def test_drop_last_works(self):
sampler = BucketBatchSampler(
batch_size=2,
padding_noise=0,
sorting_keys=["text"],
drop_last=True,
)
# We use a custom collate_fn for testing, which doesn't actually create tensors,
# just the allennlp Batches.
def collate_fn(x, **kwargs):
return Batch(x)
data_loader = MultiProcessDataLoader(
self.get_mock_reader(),
"fake_path",
batch_sampler=sampler,
)
data_loader.collate_fn = collate_fn
data_loader.index_with(self.vocab)
batches = [batch for batch in iter(data_loader)]
stats = self.get_batches_stats(batches)
# all batches have length batch_size
assert all(batch_len == 2 for batch_len in stats["batch_lengths"])
# we should have lost one instance by skipping the last batch
assert stats["total_instances"] == len(self.instances) - 1
def read_and_check_instances(self, filepath: str, num_workers: int = 0):
data_loader = MultiProcessDataLoader(self.reader,
filepath,
num_workers=num_workers,
batch_size=1,
start_method="spawn")
all_instances = []
for instance in data_loader.iter_instances():
all_instances.append(instance)
# 100 files * 4 sentences / file
assert len(all_instances) == 100 * 4
counts = Counter(fingerprint(instance) for instance in all_instances)
# should have the exact same data 100 times
assert len(counts) == 4
assert counts[("cats", "are", "animals", ".", "N", "V", "N",
"N")] == 100
assert counts[("dogs", "are", "animals", ".", "N", "V", "N",
"N")] == 100
assert counts[("snakes", "are", "animals", ".", "N", "V", "N",
"N")] == 100
assert counts[("birds", "are", "animals", ".", "N", "V", "N",
"N")] == 100
def test_language_model_data_collator():
"""
Ensure `LanguageModelingDataCollator` works
"""
norm_loader = MultiProcessDataLoader(MockDatasetReader(),
"some path",
batch_size=16)
vocab = Vocabulary.from_instances(norm_loader.iter_instances())
norm_loader.index_with(vocab)
batch0 = list(norm_loader)[0]
model_name = "epwalsh/bert-xsmall-dummy"
data_collate = LanguageModelingDataCollator(model_name)
mlm_loader = MultiProcessDataLoader(MockDatasetReader(),
"some path",
batch_size=16,
collate_fn=data_collate)
vocab = Vocabulary.from_instances(mlm_loader.iter_instances())
mlm_loader.index_with(vocab)
batch1 = list(mlm_loader)[0]
norm_inputs = batch0["source"]["tokens"]["token_ids"]
mlm_inputs = batch1["source"]["tokens"]["token_ids"]
mlm_labels = batch1["source"]["tokens"]["labels"]
# if we replace the mlm inputs with their labels, should be same as origin inputs
assert torch.where(mlm_labels != -100, mlm_labels,
mlm_inputs).tolist() == norm_inputs.tolist()
def build_data_loaders(config,
dataset_reader: DatasetReader) -> Tuple[MultiProcessDataLoader, MultiProcessDataLoader, MultiProcessDataLoader]:
train_loader = MultiProcessDataLoader(dataset_reader, data_path='train', batch_size=config.batch_size_for_train, shuffle=False)
dev_loader = MultiProcessDataLoader(dataset_reader, data_path='dev', batch_size=config.batch_size_for_eval, shuffle=False)
test_loader = MultiProcessDataLoader(dataset_reader, data_path='test', batch_size=config.batch_size_for_eval, shuffle=False)
return train_loader, dev_loader, test_loader
def test_batch_count(self):
sampler = BucketBatchSampler(batch_size=2,
padding_noise=0,
sorting_keys=["text"])
data_loader = MultiProcessDataLoader(self.get_mock_reader(),
"fake_path",
batch_sampler=sampler)
data_loader.index_with(self.vocab)
assert len(data_loader) == 3
def main():
reader = StanfordSentimentTreeBankDatasetReader()
train_path = 'https://s3.amazonaws.com/realworldnlpbook/data/stanfordSentimentTreebank/trees/train.txt'
dev_path = 'https://s3.amazonaws.com/realworldnlpbook/data/stanfordSentimentTreebank/trees/dev.txt'
sampler = BucketBatchSampler(batch_size=32, sorting_keys=["tokens"])
train_data_loader = MultiProcessDataLoader(reader,
train_path,
batch_sampler=sampler)
dev_data_loader = MultiProcessDataLoader(reader,
dev_path,
batch_sampler=sampler)
