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

[PyTorch Bart] Split Bart into different models (#9343)

Browse Source 
* first try

* remove old template

* finish bart

* finish mbart

* delete unnecessary line

* init pegasus

* save intermediate

* correct pegasus

* finish pegasus

* remove cookie cutter leftover

* add marian

* finish blenderbot

* replace in file

* correctly split blenderbot

* delete "old" folder

* correct "add statement"

* adapt config for tf comp

* correct configs for tf

* remove ipdb

* fix more stuff

* fix mbart

* push pegasus fix

* fix mbart

* more fixes

* fix research projects code

* finish docs for bart, mbart, and marian

* delete unnecessary file

* correct attn typo

* correct configs

* remove pegasus for seq class

* correct peg docs

* correct peg docs

* finish configs

* further improve docs

* add copied from statements to mbart

* fix copied from in mbart

* add copy statements to marian

* add copied from to marian

* add pegasus copied from

* finish pegasus

* finish copied from

* Apply suggestions from code review

* make style

* backward comp blenderbot

* apply lysandres and sylvains suggestions

* apply suggestions

* push last fixes

* fix docs

* fix tok tests

* fix imports code style

* fix doc
This commit is contained in:
Patrick von Platen
2021-01-05 22:00:05 +01:00
committed by GitHub
parent 4eec5d0cf6
commit eef66035a2
59 changed files with 9273 additions and 2271 deletions
Download Patch File Download Diff File Expand all files Collapse all files

287
tests/test_modeling_mbart.py
View File

@@ -1,4 +1,5 @@
# Copyright 2020 The HuggingFace Team. All rights reserved.
# 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.
@@ -11,61 +12,294 @@
# 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 MBART model. """
import copy
import tempfile
import unittest
from transformers import 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 .test_modeling_common import ModelTesterMixin
from .test_configuration_common import ConfigTester
from .test_generation_utils import GenerationTesterMixin
from .test_modeling_common import ModelTesterMixin, ids_tensor
if is_torch_available():
import torch
from transformers import (
AutoModelForSeq2SeqLM,
AutoTokenizer,
BatchEncoding,
MBartConfig,
MBartForConditionalGeneration,
MBartForQuestionAnswering,
MBartForSequenceClassification,
MBartModel,
)
from transformers.models.mbart.modeling_mbart import MBartDecoder, MBartEncoder
EN_CODE = 250004
RO_CODE = 250020
def prepare_mbart_inputs_dict(
config,
input_ids,
decoder_input_ids,
attention_mask=None,
decoder_attention_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)
return {
"input_ids": input_ids,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": attention_mask,
}
@require_torch
class ModelTester:
def __init__(self, parent):
self.config = MBartConfig(
vocab_size=99,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
add_final_layer_norm=True,
class MBartModelTester:
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 = MBartConfig(
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,
)
inputs_dict = prepare_mbart_inputs_dict(config, input_ids, decoder_input_ids)
return config, inputs_dict
def prepare_config_and_inputs_for_common(self):
return self.config, {}
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 = MBartModel(config=config).get_decoder().to(torch_device).eval()
input_ids = inputs_dict["input_ids"]
attention_mask = inputs_dict["attention_mask"]
# first forward pass
outputs = model(input_ids, attention_mask=attention_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-2))
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
model = MBartModel(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 = MBartEncoder.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 = MBartDecoder.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 SelectiveCommonTest(unittest.TestCase):
all_model_classes = (MBartForConditionalGeneration, MBartModel) if is_torch_available() else ()
test_save_load__keys_to_ignore_on_save = ModelTesterMixin.test_save_load__keys_to_ignore_on_save
class MBartModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
all_model_classes = (
(MBartModel, MBartForConditionalGeneration, MBartForSequenceClassification, MBartForQuestionAnswering)
if is_torch_available()
else ()
)
all_generative_model_classes = (MBartForConditionalGeneration,) if is_torch_available() else ()
is_encoder_decoder = True
test_pruning = False
test_head_masking = False
test_missing_keys = False
def setUp(self):
self.model_tester = ModelTester(self)
self.model_tester = MBartModelTester(self)
self.config_tester = ConfigTester(self, config_class=MBartConfig)
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)
# MBartForSequenceClassification does not support inputs_embeds
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in (MBartModel, MBartForConditionalGeneration, MBartForQuestionAnswering):
model = model_class(config)
model.to(torch_device)
model.eval()
inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
if not self.is_encoder_decoder:
input_ids = inputs["input_ids"]
del inputs["input_ids"]
else:
encoder_input_ids = inputs["input_ids"]
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
del inputs["input_ids"]
inputs.pop("decoder_input_ids", None)
wte = model.get_input_embeddings()
if not self.is_encoder_decoder:
inputs["inputs_embeds"] = wte(input_ids)
else:
inputs["inputs_embeds"] = wte(encoder_input_ids)
inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
with torch.no_grad():
model(**inputs)[0]
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 = MBartForConditionalGeneration(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
@@ -83,7 +317,7 @@ class AbstractSeq2SeqIntegrationTest(unittest.TestCase):
@cached_property
def model(self):
"""Only load the model if needed."""
model = AutoModelForSeq2SeqLM.from_pretrained(self.checkpoint_name).to(torch_device)
model = MBartForConditionalGeneration.from_pretrained(self.checkpoint_name).to(torch_device)
if "cuda" in torch_device:
model = model.half()
return model
@@ -102,7 +336,7 @@ class MBartEnroIntegrationTest(AbstractSeq2SeqIntegrationTest):
"Şeful ONU declară că nu există o soluţie militară în Siria",
'Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu vor face decât să înrăutăţească violenţa şi mizeria pentru milioane de oameni.',
]
expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, EN_CODE]
expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, 250004]
@slow
def test_enro_generate_one(self):
@@ -128,7 +362,6 @@ class MBartEnroIntegrationTest(AbstractSeq2SeqIntegrationTest):
expected = {"scale_embedding": True, "output_past": True}
for name in mbart_models:
config = MBartConfig.from_pretrained(name)
self.assertTrue(config.is_valid_mbart())
for k, v in expected.items():
try:
self.assertEqual(v, getattr(config, k))