[Test refactor 1/5] Per-folder tests reorganization (#15725)
* Per-folder tests reorganization Co-authored-by: sgugger <sylvain.gugger@gmail.com> Co-authored-by: Stas Bekman <stas@stason.org>
This commit is contained in:
Lysandre Debut
0
tests/pegasus/__init__.py
Normal file
0
tests/pegasus/__init__.py
Normal file
338
tests/pegasus/test_modeling_flax_pegasus.py
Normal file
338
tests/pegasus/test_modeling_flax_pegasus.py
Normal file
@@ -0,0 +1,338 @@
|
||||
# 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.
|
||||
|
||||
|
||||
import unittest
|
||||
|
||||
from transformers import PegasusConfig, PegasusTokenizer, is_flax_available
|
||||
from transformers.testing_utils import require_flax, slow
|
||||
|
||||
from ..test_configuration_common import ConfigTester
|
||||
from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_flax_available():
|
||||
import os
|
||||
|
||||
# The slow tests are often failing with OOM error on GPU
|
||||
# This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
|
||||
# but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
|
||||
os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
|
||||
import numpy as np
|
||||
|
||||
import jax
|
||||
import jax.numpy as jnp
|
||||
from transformers import FlaxPegasusForConditionalGeneration, FlaxPegasusModel
|
||||
|
||||
|
||||
@require_flax
|
||||
class FlaxPegasusModelTester:
|
||||
config_cls = PegasusConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size).clip(3, self.vocab_size)
|
||||
eos_tensor = np.expand_dims(np.array([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = np.concatenate([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_pegasus_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_use_cache_forward(self, model_class_name, config, inputs_dict):
|
||||
max_decoder_length = 20
|
||||
model = model_class_name(config)
|
||||
|
||||
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
||||
|
||||
decoder_input_ids, decoder_attention_mask = (
|
||||
inputs_dict["decoder_input_ids"],
|
||||
inputs_dict["decoder_attention_mask"],
|
||||
)
|
||||
|
||||
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
||||
decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4")
|
||||
|
||||
decoder_position_ids = jnp.broadcast_to(
|
||||
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
||||
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
||||
)
|
||||
outputs_cache = model.decode(
|
||||
decoder_input_ids[:, :-1],
|
||||
encoder_outputs,
|
||||
decoder_attention_mask=decoder_attention_mask,
|
||||
past_key_values=past_key_values,
|
||||
decoder_position_ids=decoder_position_ids,
|
||||
)
|
||||
|
||||
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
||||
outputs_cache_next = model.decode(
|
||||
decoder_input_ids[:, -1:],
|
||||
encoder_outputs,
|
||||
decoder_attention_mask=decoder_attention_mask,
|
||||
past_key_values=outputs_cache.past_key_values,
|
||||
decoder_position_ids=decoder_position_ids,
|
||||
)
|
||||
|
||||
outputs = model.decode(decoder_input_ids, encoder_outputs)
|
||||
|
||||
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
||||
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
||||
|
||||
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict):
|
||||
max_decoder_length = 20
|
||||
model = model_class_name(config)
|
||||
|
||||
encoder_outputs = model.encode(inputs_dict["input_ids"])
|
||||
|
||||
decoder_input_ids, decoder_attention_mask = (
|
||||
inputs_dict["decoder_input_ids"],
|
||||
inputs_dict["decoder_attention_mask"],
|
||||
)
|
||||
|
||||
decoder_attention_mask_cache = jnp.concatenate(
|
||||
[
|
||||
decoder_attention_mask,
|
||||
jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
|
||||
past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
|
||||
decoder_position_ids = jnp.broadcast_to(
|
||||
jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :],
|
||||
(decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1),
|
||||
)
|
||||
|
||||
outputs_cache = model.decode(
|
||||
decoder_input_ids[:, :-1],
|
||||
encoder_outputs,
|
||||
decoder_attention_mask=decoder_attention_mask_cache,
|
||||
past_key_values=past_key_values,
|
||||
decoder_position_ids=decoder_position_ids,
|
||||
)
|
||||
decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4")
|
||||
outputs_cache_next = model.decode(
|
||||
decoder_input_ids[:, -1:],
|
||||
encoder_outputs,
|
||||
past_key_values=outputs_cache.past_key_values,
|
||||
decoder_attention_mask=decoder_attention_mask_cache,
|
||||
decoder_position_ids=decoder_position_ids,
|
||||
)
|
||||
|
||||
outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask)
|
||||
|
||||
diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
|
||||
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
|
||||
|
||||
|
||||
def prepare_pegasus_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = np.not_equal(input_ids, config.pad_token_id).astype(np.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = np.concatenate(
|
||||
[
|
||||
np.ones(decoder_input_ids[:, :1].shape, dtype=np.int8),
|
||||
np.not_equal(decoder_input_ids[:, 1:], config.pad_token_id).astype(np.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_flax
|
||||
class FlaxPegasusModelTest(FlaxModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(
|
||||
FlaxPegasusForConditionalGeneration,
|
||||
FlaxPegasusModel,
|
||||
)
|
||||
if is_flax_available()
|
||||
else ()
|
||||
)
|
||||
all_generative_model_classes = (FlaxPegasusForConditionalGeneration,) if is_flax_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
test_head_masking = False
|
||||
test_onnx = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = FlaxPegasusModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_use_cache_forward(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
for model_class in self.all_model_classes:
|
||||
self.model_tester.check_use_cache_forward(model_class, config, inputs_dict)
|
||||
|
||||
def test_use_cache_forward_with_attn_mask(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
for model_class in self.all_model_classes:
|
||||
self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict)
|
||||
|
||||
def test_encode(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
with self.subTest(model_class.__name__):
|
||||
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
|
||||
model = model_class(config)
|
||||
|
||||
@jax.jit
|
||||
def encode_jitted(input_ids, attention_mask=None, **kwargs):
|
||||
return model.encode(input_ids=input_ids, attention_mask=attention_mask)
|
||||
|
||||
with self.subTest("JIT Enabled"):
|
||||
jitted_outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
||||
|
||||
with self.subTest("JIT Disabled"):
|
||||
with jax.disable_jit():
|
||||
outputs = encode_jitted(**prepared_inputs_dict).to_tuple()
|
||||
|
||||
self.assertEqual(len(outputs), len(jitted_outputs))
|
||||
for jitted_output, output in zip(jitted_outputs, outputs):
|
||||
self.assertEqual(jitted_output.shape, output.shape)
|
||||
|
||||
def test_decode(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
with self.subTest(model_class.__name__):
|
||||
model = model_class(config)
|
||||
encoder_outputs = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
|
||||
|
||||
prepared_inputs_dict = {
|
||||
"decoder_input_ids": inputs_dict["decoder_input_ids"],
|
||||
"decoder_attention_mask": inputs_dict["decoder_attention_mask"],
|
||||
"encoder_outputs": encoder_outputs,
|
||||
}
|
||||
|
||||
@jax.jit
|
||||
def decode_jitted(decoder_input_ids, decoder_attention_mask, encoder_outputs):
|
||||
return model.decode(
|
||||
decoder_input_ids=decoder_input_ids,
|
||||
decoder_attention_mask=decoder_attention_mask,
|
||||
encoder_outputs=encoder_outputs,
|
||||
)
|
||||
|
||||
with self.subTest("JIT Enabled"):
|
||||
jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
||||
|
||||
with self.subTest("JIT Disabled"):
|
||||
with jax.disable_jit():
|
||||
outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
|
||||
|
||||
self.assertEqual(len(outputs), len(jitted_outputs))
|
||||
for jitted_output, output in zip(jitted_outputs, outputs):
|
||||
self.assertEqual(jitted_output.shape, output.shape)
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
for model_class_name in self.all_model_classes:
|
||||
model = model_class_name.from_pretrained("google/pegasus-large", from_pt=True)
|
||||
input_ids = np.ones((1, 1))
|
||||
outputs = model(input_ids)
|
||||
self.assertIsNotNone(outputs)
|
||||
|
||||
@slow
|
||||
def test_pegasus_xsum_summary(self):
|
||||
model = FlaxPegasusForConditionalGeneration.from_pretrained("google/pegasus-xsum")
|
||||
tokenizer = PegasusTokenizer.from_pretrained("google/pegasus-xsum")
|
||||
|
||||
src_text = [
|
||||
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow.""",
|
||||
""" The London trio are up for best UK act and best album, as well as getting two nominations in the best song category."We got told like this morning 'Oh I think you're nominated'", said Dappy."And I was like 'Oh yeah, which one?' And now we've got nominated for four awards. I mean, wow!"Bandmate Fazer added: "We thought it's best of us to come down and mingle with everyone and say hello to the cameras. And now we find we've got four nominations."The band have two shots at the best song prize, getting the nod for their Tynchy Stryder collaboration Number One, and single Strong Again.Their album Uncle B will also go up against records by the likes of Beyonce and Kanye West.N-Dubz picked up the best newcomer Mobo in 2007, but female member Tulisa said they wouldn't be too disappointed if they didn't win this time around."At the end of the day we're grateful to be where we are in our careers."If it don't happen then it don't happen - live to fight another day and keep on making albums and hits for the fans."Dappy also revealed they could be performing live several times on the night.The group will be doing Number One and also a possible rendition of the War Child single, I Got Soul.The charity song is a re-working of The Killers' All These Things That I've Done and is set to feature artists like Chipmunk, Ironik and Pixie Lott.This year's Mobos will be held outside of London for the first time, in Glasgow on 30 September.N-Dubz said they were looking forward to performing for their Scottish fans and boasted about their recent shows north of the border."We just done Edinburgh the other day," said Dappy."We smashed up an N-Dubz show over there. We done Aberdeen about three or four months ago - we smashed up that show over there! Everywhere we go we smash it up!" """,
|
||||
]
|
||||
|
||||
tgt_text = [
|
||||
"California's largest electricity provider has turned off power to hundreds of thousands of customers.",
|
||||
"Pop group N-Dubz have revealed they were surprised to get four nominations for this year's Mobo Awards.",
|
||||
]
|
||||
|
||||
inputs = tokenizer(src_text, return_tensors="np", truncation=True, max_length=512, padding=True)
|
||||
translated_tokens = model.generate(**inputs, num_beams=2).sequences
|
||||
decoded = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
|
||||
assert tgt_text == decoded
|
||||
553
tests/pegasus/test_modeling_pegasus.py
Normal file
553
tests/pegasus/test_modeling_pegasus.py
Normal file
@@ -0,0 +1,553 @@
|
||||
# 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 PEGASUS model. """
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from transformers import PegasusConfig, is_torch_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
|
||||
|
||||
from ..generation.test_generation_utils import GenerationTesterMixin
|
||||
from ..mbart.test_modeling_mbart import AbstractSeq2SeqIntegrationTest
|
||||
from ..test_configuration_common import ConfigTester
|
||||
from ..test_modeling_common import ModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import AutoModelForSeq2SeqLM, PegasusForConditionalGeneration, PegasusModel
|
||||
from transformers.models.pegasus.modeling_pegasus import PegasusDecoder, PegasusEncoder, PegasusForCausalLM
|
||||
|
||||
|
||||
def prepare_pegasus_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
head_mask=None,
|
||||
decoder_head_mask=None,
|
||||
cross_attn_head_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = input_ids.ne(config.pad_token_id)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
|
||||
if head_mask is None:
|
||||
head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
|
||||
if decoder_head_mask is None:
|
||||
decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
||||
if cross_attn_head_mask is None:
|
||||
cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": attention_mask,
|
||||
"head_mask": head_mask,
|
||||
"decoder_head_mask": decoder_head_mask,
|
||||
"cross_attn_head_mask": cross_attn_head_mask,
|
||||
}
|
||||
|
||||
|
||||
class PegasusModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=16,
|
||||
num_hidden_layers=2,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=4,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_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.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
|
||||
3,
|
||||
)
|
||||
input_ids[:, -1] = self.eos_token_id # Eos Token
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.get_config()
|
||||
inputs_dict = prepare_pegasus_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def get_pipeline_config(self):
|
||||
return PegasusConfig(
|
||||
vocab_size=200,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=200,
|
||||
eos_token_id=self.eos_token_id,
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
)
|
||||
|
||||
def get_config(self):
|
||||
return PegasusConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_id=self.eos_token_id,
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config, inputs_dict = self.prepare_config_and_inputs()
|
||||
return config, inputs_dict
|
||||
|
||||
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = PegasusModel(config=config).get_decoder().to(torch_device).eval()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
attention_mask = inputs_dict["attention_mask"]
|
||||
head_mask = inputs_dict["head_mask"]
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
|
||||
# create hypothetical multiple next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
|
||||
"last_hidden_state"
|
||||
]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
||||
|
||||
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
||||
|
||||
# test that outputs are equal for slice
|
||||
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
||||
|
||||
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
|
||||
model = PegasusModel(config=config).to(torch_device).eval()
|
||||
outputs = model(**inputs_dict)
|
||||
|
||||
encoder_last_hidden_state = outputs.encoder_last_hidden_state
|
||||
last_hidden_state = outputs.last_hidden_state
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
encoder = model.get_encoder()
|
||||
encoder.save_pretrained(tmpdirname)
|
||||
encoder = PegasusEncoder.from_pretrained(tmpdirname).to(torch_device)
|
||||
|
||||
encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
|
||||
0
|
||||
]
|
||||
|
||||
self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
decoder = model.get_decoder()
|
||||
decoder.save_pretrained(tmpdirname)
|
||||
decoder = PegasusDecoder.from_pretrained(tmpdirname).to(torch_device)
|
||||
|
||||
last_hidden_state_2 = decoder(
|
||||
input_ids=inputs_dict["decoder_input_ids"],
|
||||
attention_mask=inputs_dict["decoder_attention_mask"],
|
||||
encoder_hidden_states=encoder_last_hidden_state,
|
||||
encoder_attention_mask=inputs_dict["attention_mask"],
|
||||
)[0]
|
||||
|
||||
self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
|
||||
|
||||
|
||||
@require_torch
|
||||
class PegasusModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (PegasusModel, PegasusForConditionalGeneration) if is_torch_available() else ()
|
||||
all_generative_model_classes = (PegasusForConditionalGeneration,) if is_torch_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_resize_position_embeddings = True
|
||||
test_pruning = False
|
||||
test_missing_keys = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = PegasusModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_save_load_strict(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
||||
for model_class in self.all_model_classes:
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
|
||||
self.assertEqual(info["missing_keys"], [])
|
||||
|
||||
def test_decoder_model_past_with_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_encoder_decoder_model_standalone(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
|
||||
|
||||
def test_generate_fp16(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs()
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
model = PegasusForConditionalGeneration(config).eval().to(torch_device)
|
||||
if torch_device == "cuda":
|
||||
model.half()
|
||||
model.generate(input_ids, attention_mask=attention_mask)
|
||||
model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
|
||||
|
||||
|
||||
def assert_tensors_close(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if torch.allclose(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
|
||||
if a.numel() > 100:
|
||||
msg = f"tensor values are {pct_different:.1%} percent different."
|
||||
else:
|
||||
msg = f"{a} != {b}"
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return torch.tensor(tok_lst, dtype=torch.long, device=torch_device)
|
||||
|
||||
|
||||
@require_torch
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class PegasusXSUMIntegrationTest(AbstractSeq2SeqIntegrationTest):
|
||||
checkpoint_name = "google/pegasus-xsum"
|
||||
src_text = [
|
||||
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow.""",
|
||||
""" The London trio are up for best UK act and best album, as well as getting two nominations in the best song category."We got told like this morning 'Oh I think you're nominated'", said Dappy."And I was like 'Oh yeah, which one?' And now we've got nominated for four awards. I mean, wow!"Bandmate Fazer added: "We thought it's best of us to come down and mingle with everyone and say hello to the cameras. And now we find we've got four nominations."The band have two shots at the best song prize, getting the nod for their Tynchy Stryder collaboration Number One, and single Strong Again.Their album Uncle B will also go up against records by the likes of Beyonce and Kanye West.N-Dubz picked up the best newcomer Mobo in 2007, but female member Tulisa said they wouldn't be too disappointed if they didn't win this time around."At the end of the day we're grateful to be where we are in our careers."If it don't happen then it don't happen - live to fight another day and keep on making albums and hits for the fans."Dappy also revealed they could be performing live several times on the night.The group will be doing Number One and also a possible rendition of the War Child single, I Got Soul.The charity song is a re-working of The Killers' All These Things That I've Done and is set to feature artists like Chipmunk, Ironik and Pixie Lott.This year's Mobos will be held outside of London for the first time, in Glasgow on 30 September.N-Dubz said they were looking forward to performing for their Scottish fans and boasted about their recent shows north of the border."We just done Edinburgh the other day," said Dappy."We smashed up an N-Dubz show over there. We done Aberdeen about three or four months ago - we smashed up that show over there! Everywhere we go we smash it up!" """,
|
||||
]
|
||||
|
||||
tgt_text = [
|
||||
"California's largest electricity provider has turned off power to hundreds of thousands of customers.",
|
||||
"Pop group N-Dubz have revealed they were surprised to get four nominations for this year's Mobo Awards.",
|
||||
]
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
return AutoModelForSeq2SeqLM.from_pretrained(self.checkpoint_name).to(torch_device)
|
||||
|
||||
@slow
|
||||
def test_pegasus_xsum_summary(self):
|
||||
assert self.tokenizer.model_max_length == 512
|
||||
inputs = self.tokenizer(self.src_text, return_tensors="pt", truncation=True, max_length=512, padding=True).to(
|
||||
torch_device
|
||||
)
|
||||
assert inputs.input_ids.shape == (2, 421)
|
||||
translated_tokens = self.model.generate(**inputs, num_beams=2)
|
||||
decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
|
||||
assert self.tgt_text == decoded
|
||||
|
||||
if "cuda" not in torch_device:
|
||||
return
|
||||
# Demonstrate fp16 issue, Contributions welcome!
|
||||
self.model.half()
|
||||
translated_tokens_fp16 = self.model.generate(**inputs, max_length=10)
|
||||
decoded_fp16 = self.tokenizer.batch_decode(translated_tokens_fp16, skip_special_tokens=True)
|
||||
assert decoded_fp16 == [
|
||||
"California's largest electricity provider has begun",
|
||||
"N-Dubz have revealed they were",
|
||||
]
|
||||
|
||||
|
||||
class PegasusStandaloneDecoderModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
vocab_size=99,
|
||||
batch_size=13,
|
||||
d_model=16,
|
||||
decoder_seq_length=7,
|
||||
is_training=True,
|
||||
is_decoder=True,
|
||||
use_attention_mask=True,
|
||||
use_cache=False,
|
||||
use_labels=True,
|
||||
decoder_start_token_id=2,
|
||||
decoder_ffn_dim=32,
|
||||
decoder_layers=4,
|
||||
encoder_attention_heads=4,
|
||||
decoder_attention_heads=4,
|
||||
max_position_embeddings=30,
|
||||
is_encoder_decoder=False,
|
||||
pad_token_id=0,
|
||||
bos_token_id=1,
|
||||
eos_token_id=2,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.decoder_seq_length = decoder_seq_length
|
||||
# For common tests
|
||||
self.seq_length = self.decoder_seq_length
|
||||
self.is_training = is_training
|
||||
self.use_attention_mask = use_attention_mask
|
||||
self.use_labels = use_labels
|
||||
|
||||
self.vocab_size = vocab_size
|
||||
self.d_model = d_model
|
||||
self.hidden_size = d_model
|
||||
self.num_hidden_layers = decoder_layers
|
||||
self.decoder_layers = decoder_layers
|
||||
self.decoder_ffn_dim = decoder_ffn_dim
|
||||
self.encoder_attention_heads = encoder_attention_heads
|
||||
self.decoder_attention_heads = decoder_attention_heads
|
||||
self.num_attention_heads = decoder_attention_heads
|
||||
self.eos_token_id = eos_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.decoder_start_token_id = decoder_start_token_id
|
||||
self.use_cache = use_cache
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.is_encoder_decoder = is_encoder_decoder
|
||||
|
||||
self.scope = None
|
||||
self.decoder_key_length = decoder_seq_length
|
||||
self.base_model_out_len = 2
|
||||
self.decoder_attention_idx = 1
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
||||
|
||||
attention_mask = None
|
||||
if self.use_attention_mask:
|
||||
attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
|
||||
|
||||
lm_labels = None
|
||||
if self.use_labels:
|
||||
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
|
||||
|
||||
config = PegasusConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.d_model,
|
||||
decoder_layers=self.decoder_layers,
|
||||
decoder_ffn_dim=self.decoder_ffn_dim,
|
||||
encoder_attention_heads=self.encoder_attention_heads,
|
||||
decoder_attention_heads=self.decoder_attention_heads,
|
||||
eos_token_id=self.eos_token_id,
|
||||
bos_token_id=self.bos_token_id,
|
||||
use_cache=self.use_cache,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.decoder_start_token_id,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
is_encoder_decoder=self.is_encoder_decoder,
|
||||
)
|
||||
|
||||
return (
|
||||
config,
|
||||
input_ids,
|
||||
attention_mask,
|
||||
lm_labels,
|
||||
)
|
||||
|
||||
def create_and_check_decoder_model_past(
|
||||
self,
|
||||
config,
|
||||
input_ids,
|
||||
attention_mask,
|
||||
lm_labels,
|
||||
):
|
||||
config.use_cache = True
|
||||
model = PegasusDecoder(config=config).to(torch_device).eval()
|
||||
# first forward pass
|
||||
outputs = model(input_ids, use_cache=True)
|
||||
outputs_use_cache_conf = model(input_ids)
|
||||
outputs_no_past = model(input_ids, use_cache=False)
|
||||
|
||||
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
|
||||
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
||||
|
||||
past_key_values = outputs["past_key_values"]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# test that outputs are equal for slice
|
||||
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
||||
|
||||
def create_and_check_decoder_model_attention_mask_past(
|
||||
self,
|
||||
config,
|
||||
input_ids,
|
||||
attention_mask,
|
||||
lm_labels,
|
||||
):
|
||||
model = PegasusDecoder(config=config).to(torch_device).eval()
|
||||
|
||||
# create attention mask
|
||||
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
||||
|
||||
half_seq_length = input_ids.shape[-1] // 2
|
||||
attn_mask[:, half_seq_length:] = 0
|
||||
|
||||
# first forward pass
|
||||
past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
||||
|
||||
# change a random masked slice from input_ids
|
||||
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
||||
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
||||
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
||||
|
||||
# append to next input_ids and attn_mask
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
attn_mask = torch.cat(
|
||||
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
||||
dim=1,
|
||||
)
|
||||
|
||||
# get two different outputs
|
||||
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
|
||||
"last_hidden_state"
|
||||
]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# test that outputs are equal for slice
|
||||
assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
attention_mask,
|
||||
lm_labels,
|
||||
) = config_and_inputs
|
||||
|
||||
inputs_dict = {
|
||||
"input_ids": input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
}
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
class PegasusStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (PegasusDecoder, PegasusForCausalLM) if is_torch_available() else ()
|
||||
all_generative_model_classes = (PegasusForCausalLM,) if is_torch_available() else ()
|
||||
test_resize_position_embeddings = True
|
||||
test_pruning = False
|
||||
is_encoder_decoder = False
|
||||
|
||||
def setUp(
|
||||
self,
|
||||
):
|
||||
self.model_tester = PegasusStandaloneDecoderModelTester(self, is_training=False)
|
||||
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_decoder_model_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
|
||||
|
||||
def test_decoder_model_attn_mask_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
|
||||
|
||||
def test_retain_grad_hidden_states_attentions(self):
|
||||
# decoder cannot keep gradients
|
||||
return
|
||||
374
tests/pegasus/test_modeling_tf_pegasus.py
Normal file
374
tests/pegasus/test_modeling_tf_pegasus.py
Normal file
@@ -0,0 +1,374 @@
|
||||
# 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.
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from transformers import AutoTokenizer, PegasusConfig, is_tf_available
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import require_sentencepiece, require_tf, require_tokenizers, slow
|
||||
|
||||
from ..test_configuration_common import ConfigTester
|
||||
from ..test_modeling_tf_common import TFModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
|
||||
from transformers import TFAutoModelForSeq2SeqLM, TFPegasusForConditionalGeneration, TFPegasusModel
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFPegasusModelTester:
|
||||
config_cls = PegasusConfig
|
||||
config_updates = {}
|
||||
hidden_act = "gelu"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=40,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_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_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
|
||||
eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
|
||||
input_ids = tf.concat([input_ids, eos_tensor], axis=1)
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.config_cls(
|
||||
vocab_size=self.vocab_size,
|
||||
d_model=self.hidden_size,
|
||||
encoder_layers=self.num_hidden_layers,
|
||||
decoder_layers=self.num_hidden_layers,
|
||||
encoder_attention_heads=self.num_attention_heads,
|
||||
decoder_attention_heads=self.num_attention_heads,
|
||||
encoder_ffn_dim=self.intermediate_size,
|
||||
decoder_ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_ids=[2],
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
decoder_start_token_id=self.pad_token_id,
|
||||
**self.config_updates,
|
||||
)
|
||||
inputs_dict = prepare_pegasus_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = TFPegasusModel(config=config).get_decoder()
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
|
||||
input_ids = input_ids[:1, :]
|
||||
attention_mask = inputs_dict["attention_mask"][:1, :]
|
||||
head_mask = inputs_dict["head_mask"]
|
||||
self.batch_size = 1
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
past_key_values = past_key_values[1]
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
|
||||
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]
|
||||
|
||||
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx]
|
||||
|
||||
# test that outputs are equal for slice
|
||||
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
|
||||
|
||||
|
||||
def prepare_pegasus_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
head_mask=None,
|
||||
decoder_head_mask=None,
|
||||
cross_attn_head_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
|
||||
if decoder_attention_mask is None:
|
||||
decoder_attention_mask = tf.concat(
|
||||
[
|
||||
tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
|
||||
tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
|
||||
],
|
||||
axis=-1,
|
||||
)
|
||||
if head_mask is None:
|
||||
head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
|
||||
if decoder_head_mask is None:
|
||||
decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
||||
if cross_attn_head_mask is None:
|
||||
cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"decoder_input_ids": decoder_input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"decoder_attention_mask": decoder_attention_mask,
|
||||
"head_mask": head_mask,
|
||||
"decoder_head_mask": decoder_head_mask,
|
||||
"cross_attn_head_mask": cross_attn_head_mask,
|
||||
}
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFPegasusModelTest(TFModelTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (TFPegasusForConditionalGeneration, TFPegasusModel) if is_tf_available() else ()
|
||||
all_generative_model_classes = (TFPegasusForConditionalGeneration,) if is_tf_available() else ()
|
||||
is_encoder_decoder = True
|
||||
test_pruning = False
|
||||
test_onnx = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = TFPegasusModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PegasusConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_decoder_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_compile_tf_model(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
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")
|
||||
|
||||
model_class = self.all_generative_model_classes[0]
|
||||
input_ids = {
|
||||
"decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
|
||||
"input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
|
||||
}
|
||||
|
||||
# Prepare our model
|
||||
model = model_class(config)
|
||||
model(self._prepare_for_class(inputs_dict, model_class)) # Model must be called before saving.
