def test_dpr_save_load():
d = {
'query':
'big little lies season 2 how many episodes',
'passages': [{
'title': 'Big Little Lies (TV series)',
'text':
'series garnered several accolades. It received 16 Emmy Award nominations and won eight, including Outstanding Limited Series and acting awards for Kidman, Skarsgård, and Dern. The trio also won Golden Globe Awards in addition to a Golden Globe Award for Best Miniseries or Television Film win for the series. Kidman and Skarsgård also received Screen Actors Guild Awards for their performances. Despite originally being billed as a miniseries, HBO renewed the series for a second season. Production on the second season began in March 2018 and is set to premiere in 2019. All seven episodes are being written by Kelley',
'label': 'positive',
'external_id': '18768923'
}, {
'title': 'Little People, Big World',
'text':
'final minutes of the season two-A finale, "Farm Overload". A crowd had gathered around Jacob, who was lying on the ground near the trebuchet. The first two episodes of season two-B focus on the accident, and how the local media reacted to it. The first season of "Little People, Big World" generated solid ratings for TLC (especially in the important 18–49 demographic), leading to the show\'s renewal for a second season. Critical reviews of the series have been generally positive, citing the show\'s positive portrayal of little people. Conversely, other reviews have claimed that the show has a voyeuristic bend',
'label': 'hard_negative',
'external_id': '7459116'
}, {
'title': 'Cormac McCarthy',
'text':
'chores of the house, Lee was asked by Cormac to also get a day job so he could focus on his novel writing. Dismayed with the situation, she moved to Wyoming, where she filed for divorce and landed her first job teaching. Cormac McCarthy is fluent in Spanish and lived in Ibiza, Spain, in the 1960s and later settled in El Paso, Texas, where he lived for nearly 20 years. In an interview with Richard B. Woodward from "The New York Times", "McCarthy doesn\'t drink anymore – he quit 16 years ago in El Paso, with one of his young',
'label': 'negative',
'passage_id': '2145653'
}]
}
query_tok = "facebook/dpr-question_encoder-single-nq-base"
query_tokenizer = Tokenizer.load(query_tok, use_fast=True)
passage_tok = "facebook/dpr-ctx_encoder-single-nq-base"
passage_tokenizer = Tokenizer.load(passage_tok, use_fast=True)
processor = TextSimilarityProcessor(
query_tokenizer=query_tokenizer,
passage_tokenizer=passage_tokenizer,
max_seq_len_query=256,
max_seq_len_passage=256,
data_dir="data/retriever",
train_filename="nq-train.json",
test_filename="nq-dev.json",
embed_title=True,
num_hard_negatives=1,
label_list=["hard_negative", "positive"],
metric="text_similarity_metric",
shuffle_negatives=False)
processor.save(save_dir="testsave/dpr_processor")
dataset, tensor_names, _ = processor.dataset_from_dicts(
dicts=[d], return_baskets=False)
loadedprocessor = TextSimilarityProcessor.load_from_dir(
load_dir="testsave/dpr_processor")
dataset2, tensor_names, _ = loadedprocessor.dataset_from_dicts(
dicts=[d], return_baskets=False)
assert np.array_equal(dataset.tensors[0], dataset2.tensors[0])
class DensePassageRetriever(BaseRetriever):
"""
Retriever that uses a bi-encoder (one transformer for query, one transformer for passage).
See the original paper for more details:
Karpukhin, Vladimir, et al. (2020): "Dense Passage Retrieval for Open-Domain Question Answering."
(https://arxiv.org/abs/2004.04906).
"""
def __init__(self,
document_store: BaseDocumentStore,
query_embedding_model: Union[
Path,
str] = "facebook/dpr-question_encoder-single-nq-base",
passage_embedding_model: Union[
Path, str] = "facebook/dpr-ctx_encoder-single-nq-base",
max_seq_len_query: int = 64,
max_seq_len_passage: int = 256,
use_gpu: bool = True,
batch_size: int = 16,
embed_title: bool = True,
use_fast_tokenizers: bool = True,
similarity_function: str = "dot_product"):
"""
Init the Retriever incl. the two encoder models from a local or remote model checkpoint.
The checkpoint format matches huggingface transformers' model format
**Example:**
```python
| # remote model from FAIR
| DensePassageRetriever(document_store=your_doc_store,
| query_embedding_model="facebook/dpr-question_encoder-single-nq-base",
| passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base")
| # or from local path
| DensePassageRetriever(document_store=your_doc_store,
| query_embedding_model="model_directory/question-encoder",
| passage_embedding_model="model_directory/context-encoder")
```
:param document_store: An instance of DocumentStore from which to retrieve documents.
:param query_embedding_model: Local path or remote name of question encoder checkpoint. The format equals the
one used by hugging-face transformers' modelhub models
Currently available remote names: ``"facebook/dpr-question_encoder-single-nq-base"``
:param passage_embedding_model: Local path or remote name of passage encoder checkpoint. The format equals the
one used by hugging-face transformers' modelhub models
Currently available remote names: ``"facebook/dpr-ctx_encoder-single-nq-base"``
:param max_seq_len_query: Longest length of each query sequence. Maximum number of tokens for the query text. Longer ones will be cut down."
:param max_seq_len_passage: Longest length of each passage/context sequence. Maximum number of tokens for the passage text. Longer ones will be cut down."
:param use_gpu: Whether to use gpu or not
:param batch_size: Number of questions or passages to encode at once
:param embed_title: Whether to concatenate title and passage to a text pair that is then used to create the embedding.
This is the approach used in the original paper and is likely to improve performance if your
titles contain meaningful information for retrieval (topic, entities etc.) .
The title is expected to be present in doc.meta["name"] and can be supplied in the documents
before writing them to the DocumentStore like this:
{"text": "my text", "meta": {"name": "my title"}}.
"""
self.document_store = document_store
self.batch_size = batch_size
self.max_seq_len_passage = max_seq_len_passage
self.max_seq_len_query = max_seq_len_query
if use_gpu and torch.cuda.is_available():
self.device = torch.device("cuda")
else:
self.device = torch.device("cpu")
self.embed_title = embed_title
# Init & Load Encoders
self.query_tokenizer = Tokenizer.load(
pretrained_model_name_or_path=query_embedding_model,
do_lower_case=True,
use_fast=use_fast_tokenizers)
self.query_encoder = LanguageModel.load(
pretrained_model_name_or_path=query_embedding_model,
language_model_class="DPRQuestionEncoder")
self.passage_tokenizer = Tokenizer.load(
pretrained_model_name_or_path=passage_embedding_model,
do_lower_case=True,
use_fast=use_fast_tokenizers)
self.passage_encoder = LanguageModel.load(
pretrained_model_name_or_path=passage_embedding_model,
language_model_class="DPRContextEncoder")
self.processor = TextSimilarityProcessor(
tokenizer=self.query_tokenizer,
passage_tokenizer=self.passage_tokenizer,
max_seq_len_passage=self.max_seq_len_passage,
max_seq_len_query=self.max_seq_len_query,
label_list=["hard_negative", "positive"],
metric="text_similarity_metric",
embed_title=self.embed_title,
num_hard_negatives=0,
num_negatives=0)
prediction_head = TextSimilarityHead(
similarity_function=similarity_function)
self.model = BiAdaptiveModel(
language_model1=self.query_encoder,
language_model2=self.passage_encoder,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm1_output_types=["per_sequence"],
lm2_output_types=["per_sequence"],
device=self.device,
)
self.model.connect_heads_with_processor(self.processor.tasks,
require_labels=False)
def retrieve(self,
query: str,
filters: dict = None,
top_k: int = 10,
index: str = None) -> List[Document]:
if index is None:
index = self.document_store.index
query_emb = self.embed_queries(texts=[query])
documents = self.document_store.query_by_embedding(
query_emb=query_emb[0], top_k=top_k, filters=filters, index=index)
return documents
def _get_predictions(self, dicts, tokenizer):
"""
Feed a preprocessed dataset to the model and get the actual predictions (forward pass + formatting).
