,
)
padded_input_ids = padded_sequence["input_ids"]
padded_token_type_ids = padded_sequence["token_type_ids"]
padded_attention_mask = padded_sequence["attention_mask"]
padded_special_tokens_mask = padded_sequence["special_tokens_mask"]
padded_sequence_length = len(padded_input_ids)
assert sequence_length + padding_size == padded_sequence_length
assert input_ids + [padding_idx] * padding_size == padded_input_ids
assert token_type_ids + [token_type_padding_idx] * padding_size == padded_token_type_ids
assert attention_mask + [0] * padding_size == padded_attention_mask
assert special_tokens_mask + [1] * padding_size == padded_special_tokens_mask
# Test left padding
tokenizer.padding_side = "left"
padded_sequence = tokenizer.encode_plus(
sequence,
max_length=sequence_length + padding_size,
pad_to_max_length=True,
return_special_tokens_mask=True,
)
padded_input_ids = padded_sequence["input_ids"]
padded_token_type_ids = padded_sequence["token_type_ids"]
padded_attention_mask = padded_sequence["attention_mask"]
padded_special_tokens_mask = padded_sequence["special_tokens_mask"]
padded_sequence_length = len(padded_input_ids)
assert sequence_length + padding_size == padded_sequence_length
assert [padding_idx] * padding_size + input_ids == padded_input_ids
assert [token_type_padding_idx] * padding_size + token_type_ids == padded_token_type_ids
assert [0] * padding_size + attention_mask == padded_attention_mask
assert [1] * padding_size + special_tokens_mask == padded_special_tokens_mask
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment