def question_answering():
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="Public_FARM", run_name="Run_natural_questions")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
batch_size = 24
n_epochs = 1
evaluate_every = 500
lang_model = "deepset/roberta-base-squad2" # start with a model that can already extract answers
do_lower_case = False # roberta is a cased model
train_filename = "train_medium.jsonl"
dev_filename = "dev_medium.jsonl"
keep_is_impossible = 0.15 # downsample negative examples after data conversion
downsample_context_size = 300 # reduce length of wikipedia articles to relevant part around the answer
# 1.Create a tokenizer
tokenizer = Tokenizer.load(
pretrained_model_name_or_path=lang_model, do_lower_case=do_lower_case
)
# Add HTML tag tokens to the tokenizer vocabulary, so they do not get split apart
html_tags = [
"<Th>","</Th>",
"<Td>","</Td>",
"<Tr>","</Tr>",
"<Li>","</Li>",
"<P>" ,"</P>",
"<Ul>","</Ul>",
"<H1>","</H1>",
"<H2>","</H2>",
"<H3>","</H3>",
"<H4>","</H4>",
"<H5>", "</H5>",
"<Td_colspan=",
]
tokenizer.add_tokens(html_tags)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = NaturalQuestionsProcessor(
tokenizer=tokenizer,
max_seq_len=384,
train_filename=train_filename,
dev_filename=dev_filename,
keep_no_answer=keep_is_impossible,
downsample_context_size=downsample_context_size,
data_dir=Path("../data/natural_questions"),
)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor, batch_size=batch_size, caching=True)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model,n_added_tokens=len(html_tags))
# b) and in case of Natural Questions we need two Prediction Heads
# one for extractive Question Answering
qa_head = QuestionAnsweringHead()
# another one for answering yes/no questions or deciding if the given text passage might contain an answer
classification_head = TextClassificationHead(num_labels=len(processor.answer_type_list)) # answer_type_list = ["is_impossible", "span", "yes", "no"]
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[qa_head, classification_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
schedule_opts={"name": "LinearWarmup", "warmup_proportion": 0.2},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device
)
# 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/roberta-base-squad2-nq")
model.save(save_dir)
processor.save(save_dir)
# 9. Since training on the whole NQ corpus requires substantial compute resources we trained and uploaded a model on s3
fetch_archive_from_http("https://s3.eu-central-1.amazonaws.com/deepset.ai-farm-qa/models/roberta-base-squad2-nq.zip", output_dir="../saved_models/farm")
QA_input = [
{
"qas": ["Did GameTrailers rated Twilight Princess as one of the best games ever created?"],
"context": "Twilight Princess was released to universal critical acclaim and commercial success. It received perfect scores from major publications such as 1UP.com, Computer and Video Games, Electronic Gaming Monthly, Game Informer, GamesRadar, and GameSpy. On the review aggregators GameRankings and Metacritic, Twilight Princess has average scores of 95% and 95 for the Wii version and scores of 95% and 96 for the GameCube version. GameTrailers in their review called it one of the greatest games ever created."
}
]
model = QAInferencer.load(model_name_or_path="../saved_models/farm/roberta-base-squad2-nq", batch_size=batch_size, gpu=True)
result = model.inference_from_dicts(dicts=QA_input, return_json=False) # result is a list of QAPred objects
print(f"\nQuestion: Did GameTrailers rated Twilight Princess as one of the best games ever created?"
f"\nAnswer from model: {result[0].prediction[0].answer}")
model.close_multiprcessing_pool()
def xlmr_qa_demo():
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="Public_FARM", run_name="run_xmlr_qa")
#########################
######## Settings
########################
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
batch_size = 3
grad_acc_steps = 8
n_epochs = 2
evaluate_every = 200
base_LM_model = "xlm-roberta-large"
data_dir = Path("../data/squad20")
train_filename = Path("train-v2.0.json")
dev_filename = Path("dev-v2.0.json")
save_dir = Path("../saved_models/xlmr-large-qa")
inference_file = Path("../data/MLQA_V1/dev/dev-context-de-question-de.json")
predictions_file = save_dir / "predictions.json"
full_predictions_file = save_dir / "full_predictions.json"
max_processes_for_inference = 8
train = True
inference = False
if train:
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=base_LM_model)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
label_list = ["start_token", "end_token"]
metric = "squad"
processor = SquadProcessor(
tokenizer=tokenizer,
max_seq_len=384,
label_list=label_list,
metric=metric,
train_filename=train_filename,
dev_filename=dev_filename,
test_filename=None,
data_dir=data_dir,
dev_split=0.0
)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor, batch_size=batch_size, distributed=False, max_processes=1)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(base_LM_model, n_added_tokens=3)
# b) and a prediction head on top that is suited for our task => Question Answering
prediction_head = QuestionAnsweringHead()
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
schedule_opts={"name": "LinearWarmup", "warmup_proportion": 0.2},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
grad_acc_steps=grad_acc_steps,
device=device
)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer(
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
model = trainer.train(model)
# 8. Hooray! You have a model. Store it:
model.save(save_dir)
processor.save(save_dir)
if inference:
model = Inferencer.load(save_dir, batch_size=32, gpu=True)
full_result = model.inference_from_file(
file=inference_file,
max_processes=max_processes_for_inference,
)
for x in full_result:
print(x)
print()
result = {r["id"]: r["preds"][0][0] for r in full_result}
full_result = {r["id"]: r["preds"] for r in full_result}
json.dump(result,
open(predictions_file, "w"),
indent=4,
ensure_ascii=False)
json.dump(full_result,
open(full_predictions_file, "w"),
indent=4,
ensure_ascii=False)
def run_experiment(args):
logger.info("\n***********************************************"
f"\n************* Experiment: {args.task.name} ************"
"\n************************************************")
ml_logger = MlLogger(tracking_uri=args.logging.mlflow_url)
ml_logger.init_experiment(
experiment_name=args.logging.mlflow_experiment,
run_name=args.logging.mlflow_run_name,
nested=args.logging.mlflow_nested,
)
validate_args(args)
distributed = bool(args.general.local_rank != -1)
# Init device and distributed settings
device, n_gpu = initialize_device_settings(
use_cuda=args.general.cuda,
local_rank=args.general.local_rank,
fp16=args.general.fp16,
)
args.parameter.batch_size = int(args.parameter.batch_size //
args.parameter.gradient_accumulation_steps)
if n_gpu > 1:
args.parameter.batch_size = args.parameter.batch_size * n_gpu
set_all_seeds(args.general.seed)
# Prepare Data
tokenizer = Tokenizer.load(args.parameter.model,
do_lower_case=args.parameter.lower_case)
processor = Processor.load(
tokenizer=tokenizer,
max_seq_len=args.parameter.max_seq_len,
data_dir=args.general.data_dir,
**args.task.toDict(
), # args is of type DotMap and needs conversion to std python dicts
)
data_silo = DataSilo(
processor=processor,
batch_size=args.parameter.batch_size,
distributed=distributed,
)
class_weights = None
if args.parameter.balance_classes:
task_names = list(processor.tasks.keys())
if len(task_names) > 1:
raise NotImplementedError(
f"Balancing classes is currently not supported for multitask experiments. Got tasks: {task_names} "
)
class_weights = data_silo.calculate_class_weights(
task_name=task_names[0])
model = get_adaptive_model(
lm_output_type=args.parameter.lm_output_type,
prediction_heads=args.parameter.prediction_head,
layer_dims=args.parameter.layer_dims,
model=args.parameter.model,
device=device,
class_weights=class_weights,
embeds_dropout_prob=args.parameter.embeds_dropout_prob,
)
# Init optimizer
# TODO: warmup linear is sometimes NONE depending on fp16 - is there a neater way to handle this?
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=args.parameter.learning_rate,
warmup_proportion=args.parameter.warmup_proportion,
loss_scale=args.general.loss_scale,
fp16=args.general.fp16,
n_batches=len(data_silo.loaders["train"]),
grad_acc_steps=args.parameter.gradient_accumulation_steps,
n_epochs=args.parameter.epochs,
)
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=args.parameter.epochs,
n_gpu=n_gpu,
grad_acc_steps=args.parameter.gradient_accumulation_steps,
fp16=args.general.fp16,
local_rank=args.general.local_rank,
warmup_linear=warmup_linear,
evaluate_every=args.logging.eval_every,
device=device,
)
model = trainer.train(model)
model_name = (
f"{model.language_model.name}-{model.language_model.language}-{args.task.name}"
)
processor.save(f"{args.general.output_dir}/{model_name}")
model.save(f"{args.general.output_dir}/{model_name}")
def test_ner(caplog):
if caplog:
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 3
batch_size = 2
evaluate_every = 1
lang_model = "distilbert-base-german-cased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
ner_labels = [
"[PAD]", "X", "O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG",
"I-ORG", "B-LOC", "I-LOC", "B-OTH", "I-OTH"
]
processor = NERProcessor(tokenizer=tokenizer,
max_seq_len=8,
data_dir=Path("samples/ner"),
train_filename="train-sample.txt",
dev_filename="dev-sample.txt",
test_filename=None,
delimiter=" ",
label_list=ner_labels,
metric="seq_f1")
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
language_model = LanguageModel.load(lang_model)
prediction_head = TokenClassificationHead(num_labels=13)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
#optimizer_opts={'name': 'AdamW', 'lr': 2E-05},
n_batches=len(data_silo.loaders["train"]),
n_epochs=1,
device=device,
schedule_opts={
'name': 'LinearWarmup',
'warmup_proportion': 0.1
})
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,
)
save_dir = Path("testsave/ner")
model = trainer.train()
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{
"text": "Paris is a town in France."
},
]
model = Inferencer.load(
model_name_or_path="dbmdz/bert-base-cased-finetuned-conll03-english",
num_processes=0,
task_type="ner")
# labels arent correctly inserted from transformers
# They are converted to LABEL_1 ... LABEL_N
# For the inference result to contain predictions we need them in IOB NER format
model.processor.tasks["ner"]["label_list"][-1] = "B-LOC"
result = model.inference_from_dicts(dicts=basic_texts)
assert result[0]["predictions"][0]["context"] == "Paris"
assert isinstance(result[0]["predictions"][0]["probability"], np.float32)
def test_qa(caplog=None):
if caplog:
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
batch_size = 2
n_epochs = 1
evaluate_every = 4
base_LM_model = "distilbert-base-uncased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=base_LM_model,
do_lower_case=True)
label_list = ["start_token", "end_token"]
processor = SquadProcessor(tokenizer=tokenizer,
max_seq_len=20,
doc_stride=10,
max_query_length=6,
train_filename="train-sample.json",
dev_filename="dev-sample.json",
test_filename=None,
data_dir=Path("samples/qa"),
label_list=label_list,
metric="squad")
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
language_model = LanguageModel.load(base_LM_model)
prediction_head = QuestionAnsweringHead()
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
#optimizer_opts={'name': 'AdamW', 'lr': 2E-05},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device)
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)
trainer.train()
save_dir = Path("testsave/qa")
model.save(save_dir)
processor.save(save_dir)
inferencer = Inferencer.load(save_dir,
batch_size=2,
gpu=False,
num_processes=0)
qa_format_1 = [{
"questions": ["Who counted the game among the best ever made?"],
"text":
"Twilight Princess was released to universal critical acclaim and commercial success. It received perfect scores from major publications such as 1UP.com, Computer and Video Games, Electronic Gaming Monthly, Game Informer, GamesRadar, and GameSpy. On the review aggregators GameRankings and Metacritic, Twilight Princess has average scores of 95% and 95 for the Wii version and scores of 95% and 96 for the GameCube version. GameTrailers in their review called it one of the greatest games ever created."
}]
qa_format_2 = [{
"qas": ["Who counted the game among the best ever made?"],
"context":
"Twilight Princess was released to universal critical acclaim and commercial success. It received perfect scores from major publications such as 1UP.com, Computer and Video Games, Electronic Gaming Monthly, Game Informer, GamesRadar, and GameSpy. On the review aggregators GameRankings and Metacritic, Twilight Princess has average scores of 95% and 95 for the Wii version and scores of 95% and 96 for the GameCube version. GameTrailers in their review called it one of the greatest games ever created.",
}]
result1 = inferencer.inference_from_dicts(dicts=qa_format_1)
result2 = inferencer.inference_from_dicts(dicts=qa_format_2)
assert result1 == result2
def test_ner(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 5
batch_size = 2
evaluate_every = 1
lang_model = "distilbert-base-german-cased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
ner_labels = [
"[PAD]", "X", "O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG",
"I-ORG", "B-LOC", "I-LOC", "B-OTH", "I-OTH"
]
processor = NERProcessor(tokenizer=tokenizer,
max_seq_len=8,
data_dir=Path("samples/ner"),
train_filename="train-sample.txt",
dev_filename="dev-sample.txt",
test_filename=None,
delimiter=" ",
label_list=ner_labels,
metric="seq_f1")
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
language_model = LanguageModel.load(lang_model)
prediction_head = TokenClassificationHead(num_labels=13)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
#optimizer_opts={'name': 'AdamW', 'lr': 2E-05},
n_batches=len(data_silo.loaders["train"]),
n_epochs=1,
device=device,
schedule_opts={
'name': 'LinearWarmup',
'warmup_proportion': 0.1
})
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,
)
save_dir = Path("testsave/ner")
model = trainer.train()
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{
"text": "Albrecht Lehman ist eine Person"
},
]
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts, max_processes=1)
#print(result)
#assert result[0]["predictions"][0]["context"] == "sagte"
#assert isinstance(result[0]["predictions"][0]["probability"], np.float32)
result2 = model.inference_from_dicts(dicts=basic_texts,
rest_api_schema=True)
assert result == result2
def test_doc_regression(data_dir_path, text_column_name, caplog=None):
if caplog:
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 1
evaluate_every = 2
lang_model = "bert-base-cased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
rp_params = dict(tokenizer=tokenizer,
max_seq_len=8,
data_dir=Path(data_dir_path),
train_filename="train-sample.tsv",
dev_filename="test-sample.tsv",
test_filename=None,
label_column_name="label")
if text_column_name is not None:
rp_params["text_column_name"] = text_column_name
processor = RegressionProcessor(**rp_params)
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = LanguageModel.load(lang_model)
prediction_head = RegressionHead()
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence_continuous"],
device=device)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
#optimizer_opts={'name': 'AdamW', 'lr': 2E-05},
n_batches=len(data_silo.loaders["train"]),
n_epochs=1,
device=device,
schedule_opts={
'name': 'CosineWarmup',
'warmup_proportion': 0.1
})
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)
trainer.train()
save_dir = Path("testsave/doc_regr")
model.save(save_dir)
processor.save(save_dir)
del model
del processor
del optimizer
del data_silo
del trainer
basic_texts = [
{
"text":
"The dress is just fabulous and it totally fits my size. The fabric is of great quality and the seams are really well hidden. I am super happy with this purchase and I am looking forward to trying some more from the same brand."
},
{
"text":
"it just did not fit right. The top is very thin showing everything."
},
]
model = Inferencer.load(save_dir, num_processes=0)
result = model.inference_from_dicts(dicts=basic_texts)
assert isinstance(result[0]["predictions"][0]["pred"], np.float32)
del model
def doc_classifcation():
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="Public_FARM",
run_name="Run_doc_classification_glove")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
n_epochs = 3
batch_size = 32
evaluate_every = 100
# load from a local path:
lang_model = Path("../saved_models/glove-german-uncased")
# or through s3
#lang_model = "glove-german-uncased"
do_lower_case = True
device, n_gpu = initialize_device_settings(use_cuda=True)
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# Here we load GermEval 2018 Data automaticaly if it is not available.
# GermEval 2018 only has train.tsv and test.tsv dataset - no dev.tsv
label_list = ["OTHER", "OFFENSE"]
metric = "f1_macro"
processor = TextClassificationProcessor(
tokenizer=tokenizer,
max_seq_len=128,
data_dir=Path("../data/germeval18"),
label_list=label_list,
dev_split=0,
test_filename="test.tsv",
train_filename="train.tsv",
metric=metric,
label_column_name="coarse_label")
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
# 4. Create an AdaptiveModel
# a) which consists of an embedding model as a basis.
# Word embedding models only converts words it has seen during training to embedding vectors.
language_model = LanguageModel.load(lang_model)
# b) and a prediction head on top that is suited for our task => Text classification
prediction_head = TextClassificationHead(
layer_dims=[300, 600, len(label_list)],
class_weights=data_silo.calculate_class_weights(
task_name="text_classification"),
num_labels=len(label_list))
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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
trainer.train()
delimiter="\t")
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = LanguageModel.load(lang_model)
prediction_head = RegressionHead()
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence_continuous"],
device=device)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=multitransquest_config['learning_rate'],
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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)
trainer.train()
def test_qa(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
batch_size = 32
n_epochs = 1
evaluate_every = 2
base_LM_model = "bert-base-cased"
tokenizer = BertTokenizer.from_pretrained(
pretrained_model_name_or_path=base_LM_model, do_lower_case=False)
label_list = ["start_token", "end_token"]
processor = SquadProcessor(tokenizer=tokenizer,
max_seq_len=64,
train_filename="train-sample.json",
dev_filename="dev-sample.json",
test_filename=None,
data_dir="samples/qa",
labels=label_list,
metric="squad")
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = Bert.load(base_LM_model)
prediction_head = QuestionAnsweringHead(layer_dims=[768, len(label_list)])
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=1e-5,
warmup_proportion=0.2,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
)
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
evaluate_every=evaluate_every,
device=device,
)
model = trainer.train(model)
save_dir = "testsave/qa"
model.save(save_dir)
processor.save(save_dir)
QA_input = [{
"questions": ["In what country is Normandy located?"],
"text":
'The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France. They were descended from Norse ("Norman" comes from "Norseman") raiders and pirates from Denmark, Iceland and Norway who, under their leader Rollo, agreed to swear fealty to King Charles III of West Francia. Through generations of assimilation and mixing with the native Frankish and Roman-Gaulish populations, their descendants would gradually merge with the Carolingian-based cultures of West Francia. The distinct cultural and ethnic identity of the Normans emerged initially in the first half of the 10th century, and it continued to evolve over the succeeding centuries.',
}]
model = Inferencer.load(save_dir)
result = model.run_inference(dicts=QA_input)
assert result[0]["predictions"][0]["end"] == 65
# if(__name__=="__main__"):
# test_qa()
def doc_classification_cola():
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="Public_FARM",
run_name="Run_cola")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 5
batch_size = 100
evaluate_every = 20
lang_model = "bert-base-cased"
do_lower_case = False
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# Here we load Cola 2018 Data.
label_list = ["0", "1"]
metric = "mcc"
processor = TextClassificationProcessor(tokenizer=tokenizer,
max_seq_len=64,
data_dir=Path("../data/cola"),
dev_filename=Path("dev.tsv"),
dev_split=None,
test_filename=None,
label_list=label_list,
metric=metric,
label_column_name="label")
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor, batch_size=batch_size)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# language_model = Roberta.load(lang_model)
# b) and a prediction head on top that is suited for our task => Text classification
prediction_head = TextClassificationHead(
num_labels=len(label_list),
class_weights=data_silo.calculate_class_weights(
task_name="text_classification"))
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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
trainer.train()
# 8. Hooray! You have a model. Store it:
save_dir = Path("saved_models/bert-doc-tutorial")
model.save(save_dir)
processor.save(save_dir)
# 9. Load it & harvest your fruits (Inference)
basic_texts = [
{
"text": "The box contained the ball from the tree."
},
{
"text": "I'll fix you a drink."
},
]
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
print(result)
def doc_classification(task,
model_type,
n_epochs,
batch_size,
embeds_dropout,
evaluate_every,
use_cuda,
max_seq_len,
learning_rate,
do_lower_case,
register_model,
save_model=True,
early_stopping=False):
language = cu.params.get('language')
# Check task
if cu.tasks.get(str(task)).get('type') != 'classification':
raise Exception('NOT A CLASSIFICATION TASK')
# Data
dt_task = dt.Data(task=task)
## Download training files
if not os.path.isfile(dt_task.get_path('fn_train', dir='data_dir')):
dt_task.download('data_dir', dir='data_dir', source='datastore')
# Settings
set_all_seeds(seed=42)
use_amp = None
device, n_gpu = initialize_device_settings(use_cuda=use_cuda,
use_amp=use_amp)
lang_model = he.get_farm_model(model_type, language)
save_dir = dt_task.get_path('model_dir')
label_list = dt_task.load('fn_label', dir='data_dir',
header=None)[0].to_list()
# AML log
try:
aml_run.log('task', task)
aml_run.log('language', language)
aml_run.log('n_epochs', n_epochs)
aml_run.log('batch_size', batch_size)
aml_run.log('learning_rate', learning_rate)
aml_run.log('embeds_dropout', embeds_dropout)
aml_run.log('max_seq_len', max_seq_len)
aml_run.log('lang_model', lang_model)
aml_run.log_list('label_list', label_list)
except:
pass
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# The evaluation on the dev-set can be done with one of the predefined metrics or with a
# metric defined as a function from (preds, labels) to a dict that contains all the actual
# metrics values. The function must get registered under a string name and the string name must
# be used.
def mymetrics(preds, labels):
acc = simple_accuracy(preds, labels)
f1macro = f1_score(y_true=labels, y_pred=preds, average="macro")
f1micro = f1_score(y_true=labels, y_pred=preds, average="micro")
# AML log
try:
aml_run.log('acc', acc.get('acc'))
aml_run.log('f1macro', f1macro)
aml_run.log('f1micro', f1micro)
except:
pass
return {"acc": acc, "f1_macro": f1macro, "f1_micro": f1micro}
register_metrics('mymetrics', mymetrics)
metric = 'mymetrics'
processor = TextClassificationProcessor(
tokenizer=tokenizer,
max_seq_len=max_seq_len,
data_dir=dt_task.data_dir,
label_list=label_list,
metric=metric,
label_column_name="label",
train_filename=dt_task.get_path('fn_train', dir='data_dir'),
test_filename=dt_task.get_path('fn_test', dir='data_dir'))
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor, batch_size=batch_size)
# 4. Create an AdaptiveModel
## Pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
## Prediction head on top that is suited for our task => Text classification
prediction_head = TextClassificationHead(
num_labels=len(processor.tasks["text_classification"]["label_list"]),
class_weights=data_silo.calculate_class_weights(
task_name="text_classification"))
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=embeds_dropout,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device,
learning_rate=learning_rate,
use_amp=use_amp)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant and evaluates it from time to time
# Also create an EarlyStopping instance and pass it on to the trainer
# An early stopping instance can be used to save the model that performs best on the dev set
# according to some metric and stop training when no improvement is happening for some iterations.
if early_stopping:
earlystopping = EarlyStopping(
metric="f1_macro",
mode="max", # use f1_macro from the dev evaluator of the trainer
# metric="loss", mode="min", # use loss from the dev evaluator of the trainer
save_dir=save_dir, # where to save the best model
patience=
2 # number of evaluations to wait for improvement before terminating the training
)
else:
earlystopping = None
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,
early_stopping=earlystopping)
# 7. Let it grow
trainer.train()
# 8. Store it:
# NOTE: if early stopping is used, the best model has been stored already in the directory
# defined with the EarlyStopping instance
# The model we have at this moment is the model from the last training epoch that was carried
# out before early stopping terminated the training
if save_model:
model.save(save_dir)
processor.save(save_dir)
if register_model:
dt_task.upload('model_dir', destination='model')
def doc_classification_multilabel():
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="Public_FARM", run_name="Run_doc_classification")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 1
batch_size = 32
evaluate_every = 500
lang_model = "bert-base-uncased"
do_lower_case = True
# 1.Create a tokenizer
tokenizer = Tokenizer.load(
pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# Here we load GermEval 2018 Data.
label_list = ["toxic","severe_toxic","obscene","threat","insult","identity_hate"]
metric = "acc"
processor = TextClassificationProcessor(tokenizer=tokenizer,
max_seq_len=128,
data_dir=Path("../data/toxic-comments"),
label_list=label_list,
label_column_name="label",
metric=metric,
quote_char='"',
multilabel=True,
train_filename="train.tsv",
dev_filename="val.tsv",
test_filename=None,
dev_split=0,
)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(
processor=processor,
batch_size=batch_size)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# b) and a prediction head on top that is suited for our task => Text classification
prediction_head = MultiLabelTextClassificationHead(num_labels=len(label_list))
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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
trainer.train()
# 8. Hooray! You have a model. Store it:
save_dir = Path("../saved_models/bert-german-multi-doc-tutorial")
model.save(save_dir)
processor.save(save_dir)
# 9. Load it & harvest your fruits (Inference)
basic_texts = [
{"text": "You f*****g bastards"},
{"text": "What a lovely world"},
]
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
print(result)
def main(args):
print(f"[INFO] PyTorch Version: {torch.__version__}")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("[INFO] Devices available: {}".format(device))
checkpoint_path = Path(args.ckpt_path) / args.run_name
ml_logger = MLFlowLogger(tracking_uri=args.tracking_uri)
ml_logger.init_experiment(experiment_name=args.experiment_name,
run_name=args.run_name)
tokenizer = Tokenizer.load(
pretrained_model_name_or_path=args.pretrained_model_name_or_path,
do_lower_case=False)
# Processor
if args.task_name == "text_classification":
processor = TextClassificationProcessor(
tokenizer=tokenizer,
train_filename=args.train_filename,
dev_filename=None,
test_filename=args.test_filename,
header=0,
max_seq_len=args.max_seq_len,
data_dir=args.data_dir,
label_list=args.label_list,
metric=args.metric,
label_column_name=args.label_column_name,
text_column_name=args.text_column_name)
elif args.task_name == "question_answering":
processor = SquadProcessor(tokenizer=tokenizer,
train_filename=args.train_filename,
dev_filename=args.test_filename,
test_filename=args.test_filename,
max_seq_len=args.max_seq_len,
data_dir=args.data_dir,
label_list=args.label_list,
metric=args.metric,
max_query_length=64,
doc_stride=128,
max_answers=1)
else:
raise ValueError("task name error")
processor.save(checkpoint_path)
# DataSilo
data_silo = DataSilo(processor=processor,
batch_size=args.batch_size,
eval_batch_size=args.eval_batch_size,
caching=True,
cache_path=checkpoint_path)
# LanguageModel: Build pretrained language model
language_model = LanguageModel.load(args.pretrained_model_name_or_path,
language="korean")
# PredictionHead: Build predictor layer
if args.task_name == "text_classification":
# If you do classification on imbalanced classes, consider using class weights.
# They change the loss function to down-weight frequent classes.
prediction_head = TextClassificationHead(
num_labels=len(args.label_list),
class_weights=data_silo.calculate_class_weights(
task_name=args.task_name))
elif args.task_name == "question_answering":
prediction_head = QuestionAnsweringHead(
layer_dims=[768, 2],
task_name=args.task_name,
)
else:
raise ValueError("task name error")
# AdaptiveModel: Combine all
if args.task_name == "text_classification":
lm_output_types = ["per_sequence"]
elif args.task_name == "question_answering":
lm_output_types = ["per_token"]
else:
raise ValueError("task name error")
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=args.embeds_dropout_prob,
lm_output_types=lm_output_types,
device=device)
# Initialize Optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
device=device,
learning_rate=args.learning_rate,
n_batches=len(data_silo.loaders["train"]),
n_epochs=args.n_epochs)
# EarlyStopping
earlymetric = "f1" if args.task_name == "question_answering" else "acc"
mode = "max" if args.task_name in [
"text_classification", "question_answering"
] else "min"
earlystop = EarlyStopping(save_dir=checkpoint_path,
metric=earlymetric,
mode=mode,
patience=5)
# Trainer
trainer = Trainer(
model=model,
optimizer=optimizer,
lr_schedule=lr_schedule,
data_silo=data_silo,
early_stopping=earlystop,
evaluate_every=args.evaluate_every,
checkpoints_to_keep=args.checkpoints_to_keep,
checkpoint_root_dir=checkpoint_path,
checkpoint_every=args.checkpoint_every,
epochs=args.n_epochs,
n_gpu=args.n_gpu,
device=device,
)
# now train!
model = trainer.train()
def train(
self,
data_dir: str,
train_filename: str,
dev_filename: Optional[str] = None,
test_file_name: Optional[str] = None,
use_gpu: Optional[bool] = None,
batch_size: int = 10,
n_epochs: int = 2,
learning_rate: float = 1e-5,
max_seq_len: Optional[int] = None,
warmup_proportion: float = 0.2,
dev_split: Optional[float] = 0.1,
evaluate_every: int = 300,
save_dir: Optional[str] = None,
):
"""
Fine-tune a model on a QA dataset. Options:
- Take a plain language model (e.g. `bert-base-cased`) and train it for QA (e.g. on SQuAD data)
- Take a QA model (e.g. `deepset/bert-base-cased-squad2`) and fine-tune it for your domain (e.g. using your labels collected via the haystack annotation tool)
:param data_dir: Path to directory containing your training data in SQuAD style
:param train_filename: filename of training data
:param dev_filename: filename of dev / eval data
:param test_file_name: filename of test data
:param dev_split: Instead of specifying a dev_filename you can also specify a ratio (e.g. 0.1) here
that get's split off from training data for eval.
:param use_gpu: Whether to use GPU (if available)
:param batch_size: Number of samples the model receives in one batch for training
:param n_epochs: number of iterations on the whole training data set
:param learning_rate: learning rate of the optimizer
:param max_seq_len: maximum text length (in tokens). Everything longer gets cut down.
:param warmup_proportion: Proportion of training steps until maximum learning rate is reached.
Until that point LR is increasing linearly. After that it's decreasing again linearly.
Options for different schedules are available in FARM.
:param evaluate_every: Evaluate the model every X steps on the hold-out eval dataset
:param save_dir: Path to store the final model
:return: None
"""
if dev_filename:
dev_split = None
set_all_seeds(seed=42)
# For these variables, by default, we use the value set when initializing the FARMReader.
# These can also be manually set when train() is called if you want a different value at train vs inference
if use_gpu is None:
use_gpu = self.use_gpu
if max_seq_len is None:
max_seq_len = self.max_seq_len
device, n_gpu = initialize_device_settings(use_cuda=use_gpu)
if not save_dir:
save_dir = f"../../saved_models/{self.inferencer.model.language_model.name}"
# 1. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
label_list = ["start_token", "end_token"]
metric = "squad"
processor = SquadProcessor(
tokenizer=self.inferencer.processor.tokenizer,
max_seq_len=max_seq_len,
label_list=label_list,
metric=metric,
train_filename=train_filename,
dev_filename=dev_filename,
dev_split=dev_split,
test_filename=test_file_name,
data_dir=Path(data_dir),
)
# 2. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them
# and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
distributed=False)
# 3. Create an optimizer and pass the already initialized model
model, optimizer, lr_schedule = initialize_optimizer(
model=self.inferencer.model,
learning_rate=learning_rate,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": warmup_proportion
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device)
# 4. 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,
)
# 5. Let it grow!
self.inferencer.model = trainer.train()
self.save(Path(save_dir))
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_positives: int = 1,
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",
query_encoder_save_dir: str = "query_encoder",
passage_encoder_save_dir: str = "passage_encoder"):
"""
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_positives: number of positive passages
: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_positives=num_positives)
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.query_tokenizer.save_pretrained(
f"{save_dir}/{query_encoder_save_dir}")
self.passage_tokenizer.save_pretrained(
f"{save_dir}/{passage_encoder_save_dir}")
def doc_classifcation():
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="Public_FARM", run_name="Run_doc_classification")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
n_epochs = 1
batch_size = 32
evaluate_every = 100
lang_model = "bert-base-german-cased"
do_lower_case = False
dev_split = 0.1
dev_stratification = True
max_processes = 1 # 128 is default
# or a local path:
# lang_model = Path("../saved_models/farm-bert-base-cased")
use_amp = None
device, n_gpu = initialize_device_settings(use_cuda=True, use_amp=use_amp)
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model, do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# Here we load GermEval 2018 Data automaticaly if it is not available.
# GermEval 2018 only has train.tsv and test.tsv dataset - no dev.tsv
label_list = ["OTHER", "OFFENSE"]
metric = "f1_macro"
processor = TextClassificationProcessor(tokenizer=tokenizer,
max_seq_len=128,
data_dir=Path("../data/germeval18"),
label_list=label_list,
metric=metric,
dev_split=dev_split,
dev_stratification=dev_stratification,
label_column_name="coarse_label"
)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a
# few descriptive statistics of our datasets
data_silo = DataSilo(
processor=processor,
max_processes=max_processes,
batch_size=batch_size)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# b) and a prediction head on top that is suited for our task => Text classification
prediction_head = TextClassificationHead(
class_weights=data_silo.calculate_class_weights(task_name="text_classification"),
num_labels=len(label_list))
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
use_amp=use_amp)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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
trainer.train()
# 8. Hooray! You have a model. Store it:
save_dir = Path("saved_models/bert-german-doc-tutorial")
model.save(save_dir)
processor.save(save_dir)
# 9. Load it & harvest your fruits (Inference)
basic_texts = [
{"text": "Schartau sagte dem Tagesspiegel, dass Fischer ein Idiot sei"},
{"text": "Martin Müller spielt Handball in Berlin"},
]
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
print(result)
model.close_multiprocessing_pool()
def test_ner_amp(caplog):
if caplog:
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 1
batch_size = 2
evaluate_every = 1
lang_model = "bert-base-german-cased"
if AMP_AVAILABLE:
use_amp = 'O1'
else:
use_amp = None
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
ner_labels = [
"[PAD]", "X", "O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG",
"I-ORG", "B-LOC", "I-LOC", "B-OTH", "I-OTH"
]
processor = NERProcessor(tokenizer=tokenizer,
max_seq_len=8,
data_dir=Path("samples/ner"),
train_filename=Path("train-sample.txt"),
dev_filename=Path("dev-sample.txt"),
test_filename=None,
delimiter=" ",
label_list=ner_labels,
metric="seq_f1")
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
language_model = LanguageModel.load(lang_model)
prediction_head = TokenClassificationHead(num_labels=13)
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-05,
schedule_opts=None,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device,
use_amp=use_amp)
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,
)
save_dir = Path("testsave/ner")
trainer.train()
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{
"text": "1980 kam der Crown von Toyota"
},
]
model = Inferencer.load(save_dir, num_processes=0)
result = model.inference_from_dicts(dicts=basic_texts)
assert result[0]["predictions"][0]["context"] == "Crown"
assert isinstance(result[0]["predictions"][0]["probability"], np.float32)
def doc_classification_with_earlystopping():
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/")
# for local logging instead:
# ml_logger = MLFlowLogger(tracking_uri="logs")
ml_logger.init_experiment(experiment_name="Public_FARM",
run_name="DocClassification_ES_f1_1")
##########################
########## Settings
##########################
set_all_seeds(seed=42)
use_amp = None
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 20
batch_size = 32
evaluate_every = 100
lang_model = "bert-base-german-cased"
do_lower_case = False
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
# Here we load GermEval 2018 Data automaticaly if it is not available.
# GermEval 2018 only has train.tsv and test.tsv dataset - no dev.tsv
# The processor wants to know the possible labels ...
label_list = ["OTHER", "OFFENSE"]
# The evaluation on the dev-set can be done with one of the predefined metrics or with a
# metric defined as a function from (preds, labels) to a dict that contains all the actual
# metrics values. The function must get registered under a string name and the string name must
# be used.
def mymetrics(preds, labels):
acc = simple_accuracy(preds, labels)
f1other = f1_score(y_true=labels, y_pred=preds, pos_label="OTHER")
f1offense = f1_score(y_true=labels, y_pred=preds, pos_label="OFFENSE")
f1macro = f1_score(y_true=labels, y_pred=preds, average="macro")
f1micro = f1_score(y_true=labels, y_pred=preds, average="micro")
return {
"acc": acc,
"f1_other": f1other,
"f1_offense": f1offense,
"f1_macro": f1macro,
"f1_micro": f1micro
}
register_metrics('mymetrics', mymetrics)
metric = 'mymetrics'
processor = TextClassificationProcessor(
tokenizer=tokenizer,
max_seq_len=64,
data_dir=Path("../data/germeval18"),
label_list=label_list,
metric=metric,
label_column_name="coarse_label")
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor, batch_size=batch_size)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# b) and a prediction head on top that is suited for our task => Text classification
prediction_head = TextClassificationHead(
num_labels=len(label_list),
class_weights=data_silo.calculate_class_weights(
task_name="text_classification"))
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.2,
lm_output_types=["per_sequence"],
device=device)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=0.5e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
use_amp=use_amp)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant and evaluates it from time to time
# Also create an EarlyStopping instance and pass it on to the trainer
# An early stopping instance can be used to save the model that performs best on the dev set
# according to some metric and stop training when no improvement is happening for some iterations.
earlystopping = EarlyStopping(
metric="f1_offense",
mode=
"max", # use the metric from our own metrics function instead of loss
# metric="f1_macro", mode="max", # use f1_macro from the dev evaluator of the trainer
# metric="loss", mode="min", # use loss from the dev evaluator of the trainer
save_dir=Path("saved_models/bert-german-doc-tutorial-es"
), # where to save the best model
patience=
5 # number of evaluations to wait for improvement before terminating the training
)
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,
early_stopping=earlystopping)
# 7. Let it grow
trainer.train()
# 8. Hooray! You have a model.
# NOTE: if early stopping is used, the best model has been stored already in the directory
# defined with the EarlyStopping instance
# The model we have at this moment is the model from the last training epoch that was carried
# out before early stopping terminated the training
save_dir = Path("saved_models/bert-german-doc-tutorial")
model.save(save_dir)
processor.save(save_dir)
# 9. Load it & harvest your fruits (Inference)
basic_texts = [
{
"text":
"Schartau sagte dem Tagesspiegel, dass Fischer ein Idiot sei"
},
{
"text": "Martin Müller spielt Handball in Berlin"
},
]
# Load from the final epoch directory and apply
print("LOADING INFERENCER FROM FINAL MODEL DURING TRAINING")
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
print(result)
model.close_multiprocessing_pool()
# Load from saved best model
print("LOADING INFERENCER FROM BEST MODEL DURING TRAINING")
model = Inferencer.load(earlystopping.save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
print("APPLICATION ON BEST MODEL")
print(result)
model.close_multiprocessing_pool()
def train_from_scratch():
args = parse_arguments()
use_amp = "O2" # using "O2" here allows roughly 30% larger batch_sizes and 45% speed up
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Only the main process should log here
if args.local_rank in [-1, 0]:
ml_logger = MLFlowLogger(
tracking_uri="https://public-mlflow.deepset.ai/")
ml_logger.init_experiment(experiment_name="train_from_scratch",
run_name="run")
set_all_seeds(seed=39)
device, n_gpu = initialize_device_settings(use_cuda=True,
local_rank=args.local_rank,
use_amp=use_amp)
save_dir = Path("saved_models/train_from_scratch")
data_dir = Path("data/test")
# Option A) just using a single file
# train_filename = "train.txt"
# Option B) (recommended when using StreamingDataSilo):
# split and shuffle that file to have random order within and across epochs
randomize_and_split_file(data_dir / "train.txt",
output_dir=Path("data/split_files"),
docs_per_file=1000)
train_filename = Path("data/split_files")
dev_filename = "dev.txt"
distributed = args.local_rank != -1
max_seq_len = 128
batch_size = 8 #if distributed: this is per_gpu
grad_acc = 1
learning_rate = 1e-4
warmup_proportion = 0.05
n_epochs = 2
evaluate_every = 15000
log_loss_every = 2
checkpoint_every = 500
checkpoint_root_dir = Path("checkpoints")
checkpoints_to_keep = 4
next_sent_pred_style = "bert-style" #or "sentence"
max_docs = None
# Choose enough workers to queue sufficient batches during training.
# Optimal number depends on your GPU speed, CPU speed and number of cores
# 16 works well on a 4x V100 machine with 16 cores (AWS: p3.8xlarge). For a single GPU you will need less.
data_loader_workers = 1
# 1.Create a tokenizer
tokenizer = Tokenizer.load("bert-base-uncased", do_lower_case=True)
# 2. Create a DataProcessor that handles all the conversion from raw text into a PyTorch Dataset
processor = BertStyleLMProcessor(data_dir=data_dir,
tokenizer=tokenizer,
max_seq_len=max_seq_len,
train_filename=train_filename,
dev_filename=dev_filename,
test_filename=None,
next_sent_pred_style=next_sent_pred_style,
max_docs=max_docs)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and
# calculates a few descriptive statistics of our datasets
# stream_data_silo = DataSilo(processor=processor, batch_size=batch_size, distributed=distributed)
stream_data_silo = StreamingDataSilo(
processor=processor,
batch_size=batch_size,
distributed=distributed,
dataloader_workers=data_loader_workers)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.from_scratch("bert", tokenizer.vocab_size)
# b) and *two* prediction heads on top that are suited for our task => Language Model finetuning
lm_prediction_head = BertLMHead(768, tokenizer.vocab_size)
next_sentence_head = NextSentenceHead(num_labels=2,
task_name="nextsentence")
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=learning_rate,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": warmup_proportion
},
n_batches=len(stream_data_silo.get_data_loader("train")),
n_epochs=n_epochs,
device=device,
grad_acc_steps=grad_acc,
distributed=distributed,
use_amp=use_amp,
local_rank=args.local_rank)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer.create_or_load_checkpoint(
model=model,
optimizer=optimizer,
data_silo=stream_data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
log_loss_every=log_loss_every,
device=device,
grad_acc_steps=grad_acc,
local_rank=args.local_rank,
checkpoint_every=checkpoint_every,
checkpoint_root_dir=checkpoint_root_dir,
checkpoints_to_keep=checkpoints_to_keep,
use_amp=use_amp)
# 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:
model.save(save_dir)
processor.save(save_dir)
if args.local_rank != -1:
torch.distributed.destroy_process_group()
def train(
self,
data_dir: str,
train_filename: str,
dev_filename: Optional[str] = None,
test_filename: Optional[str] = None,
use_gpu: Optional[bool] = None,
batch_size: int = 10,
n_epochs: int = 2,
learning_rate: float = 1e-5,
max_seq_len: Optional[int] = None,
warmup_proportion: float = 0.2,
dev_split: float = 0,
evaluate_every: int = 300,
save_dir: Optional[str] = None,
num_processes: Optional[int] = None,
use_amp: str = None,
):
"""
Fine-tune a model on a QA dataset. Options:
- Take a plain language model (e.g. `bert-base-cased`) and train it for QA (e.g. on SQuAD data)
- Take a QA model (e.g. `deepset/bert-base-cased-squad2`) and fine-tune it for your domain (e.g. using your labels collected via the haystack annotation tool)
:param data_dir: Path to directory containing your training data in SQuAD style
:param train_filename: Filename of training data
:param dev_filename: Filename of dev / eval data
:param test_filename: Filename of test data
:param dev_split: Instead of specifying a dev_filename, you can also specify a ratio (e.g. 0.1) here
that gets split off from training data for eval.