|
||||
# Let's load it from the disk to be sure we can use pretrained weights
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
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 integration 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_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)
|
||||
|
||||
if model_class in self.all_generative_model_classes:
|
||||
x = model.get_output_embeddings()
|
||||
assert isinstance(x, tf.keras.layers.Layer)
|
||||
name = model.get_bias()
|
||||
assert isinstance(name, dict)
|
||||
for k, v in name.items():
|
||||
assert isinstance(v, tf.Variable)
|
||||
else:
|
||||
x = model.get_output_embeddings()
|
||||
assert x is None
|
||||
name = model.get_bias()
|
||||
assert name is None
|
||||
|
||||
def test_saved_model_creation(self):
|
||||
# This test is too long (>30sec) and makes fail the CI
|
||||
pass
|
||||
|
||||
def test_resize_token_embeddings(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
def _get_word_embedding_weight(model, embedding_layer):
|
||||
if hasattr(embedding_layer, "weight"):
|
||||
return embedding_layer.weight
|
||||
else:
|
||||
# Here we build the word embeddings weights if not exists.
|
||||
# And then we retry to get the attribute once built.
|
||||
model(model.dummy_inputs)
|
||||
if hasattr(embedding_layer, "weight"):
|
||||
return embedding_layer.weight
|
||||
else:
|
||||
return None
|
||||
|
||||
for model_class in self.all_model_classes:
|
||||
for size in [config.vocab_size - 10, config.vocab_size + 10, None]:
|
||||
# build the embeddings
|
||||
model = model_class(config=config)
|
||||
old_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
||||
old_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
||||
old_final_logits_bias = model.get_bias()
|
||||
|
||||
# reshape the embeddings
|
||||
model.resize_token_embeddings(size)
|
||||
new_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
|
||||
new_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
|
||||
new_final_logits_bias = model.get_bias()
|
||||
|
||||
# check that the resized embeddings size matches the desired size.
|
||||
assert_size = size if size is not None else config.vocab_size
|
||||
|
||||
self.assertEqual(new_input_embeddings.shape[0], assert_size)
|
||||
|
||||
# check that weights remain the same after resizing
|
||||
models_equal = True
|
||||
for p1, p2 in zip(old_input_embeddings.value(), new_input_embeddings.value()):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
if old_output_embeddings is not None and new_output_embeddings is not None:
|
||||
self.assertEqual(new_output_embeddings.shape[0], assert_size)
|
||||
|
||||
models_equal = True
|
||||
for p1, p2 in zip(old_output_embeddings.value(), new_output_embeddings.value()):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
if old_final_logits_bias is not None and new_final_logits_bias is not None:
|
||||
old_final_logits_bias = old_final_logits_bias["final_logits_bias"]
|
||||
new_final_logits_bias = new_final_logits_bias["final_logits_bias"]
|
||||
self.assertEqual(new_final_logits_bias.shape[0], 1)
|
||||
self.assertEqual(new_final_logits_bias.shape[1], assert_size)
|
||||
|
||||
models_equal = True
|
||||
for old, new in zip(old_final_logits_bias.value(), new_final_logits_bias.value()):
|
||||
for p1, p2 in zip(old, new):
|
||||
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
|
||||
models_equal = False
|
||||
self.assertTrue(models_equal)
|
||||
|
||||
|
||||
def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors not close, or a and b arent both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if tf.debugging.assert_near(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
if len(prefix) > 0:
|
||||
prefix = f"{prefix}: "
|
||||
raise AssertionError(f"{prefix}{a} != {b}")
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return tf.constant(tok_lst, dtype=tf.int32)
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
@require_tf
|
||||
class TFPegasusIntegrationTests(unittest.TestCase):
|
||||
src_text = [
|
||||
""" PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow.""",
|
||||
""" The London trio are up for best UK act and best album, as well as getting two nominations in the best song category."We got told like this morning 'Oh I think you're nominated'", said Dappy."And I was like 'Oh yeah, which one?' And now we've got nominated for four awards. I mean, wow!"Bandmate Fazer added: "We thought it's best of us to come down and mingle with everyone and say hello to the cameras. And now we find we've got four nominations."The band have two shots at the best song prize, getting the nod for their Tynchy Stryder collaboration Number One, and single Strong Again.Their album Uncle B will also go up against records by the likes of Beyonce and Kanye West.N-Dubz picked up the best newcomer Mobo in 2007, but female member Tulisa said they wouldn't be too disappointed if they didn't win this time around."At the end of the day we're grateful to be where we are in our careers."If it don't happen then it don't happen - live to fight another day and keep on making albums and hits for the fans."Dappy also revealed they could be performing live several times on the night.The group will be doing Number One and also a possible rendition of the War Child single, I Got Soul.The charity song is a re-working of The Killers' All These Things That I've Done and is set to feature artists like Chipmunk, Ironik and Pixie Lott.This year's Mobos will be held outside of London for the first time, in Glasgow on 30 September.N-Dubz said they were looking forward to performing for their Scottish fans and boasted about their recent shows north of the border."We just done Edinburgh the other day," said Dappy."We smashed up an N-Dubz show over there. We done Aberdeen about three or four months ago - we smashed up that show over there! Everywhere we go we smash it up!" """,
|
||||
]
|
||||
expected_text = [
|
||||
"California's largest electricity provider has cut power to hundreds of thousands of customers in an effort to reduce the risk of wildfires.",