# You can optionally specify the minimum count of tokens/labels.
# `min_count={'tokens':3}` here means that any tokens that appear less than three times
# will be ignored and not included in the vocabulary.
vocab = Vocabulary.from_instances(chain(train_data_loader.iter_instances(),
dev_data_loader.iter_instances()),
min_count={'tokens': 3})
train_data_loader.index_with(vocab)
dev_data_loader.index_with(vocab)
token_embedding = Embedding(num_embeddings=vocab.get_vocab_size('tokens'),
embedding_dim=EMBEDDING_DIM)
# BasicTextFieldEmbedder takes a dict - we need an embedding just for tokens,
# not for labels, which are used as-is as the "answer" of the sentence classification
word_embeddings = BasicTextFieldEmbedder({"tokens": token_embedding})
# Seq2VecEncoder is a neural network abstraction that takes a sequence of something
# (usually a sequence of embedded word vectors), processes it, and returns a single
# vector. Oftentimes this is an RNN-based architecture (e.g., LSTM or GRU), but
# AllenNLP also supports CNNs and other simple architectures (for example,
# just averaging over the input vectors).
encoder = PytorchSeq2VecWrapper(
torch.nn.LSTM(EMBEDDING_DIM, HIDDEN_DIM, batch_first=True))
model = LstmClassifier(word_embeddings, encoder, vocab)
optimizer = optim.Adam(model.parameters(), lr=1e-4, weight_decay=1e-5)
trainer = GradientDescentTrainer(model=model,
optimizer=optimizer,
data_loader=train_data_loader,
validation_data_loader=dev_data_loader,
patience=10,
num_epochs=20,
cuda_device=-1)
trainer.train()
predictor = SentenceClassifierPredictor(model, dataset_reader=reader)
logits = predictor.predict('This is the best movie ever!')['logits']
label_id = np.argmax(logits)
print(model.vocab.get_token_from_index(label_id, 'labels'))
def test_batches_per_epoch():
loader = MultiProcessDataLoader(MockDatasetReader(),
"some path",
batch_size=4,
batches_per_epoch=10)
vocab = Vocabulary.from_instances(loader.iter_instances())
loader.index_with(vocab)
assert len(loader) == 10
assert len(list(loader)) == 10
def test_load_to_cuda(options):
reader = MockDatasetReader()
loader = MultiProcessDataLoader(
reader=reader,
data_path="this doens't matter",
cuda_device=0,
**options,
)
vocab = Vocabulary.from_instances(loader.iter_instances())
loader.index_with(vocab)
for batch in loader:
assert batch["tensor"].device == torch.device("cuda:0")
def build_data_loaders(
reader,
train_data_path: str,
validation_data_path: str,
) -> Tuple[DataLoader, DataLoader]:
train_loader = MultiProcessDataLoader(reader,
train_data_path,
batch_size=8,
shuffle=True)
dev_loader = MultiProcessDataLoader(reader,
validation_data_path,
batch_size=8,
shuffle=False)
return train_loader, dev_loader
def test_batch_count(self):
sampler = MaxTokensBatchSampler(max_tokens=8,
padding_noise=0,
sorting_keys=["text"])
data_loader = MultiProcessDataLoader(self.get_mock_reader(),
"fake_path",
batch_sampler=sampler)
assert len(data_loader) == 3
class TrainerTestBase(AllenNlpTestCase):
def setup_method(self):
super().setup_method()
self.data_path = str(self.FIXTURES_ROOT / "data" / "sequence_tagging.tsv")
self.reader = SequenceTaggingDatasetReader()
self.data_loader = MultiProcessDataLoader(self.reader, self.data_path, batch_size=2)
self.data_loader_lazy = MultiProcessDataLoader(
self.reader, self.data_path, batch_size=2, max_instances_in_memory=10
)
self.instances = list(self.data_loader.iter_instances())
self.vocab = Vocabulary.from_instances(self.instances)
self.data_loader.index_with(self.vocab)
self.data_loader_lazy.index_with(self.vocab)
self.model_params = Params(
{
"text_field_embedder": {
"token_embedders": {"tokens": {"type": "embedding", "embedding_dim": 5}}
},
"encoder": {"type": "lstm", "input_size": 5, "hidden_size": 7, "num_layers": 2},
}
)
self.model = SimpleTagger.from_params(vocab=self.vocab, params=self.model_params)
self.optimizer = torch.optim.SGD(self.model.parameters(), 0.01, momentum=0.9)
self.validation_data_loader = MultiProcessDataLoader(
self.reader, self.data_path, batch_size=2
)
self.validation_data_loader.index_with(self.vocab)
def test_batch_count_with_drop_last(self):
sampler = BucketBatchSampler(
batch_size=2,
padding_noise=0,
sorting_keys=["text"],
drop_last=True,
)
data_loader = MultiProcessDataLoader(self.get_mock_reader(),
"fake_path",
batch_sampler=sampler)
assert len(data_loader) == 2
def test_with_multi_process_loading(self, lazy):
readers = {"a": PlainTextReader(), "b": PlainTextReader(), "c": PlainTextReader()}
reader = InterleavingDatasetReader(readers)
data_dir = self.FIXTURES_ROOT / "data"
file_path = {
"a": data_dir / "babi.txt",
"b": data_dir / "conll2003.txt",
"c": data_dir / "conll2003.txt",
}
vocab = Vocabulary.from_instances(reader.read(file_path))
loader = MultiProcessDataLoader(
reader,
file_path,
num_workers=1,
batch_size=1,
max_instances_in_memory=2 if lazy else None,
)
loader.index_with(vocab)
list(loader.iter_instances())
list(loader)
def test_error_raised_when_text_fields_contain_token_indexers(
max_instances_in_memory):
"""
This tests that the MultiProcessDataLoader raises an error when num_workers > 0
but the dataset reader doesn't implement apply_token_indexers().
It also tests that errors raised within a worker process are propogated upwards
to the main process, and that when that happens, all workers will be successfully
killed.
"""
with pytest.raises(
WorkerError,
match="Make sure your dataset reader's text_to_instance()"):
loader = MultiProcessDataLoader(
MockOldDatasetReader(),
"this-path-doesn't-matter",
num_workers=2,
max_instances_in_memory=max_instances_in_memory,
batch_size=1,
)
list(loader.iter_instances())
def test_multiprocess_data_loader(options):
reader = MockDatasetReader()
data_path = "this doesn't matter"
loader = MultiProcessDataLoader(reader=reader,
data_path=data_path,
**options)
if not options.get("max_instances_in_memory"):
# Instances should be loaded immediately if max_instances_in_memory is None.
assert loader._instances
instances: Iterable[Instance] = loader.iter_instances()
# This should be a generator.
assert not isinstance(instances, (list, tuple))
instances = list(instances)
assert len(instances) == MockDatasetReader.NUM_INSTANCES
# Now build vocab.
vocab = Vocabulary.from_instances(instances)
# Before indexing the loader, trying to iterate through batches will raise an error.
with pytest.raises(ValueError,
match="Did you forget to call DataLoader.index_with"):
list(loader)
loader.index_with(vocab)
# Run through a couple epochs to make sure we collect all of the instances.
for epoch in range(2):
indices: List[int] = []
for batch in loader:
for index in batch["index"]:
indices.append(index) # type: ignore
# Ensure no duplicates.
assert len(indices) == len(set(indices)), indices
# Ensure all collected.
assert len(indices) == MockDatasetReader.NUM_INSTANCES, epoch
def build_data_loader(data_reader: DatasetReader,
data_path: Path,
batch_size: int,
shuffle: bool = True) -> DataLoader:
"""
Build an AllenNLP DataLoader.
:param train_data: The training dataset, torch object.
:param dev_data: The dev dataset, torch object.
:return train_loader, dev_loader: The train and dev data loaders as a
tuple.
"""
# Note that DataLoader is imported from allennlp above, *not* torch.
# We need to get the allennlp-specific collate function, which is
# what actually does indexing and batching.
# log.debug("Building DataLoader.")
loader = MultiProcessDataLoader(reader=data_reader,
data_path=data_path,
batch_size=batch_size,
shuffle=shuffle)
# log.debug("DataLoader built.")
return loader
def test_drop_last():
"""
Ensures that the `drop_last` option is respected.
"""
loader = MultiProcessDataLoader(MockDatasetReader(),
"some path",
batch_size=16,
drop_last=True)
vocab = Vocabulary.from_instances(loader.iter_instances())
loader.index_with(vocab)
# Should still load all instances. `drop_last` only affects batches.
assert len(list(
loader.iter_instances())) == MockDatasetReader.NUM_INSTANCES
# Just here because the assertions below depend on the exact value of NUM_INSTANCES.
assert MockDatasetReader.NUM_INSTANCES == 100
batches = list(loader)
for batch in batches:
assert len(batch["index"]) == 16
assert len(batches) == 6
def evaluate_transformers_checkpoint(
data_path: str,
model_config_path: str,
checkpoint_model_name: str,
checkpoint_tokenizer_name: str,
batch_size: int,
cuda_device: int,
result_save_path: str,
):
"""
Expected results for ``test.json`` from the Open Entity dataset:
{'micro_precision': 0.7997806072235107, 'micro_recall': 0.7657563090324402, 'micro_fscore': 0.7823987007141113}.
Parameters
----------
data_path : str
Data path to the input file.
model_config_path : str
A config file that defines the model architecture to evaluate.
checkpoint_model_name : str
The name of the checkpoint in Hugging Face Model Hub.
checkpoint_tokenizer_name : str
This should be the name of the base pre-training model because sometimes
the tokenizer of downstream task is not compatible with allennlp.
batch_size : int
cuda_device : int
result_save_path : str
"""
import_module_and_submodules("examples_allennlp")
tokenizer_kwargs = {"additional_special_tokens": [ENT]}
reader = EntityTypingReader(
tokenizer=PretrainedTransformerTokenizer(
model_name=checkpoint_tokenizer_name,
add_special_tokens=True,
tokenizer_kwargs=tokenizer_kwargs),
token_indexers={
"tokens":
PretrainedTransformerIndexer(model_name=checkpoint_tokenizer_name,
tokenizer_kwargs=tokenizer_kwargs)
},
use_entity_feature=True,
)
transformers_tokenizer = LukeTokenizer.from_pretrained(
checkpoint_model_name)
transformers_model = LukeForEntityClassification.from_pretrained(
checkpoint_model_name)
vocab = Vocabulary()
vocab.add_transformer_vocab(transformers_tokenizer, "tokens")
num_labels = len(transformers_model.config.id2label)
labels = [transformers_model.config.id2label[i] for i in range(num_labels)]
vocab.add_tokens_to_namespace(labels, namespace="labels")
# read model
params = Params.from_file(
model_config_path,
ext_vars={"TRANSFORMERS_MODEL_NAME": checkpoint_model_name})
model = Model.from_params(params, vocab=vocab)
model.classifier = transformers_model.classifier
model.eval()
# set the GPU device to use
if cuda_device < 0:
device = torch.device("cpu")
else:
device = torch.device(f"cuda:{cuda_device}")
model = model.to(device)
loader = MultiProcessDataLoader(reader,
data_path,
batch_size=batch_size,
shuffle=False)
loader.index_with(model.vocab)
with torch.no_grad():
for batch in tqdm.tqdm(loader):
batch = nn_util.move_to_device(batch, device)
output_dict = model(**batch)
metrics = model.get_metrics(reset=True)
print(metrics)
if result_save_path is not None:
with open(result_save_path, "w") as f:
json.dump(metrics, f)
def evaluate_transformers_checkpoint(
data_path: str,
model_config_path: str,
checkpoint_model_name: str,
checkpoint_tokenizer_name: str,
batch_size: int,
cuda_device: int,
result_save_path: str,
prediction_save_path: str,
):
"""
Expected results for CoNLL-2003 NER English test set.
{'f1': 0.9461946902654867, 'precision': 0.945859872611465, 'recall': 0.9465297450424929}
Parameters
----------
data_path : str
Data path to the input file.
model_config_path : str
A config file that defines the model architecture to evaluate.
checkpoint_model_name : str
The name of the checkpoint in Hugging Face Model Hub.
checkpoint_tokenizer_name : str
This should be the name of the base pre-training model because sometimes
the tokenizer of downstream task is not compatible with allennlp.
batch_size : int
cuda_device : int
result_save_path : str
"""
import_module_and_submodules("examples_allennlp")
reader = ConllSpanReader(
tokenizer=PretrainedTransformerTokenizer(
model_name=checkpoint_tokenizer_name,
add_special_tokens=False,
tokenizer_kwargs={"add_prefix_space": True}),
token_indexers={
"tokens":
PretrainedTransformerIndexer(model_name=checkpoint_tokenizer_name)
},
use_entity_feature=True,
)
transformers_tokenizer = LukeTokenizer.from_pretrained(
checkpoint_model_name)
transformers_model = LukeForEntitySpanClassification.from_pretrained(
checkpoint_model_name)
vocab = Vocabulary()
vocab.add_transformer_vocab(transformers_tokenizer, "tokens")
num_labels = len(transformers_model.config.id2label)
labels = [transformers_model.config.id2label[i] for i in range(num_labels)]
labels = ["O" if l == "NIL" else l for l in labels]
vocab.add_tokens_to_namespace(labels, namespace="labels")
# read model
params = Params.from_file(
model_config_path,
ext_vars={"TRANSFORMERS_MODEL_NAME": checkpoint_model_name})
if prediction_save_path is not None:
params["prediction_save_path"] = prediction_save_path
model = Model.from_params(params, vocab=vocab)
model.classifier = transformers_model.classifier
model.eval()
# set the GPU device to use
if cuda_device < 0:
device = torch.device("cpu")
else:
device = torch.device(f"cuda:{cuda_device}")
model = model.to(device)
loader = MultiProcessDataLoader(reader,
data_path,
batch_size=batch_size,
shuffle=False)
loader.index_with(model.vocab)
with torch.no_grad():
for batch in tqdm.tqdm(loader):
batch = nn_util.move_to_device(batch, device)
output_dict = model(**batch)
metrics = model.get_metrics(reset=True)
print(metrics)
if result_save_path is not None:
with open(result_save_path, "w") as f:
json.dump(metrics, f)