:param dicts: list of dictionaries
examples:[{'query': "where is florida?"}, {'query': "who wrote lord of the rings?"}, ...]
[{'passages': [{
"title": 'Big Little Lies (TV series)',
"text": 'series garnered several accolades. It received..',
"label": 'positive',
"external_id": '18768923'},
{"title": 'Framlingham Castle',
"text": 'Castle on the Hill "Castle on the Hill" is a song by English..',
"label": 'positive',
"external_id": '19930582'}, ...]
:return: dictionary of embeddings for "passages" and "query"
"""
dataset, tensor_names, baskets = self.processor.dataset_from_dicts(
dicts, indices=[i for i in range(len(dicts))], return_baskets=True)
data_loader = NamedDataLoader(dataset=dataset,
sampler=SequentialSampler(dataset),
batch_size=self.batch_size,
tensor_names=tensor_names)
all_embeddings = {"query": [], "passages": []}
self.model.eval()
for i, batch in enumerate(
tqdm(data_loader,
desc=f"Inferencing Samples",
unit=" Batches",
disable=False)):
batch = {key: batch[key].to(self.device) for key in batch}
# get logits
with torch.no_grad():
query_embeddings, passage_embeddings = self.model.forward(
**batch)[0]
if query_embeddings is not None:
all_embeddings["query"].append(
query_embeddings.cpu().numpy())
if passage_embeddings is not None:
all_embeddings["passages"].append(
passage_embeddings.cpu().numpy())
if all_embeddings["passages"]:
all_embeddings["passages"] = np.concatenate(
all_embeddings["passages"])
if all_embeddings["query"]:
all_embeddings["query"] = np.concatenate(all_embeddings["query"])
return all_embeddings
def embed_queries(self, texts: List[str]) -> List[np.array]:
"""
Create embeddings for a list of queries using the query encoder
:param texts: Queries to embed
:return: Embeddings, one per input queries
"""
queries = [{'query': q} for q in texts]
result = self._get_predictions(queries, self.query_tokenizer)["query"]
return result
def embed_passages(self, docs: List[Document]) -> List[np.array]:
"""
Create embeddings for a list of passages using the passage encoder
:param docs: List of Document objects used to represent documents / passages in a standardized way within Haystack.
:return: Embeddings of documents / passages shape (batch_size, embedding_dim)
"""
passages = [{
'passages': [{
"title":
d.meta["name"] if d.meta and "name" in d.meta else "",
"text":
d.text,
"label":
d.meta["label"]
if d.meta and "label" in d.meta else "positive",
"external_id":
d.id
}]
} for d in docs]
embeddings = self._get_predictions(passages,
self.passage_tokenizer)["passages"]
return embeddings
def train(self,
data_dir: str,
train_filename: str,
dev_filename: str = None,
test_filename: str = None,
batch_size: int = 2,
embed_title: bool = True,
num_hard_negatives: int = 1,
num_negatives: int = 0,
n_epochs: int = 3,
evaluate_every: int = 1000,
n_gpu: int = 1,
learning_rate: float = 1e-5,
epsilon: float = 1e-08,
weight_decay: float = 0.0,
num_warmup_steps: int = 100,
grad_acc_steps: int = 1,
optimizer_name: str = "TransformersAdamW",
optimizer_correct_bias: bool = True,
save_dir: str = "../saved_models/dpr-tutorial",
query_encoder_save_dir: str = "lm1",
passage_encoder_save_dir: str = "lm2"):
"""
train a DensePassageRetrieval model
:param data_dir: Directory where training file, dev file and test file are present
:param train_filename: training filename
:param dev_filename: development set filename, file to be used by model in eval step of training
:param test_filename: test set filename, file to be used by model in test step after training
:param batch_size: total number of samples in 1 batch of data
:param embed_title: whether to concatenate passage title with each passage. The default setting in official DPR embeds passage title with the corresponding passage
:param num_hard_negatives: number of hard negative passages(passages which are very similar(high score by BM25) to query but do not contain the answer
:param num_negatives: number of negative passages(any random passage from dataset which do not contain answer to query)
:param n_epochs: number of epochs to train the model on
:param evaluate_every: number of training steps after evaluation is run
:param n_gpu: number of gpus to train on
:param learning_rate: learning rate of optimizer
:param epsilon: epsilon parameter of optimizer
:param weight_decay: weight decay parameter of optimizer
:param grad_acc_steps: number of steps to accumulate gradient over before back-propagation is done
:param optimizer_name: what optimizer to use (default: TransformersAdamW)
:param num_warmup_steps: number of warmup steps
:param optimizer_correct_bias: Whether to correct bias in optimizer
:param save_dir: directory where models are saved
:param query_encoder_save_dir: directory inside save_dir where query_encoder model files are saved
:param passage_encoder_save_dir: directory inside save_dir where passage_encoder model files are saved
"""
self.embed_title = embed_title
self.processor = TextSimilarityProcessor(
tokenizer=self.query_tokenizer,
passage_tokenizer=self.passage_tokenizer,
max_seq_len_passage=self.max_seq_len_passage,
max_seq_len_query=self.max_seq_len_query,
label_list=["hard_negative", "positive"],
metric="text_similarity_metric",
data_dir=data_dir,
train_filename=train_filename,
dev_filename=dev_filename,
test_filename=test_filename,
embed_title=self.embed_title,
num_hard_negatives=num_hard_negatives,
num_negatives=num_negatives)
self.model.connect_heads_with_processor(self.processor.tasks,
require_labels=True)
data_silo = DataSilo(processor=self.processor,
batch_size=batch_size,
distributed=False)
# 5. Create an optimizer
self.model, optimizer, lr_schedule = initialize_optimizer(
model=self.model,
learning_rate=learning_rate,
optimizer_opts={
"name": optimizer_name,
"correct_bias": optimizer_correct_bias,
"weight_decay": weight_decay,
"eps": epsilon
},
schedule_opts={
"name": "LinearWarmup",
"num_warmup_steps": num_warmup_steps
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
grad_acc_steps=grad_acc_steps,
device=self.device)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer(
model=self.model,
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
device=self.device,
)
# 7. Let it grow! Watch the tracked metrics live on the public mlflow server: https://public-mlflow.deepset.ai
trainer.train()
self.model.save(Path(save_dir),
lm1_name=query_encoder_save_dir,
lm2_name=passage_encoder_save_dir)
self.processor.save(Path(save_dir))
def save(self, save_dir: Union[Path, str]):
save_dir = Path(save_dir)
self.model.save(save_dir,
lm1_name="query_encoder",
lm2_name="passage_encoder")
save_dir = str(save_dir)
self.query_tokenizer.save_pretrained(save_dir + "/query_encoder")
self.passage_tokenizer.save_pretrained(save_dir + "/passage_encoder")
@classmethod
def load(
cls,
load_dir: Union[Path, str],
document_store: BaseDocumentStore,
max_seq_len_query: int = 64,
max_seq_len_passage: int = 256,
use_gpu: bool = True,
batch_size: int = 16,
embed_title: bool = True,
use_fast_tokenizers: bool = True,
similarity_function: str = "dot_product",
):
load_dir = Path(load_dir)
dpr = cls(document_store=document_store,
query_embedding_model=Path(load_dir) / "query_encoder",
passage_embedding_model=Path(load_dir) / "passage_encoder",
max_seq_len_query=max_seq_len_query,
max_seq_len_passage=max_seq_len_passage,
use_gpu=use_gpu,
batch_size=batch_size,
embed_title=embed_title,
use_fast_tokenizers=use_fast_tokenizers,
similarity_function=similarity_function)
return dpr
def dense_passage_retrieval():
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
ml_logger = MLFlowLogger(tracking_uri="https://public-mlflow.deepset.ai/")
ml_logger.init_experiment(experiment_name="FARM-dense_passage_retrieval",
run_name="Run_dpr")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
batch_size = 4
n_epochs = 3
distributed = False # enable for multi GPU training via DDP
evaluate_every = 1000
question_lang_model = "facebook/dpr-question_encoder-single-nq-base"
passage_lang_model = "facebook/dpr-ctx_encoder-single-nq-base"
do_lower_case = True
use_fast = True
embed_title = True
num_hard_negatives = 1
similarity_function = "dot_product"
train_filename = "nq-train.json"
dev_filename = "nq-dev.json"
test_filename = "nq-dev.json"
max_samples = None # load a smaller dataset (e.g. for debugging)
# For multi GPU Training via DDP we need to get the local rank
args = parse_arguments()
device, n_gpu = initialize_device_settings(use_cuda=True,
local_rank=args.local_rank)
# 1.Create question and passage tokenizers
query_tokenizer = Tokenizer.load(
pretrained_model_name_or_path=question_lang_model,
do_lower_case=do_lower_case,
use_fast=use_fast)
passage_tokenizer = Tokenizer.load(
pretrained_model_name_or_path=passage_lang_model,
do_lower_case=do_lower_case,
use_fast=use_fast)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# data_dir "data/retriever" should contain DPR training and dev files downloaded from https://github.com/facebookresearch/DPR
# i.e., nq-train.json, nq-dev.json or trivia-train.json, trivia-dev.json
label_list = ["hard_negative", "positive"]
metric = "text_similarity_metric"
processor = TextSimilarityProcessor(query_tokenizer=query_tokenizer,
passage_tokenizer=passage_tokenizer,
max_seq_len_query=64,
max_seq_len_passage=256,
label_list=label_list,
metric=metric,
data_dir="../data/retriever",
train_filename=train_filename,
dev_filename=dev_filename,
test_filename=test_filename,
embed_title=embed_title,
num_hard_negatives=num_hard_negatives,
max_samples=max_samples)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
# NOTE: In FARM, the dev set metrics differ from test set metrics in that they are calculated on a token level instead of a word level
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
distributed=distributed)
# 4. Create an BiAdaptiveModel+
# a) which consists of 2 pretrained language models as a basis
question_language_model = LanguageModel.load(
pretrained_model_name_or_path="bert-base-uncased",
language_model_class="DPRQuestionEncoder")
passage_language_model = LanguageModel.load(
pretrained_model_name_or_path="bert-base-uncased",
language_model_class="DPRContextEncoder")
# b) and a prediction head on top that is suited for our task => Question Answering
prediction_head = TextSimilarityHead(
similarity_function=similarity_function)
model = BiAdaptiveModel(
language_model1=question_language_model,
language_model2=passage_language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm1_output_types=["per_sequence"],
lm2_output_types=["per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=1e-5,
optimizer_opts={"name": "TransformersAdamW", "correct_bias": True, "weight_decay": 0.0, \
"eps": 1e-08},
schedule_opts={"name": "LinearWarmup", "num_warmup_steps": 100},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
grad_acc_steps=1,
device=device,
distributed=distributed
)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer(
model=model,
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
device=device,
)
# 7. Let it grow! Watch the tracked metrics live on the public mlflow server: https://public-mlflow.deepset.ai
trainer.train()
# 8. Hooray! You have a model. Store it:
save_dir = Path("../saved_models/dpr-tutorial")
model.save(save_dir)
processor.save(save_dir)
# 9. Evaluate
test_data_loader = data_silo.get_data_loader("test")
if test_data_loader is not None:
evaluator_test = Evaluator(data_loader=test_data_loader,
tasks=data_silo.processor.tasks,
device=device)
model.connect_heads_with_processor(processor.tasks)
test_result = evaluator_test.eval(model)