|
||||
'N-Dubz have revealed they\'re "grateful" to have been nominated for four Mobo Awards.',
|
||||
] # differs slightly from pytorch, likely due to numerical differences in linear layers
|
||||
model_name = "google/pegasus-xsum"
|
||||
|
||||
@cached_property
|
||||
def tokenizer(self):
|
||||
return AutoTokenizer.from_pretrained(self.model_name)
|
||||
|
||||
@cached_property
|
||||
def model(self):
|
||||
model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
|
||||
return model
|
||||
|
||||
def _assert_generated_batch_equal_expected(self, **tokenizer_kwargs):
|
||||
generated_words = self.translate_src_text(**tokenizer_kwargs)
|
||||
assert self.expected_text == generated_words
|
||||
|
||||
def translate_src_text(self, **tokenizer_kwargs):
|
||||
model_inputs = self.tokenizer(self.src_text, **tokenizer_kwargs, padding=True, return_tensors="tf")
|
||||
generated_ids = self.model.generate(
|
||||
model_inputs.input_ids,
|
||||
attention_mask=model_inputs.attention_mask,
|
||||
num_beams=2,
|
||||
use_cache=True,
|
||||
)
|
||||
generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)
|
||||
return generated_words
|
||||
|
||||
@slow
|
||||
def test_batch_generation(self):
|
||||
self._assert_generated_batch_equal_expected()
|
||||
208
tests/pegasus/test_tokenization_pegasus.py
Normal file
208
tests/pegasus/test_tokenization_pegasus.py
Normal file
@@ -0,0 +1,208 @@
|
||||
# Copyright 2020 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 PegasusTokenizer, PegasusTokenizerFast
|
||||
from transformers.file_utils import cached_property
|
||||
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, require_torch, slow
|
||||
|
||||
from ..test_tokenization_common import TokenizerTesterMixin
|
||||
|
||||
|
||||
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece_no_bos.model")
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class PegasusTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
|
||||
tokenizer_class = PegasusTokenizer
|
||||
rust_tokenizer_class = PegasusTokenizerFast
|
||||
test_rust_tokenizer = True
|
||||
test_sentencepiece = True
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
|
||||
# We have a SentencePiece fixture for testing
|
||||
tokenizer = PegasusTokenizer(SAMPLE_VOCAB)
|
||||
tokenizer.save_pretrained(self.tmpdirname)
|
||||
|
||||
@cached_property
|
||||
def _large_tokenizer(self):
|
||||
return PegasusTokenizer.from_pretrained("google/pegasus-large")
|
||||
|
||||
def get_tokenizer(self, **kwargs) -> PegasusTokenizer:
|
||||
return PegasusTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_input_output_texts(self, tokenizer):
|
||||
return ("This is a test", "This is a test")
|
||||
|
||||
def test_convert_token_and_id(self):
|
||||
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
|
||||
token = "</s>"
|
||||
token_id = 1
|
||||
|
||||
self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
|
||||
self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
|
||||
|
||||
def test_get_vocab(self):
|
||||
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
|
||||
|
||||
self.assertEqual(vocab_keys[0], "<pad>")
|
||||
self.assertEqual(vocab_keys[1], "</s>")
|
||||
self.assertEqual(vocab_keys[-1], "v")
|
||||
self.assertEqual(len(vocab_keys), 1_103)
|
||||
|
||||
def test_vocab_size(self):
|
||||
self.assertEqual(self.get_tokenizer().vocab_size, 1_103)
|
||||
|
||||
def test_mask_tokens_rust_pegasus(self):
|
||||
rust_tokenizer = self.rust_tokenizer_class.from_pretrained(self.tmpdirname)
|
||||
py_tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname)
|
||||
raw_input_str = "Let's see which <unk> is the better <unk_token_11> one <mask_1> It seems like this <mask_2> was important </s> <pad> <pad> <pad>"
|
||||
rust_ids = rust_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
|
||||
py_ids = py_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
|
||||
self.assertListEqual(py_ids, rust_ids)
|
||||
|
||||
def test_large_mask_tokens(self):
|
||||
tokenizer = self._large_tokenizer
|
||||
# <mask_1> masks whole sentence while <mask_2> masks single word
|
||||
raw_input_str = "<mask_1> To ensure a <mask_2> flow of bank resolutions."
|
||||
desired_result = [2, 413, 615, 114, 3, 1971, 113, 1679, 10710, 107, 1]
|
||||
ids = tokenizer([raw_input_str], return_tensors=None).input_ids[0]
|
||||
self.assertListEqual(desired_result, ids)
|
||||
|
||||
def test_large_tokenizer_settings(self):
|
||||
tokenizer = self._large_tokenizer
|
||||
# The tracebacks for the following asserts are **better** without messages or self.assertEqual
|
||||
assert tokenizer.vocab_size == 96103
|
||||
assert tokenizer.pad_token_id == 0
|
||||
assert tokenizer.eos_token_id == 1
|
||||
assert tokenizer.offset == 103
|
||||
assert tokenizer.unk_token_id == tokenizer.offset + 2 == 105
|
||||
assert tokenizer.unk_token == "<unk>"
|
||||
assert tokenizer.model_max_length == 1024
|
||||
raw_input_str = "To ensure a smooth flow of bank resolutions."
|
||||
desired_result = [413, 615, 114, 2291, 1971, 113, 1679, 10710, 107, 1]
|
||||
ids = tokenizer([raw_input_str], return_tensors=None).input_ids[0]
|
||||
self.assertListEqual(desired_result, ids)
|
||||
assert tokenizer.convert_ids_to_tokens([0, 1, 2, 3]) == ["<pad>", "</s>", "<mask_1>", "<mask_2>"]
|
||||
|
||||
@require_torch
|
||||
def test_large_seq2seq_truncation(self):
|
||||
src_texts = ["This is going to be way too long." * 150, "short example"]
|
||||
tgt_texts = ["not super long but more than 5 tokens", "tiny"]
|
||||
batch = self._large_tokenizer(src_texts, padding=True, truncation=True, return_tensors="pt")
|
||||
with self._large_tokenizer.as_target_tokenizer():
|
||||
targets = self._large_tokenizer(
|
||||
tgt_texts, max_length=5, padding=True, truncation=True, return_tensors="pt"
|
||||
)
|
||||
|
||||
assert batch.input_ids.shape == (2, 1024)
|
||||
assert batch.attention_mask.shape == (2, 1024)
|
||||
assert targets["input_ids"].shape == (2, 5)
|
||||
assert len(batch) == 2 # input_ids, attention_mask.
|
||||
|
||||
@slow
|
||||
def test_tokenizer_integration(self):
|
||||
# fmt: off
|
||||
expected_encoding = {'input_ids': [[38979, 143, 18485, 606, 130, 26669, 87686, 121, 54189, 1129, 111, 26669, 87686, 121, 9114, 14787, 121, 13249, 158, 592, 956, 121, 14621, 31576, 143, 62613, 108, 9688, 930, 43430, 11562, 62613, 304, 108, 11443, 897, 108, 9314, 17415, 63399, 108, 11443, 7614, 18316, 118, 4284, 7148, 12430, 143, 1400, 25703, 158, 111, 4284, 7148, 11772, 143, 21297, 1064, 158, 122, 204, 3506, 1754, 1133, 14787, 1581, 115, 33224, 4482, 111, 1355, 110, 29173, 317, 50833, 108, 20147, 94665, 111, 77198, 107, 1], [110, 62613, 117, 638, 112, 1133, 121, 20098, 1355, 79050, 13872, 135, 1596, 53541, 1352, 141, 13039, 5542, 124, 302, 518, 111, 268, 2956, 115, 149, 4427, 107, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [139, 1235, 2799, 18289, 17780, 204, 109, 9474, 1296, 107, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], 'attention_mask': [[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, 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, 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, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]} # noqa: E501
|
||||
# fmt: on
|
||||
|
||||
self.tokenizer_integration_test_util(
|
||||
expected_encoding=expected_encoding,
|
||||
model_name="google/bigbird-pegasus-large-arxiv",
|
||||
revision="ba85d0851d708441f91440d509690f1ab6353415",
|
||||
)
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class BigBirdPegasusTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
|
||||
tokenizer_class = PegasusTokenizer
|
||||
rust_tokenizer_class = PegasusTokenizerFast
|
||||
test_rust_tokenizer = True
|
||||
test_sentencepiece = True
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
|
||||
# We have a SentencePiece fixture for testing
|
||||
tokenizer = PegasusTokenizer(SAMPLE_VOCAB, offset=0, mask_token_sent=None, mask_token="[MASK]")
|
||||
tokenizer.save_pretrained(self.tmpdirname)
|
||||
|
||||
@cached_property
|
||||
def _large_tokenizer(self):
|
||||
return PegasusTokenizer.from_pretrained("google/bigbird-pegasus-large-arxiv")
|
||||
|
||||
def get_tokenizer(self, **kwargs) -> PegasusTokenizer:
|
||||
return PegasusTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_input_output_texts(self, tokenizer):
|
||||
return ("This is a test", "This is a test")
|
||||
|
||||
def test_mask_tokens_rust_pegasus(self):
|
||||
rust_tokenizer = self.rust_tokenizer_class.from_pretrained(self.tmpdirname)
|
||||
py_tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname)
|
||||
raw_input_str = "Let's see which <unk> is the better <unk_token> one [MASK] It seems like this [MASK] was important </s> <pad> <pad> <pad>"
|
||||
rust_ids = rust_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
|
||||
py_ids = py_tokenizer([raw_input_str], return_tensors=None, add_special_tokens=False).input_ids[0]
|
||||
self.assertListEqual(py_ids, rust_ids)
|
||||
|
||||
@require_torch
|
||||
def test_large_seq2seq_truncation(self):
|
||||
src_texts = ["This is going to be way too long." * 1000, "short example"]
|
||||
tgt_texts = ["not super long but more than 5 tokens", "tiny"]
|
||||
batch = self._large_tokenizer(src_texts, padding=True, truncation=True, return_tensors="pt")
|
||||
with self._large_tokenizer.as_target_tokenizer():
|
||||
targets = self._large_tokenizer(
|
||||
tgt_texts, max_length=5, padding=True, truncation=True, return_tensors="pt"
|
||||
)
|
||||
|
||||
assert batch.input_ids.shape == (2, 4096)
|
||||
assert batch.attention_mask.shape == (2, 4096)
|
||||
assert targets["input_ids"].shape == (2, 5)
|
||||
assert len(batch) == 2 # input_ids, attention_mask.
|
||||
|
||||
def test_equivalence_to_orig_tokenizer(self):
|
||||
"""
|
||||
To run with original TF tokenizer:
|
||||
|
||||
!wget https://github.com/google-research/bigbird/raw/master/bigbird/vocab/pegasus.model
|
||||
!pip install tensorflow-text
|
||||
|
||||
import tensorflow.compat.v2 as tf
|
||||
import tensorflow_text as tft
|
||||
|
||||
VOCAB_FILE = "./pegasus.model"
|
||||
|
||||
tf.enable_v2_behavior()
|
||||
|
||||
test_str = "This is an example string that is used to test the original TF implementation against the HF implementation"
|
||||
tokenizer = tft.SentencepieceTokenizer(model=tf.io.gfile.GFile(VOCAB_FILE, "rb").read())
|
||||
|
||||
tokenizer.tokenize(test_str)
|
||||
"""
|
||||
|
||||
test_str = "This is an example string that is used to test the original TF implementation against the HF implementation"
|
||||
|
||||
token_ids = self._large_tokenizer(test_str).input_ids
|
||||
|
||||
self.assertListEqual(
|
||||
token_ids,
|
||||
[182, 117, 142, 587, 4211, 120, 117, 263, 112, 804, 109, 856, 25016, 3137, 464, 109, 26955, 3137, 1],
|
||||
)
|
||||
Reference in New Issue
Block a user