:param use_gpu: Whether to use GPU (if available)
:param batch_size: Number of samples the model receives in one batch for training
:param n_epochs: Number of iterations on the whole training data set
:param learning_rate: Learning rate of the optimizer
:param max_seq_len: Maximum text length (in tokens). Everything longer gets cut down.
:param warmup_proportion: Proportion of training steps until maximum learning rate is reached.
Until that point LR is increasing linearly. After that it's decreasing again linearly.
Options for different schedules are available in FARM.
:param evaluate_every: Evaluate the model every X steps on the hold-out eval dataset
:param save_dir: Path to store the final model
:param num_processes: The number of processes for `multiprocessing.Pool` during preprocessing.
Set to value of 1 to disable multiprocessing. When set to 1, you cannot split away a dev set from train set.
Set to None to use all CPU cores minus one.
:param use_amp: Optimization level of NVIDIA's automatic mixed precision (AMP). The higher the level, the faster the model.
Available options:
None (Don't use AMP)
"O0" (Normal FP32 training)
"O1" (Mixed Precision => Recommended)
"O2" (Almost FP16)
"O3" (Pure FP16).
See details on: https://nvidia.github.io/apex/amp.html
:return: None
"""
if dev_filename:
dev_split = 0
if num_processes is None:
num_processes = multiprocessing.cpu_count() - 1 or 1
set_all_seeds(seed=42)
# For these variables, by default, we use the value set when initializing the FARMReader.
# These can also be manually set when train() is called if you want a different value at train vs inference
if use_gpu is None:
use_gpu = self.use_gpu
if max_seq_len is None:
max_seq_len = self.max_seq_len
device, n_gpu = initialize_device_settings(use_cuda=use_gpu,
use_amp=use_amp)
if not save_dir:
save_dir = f"../../saved_models/{self.inferencer.model.language_model.name}"
# 1. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
label_list = ["start_token", "end_token"]
metric = "squad"
processor = SquadProcessor(
tokenizer=self.inferencer.processor.tokenizer,
max_seq_len=max_seq_len,
label_list=label_list,
metric=metric,
train_filename=train_filename,
dev_filename=dev_filename,
dev_split=dev_split,
test_filename=test_filename,
data_dir=Path(data_dir),
)
# 2. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them
# and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
distributed=False,
max_processes=num_processes)
# 3. Create an optimizer and pass the already initialized model
model, optimizer, lr_schedule = initialize_optimizer(
model=self.inferencer.model,
# model=self.inferencer.model,
learning_rate=learning_rate,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": warmup_proportion
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device,
use_amp=use_amp,
)
# 4. 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,
use_amp=use_amp,
disable_tqdm=not self.progress_bar)
# 5. Let it grow!
self.inferencer.model = trainer.train()
self.save(Path(save_dir))
def test_doc_classification(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 8
evaluate_every = 5
lang_model = "bert-base-german-cased"
tokenizer = BertTokenizer.from_pretrained(
pretrained_model_name_or_path=lang_model, do_lower_case=False)
processor = TextClassificationProcessor(tokenizer=tokenizer,
max_seq_len=128,
data_dir="samples/doc_class",
train_filename="train-sample.tsv",
label_list=["OTHER", "OFFENSE"],
metric="f1_macro",
dev_filename="test-sample.tsv",
test_filename=None,
dev_split=0.0,
label_column_name="coarse_label")
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = Bert.load(lang_model)
prediction_head = TextClassificationHead(layer_dims=[
768, len(processor.tasks["text_classification"]["label_list"])
])
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=2e-5,
warmup_proportion=0.1,
n_batches=len(data_silo.loaders["train"]),
n_epochs=1)
trainer = Trainer(optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
evaluate_every=evaluate_every,
device=device)
model = trainer.train(model)
save_dir = "testsave/doc_class"
model.save(save_dir)
processor.save(save_dir)
basic_texts = [{
"text": "Martin Müller spielt Handball in Berlin."
}, {
"text":
"Schartau sagte dem Tagesspiegel, dass Fischer ein Idiot sei."
}, {
"text":
"Franzosen verteidigen 2:1-Führung – Kritische Stimmen zu Schwedens Superstar"
}, {
"text": "Neues Video von Designern macht im Netz die Runde"
}, {
"text":
"23-jähriger Brasilianer muss vier Spiele pausieren – Entscheidung kann noch angefochten werden"
}, {
"text":
"Aufständische verwendeten Chemikalie bei Gefechten im August."
}, {
"text":
"Bewährungs- und Geldstrafe für 26-Jährigen wegen ausländerfeindlicher Äußerung"
}, {
"text":
"ÖFB-Teamspieler nur sechs Minuten nach seinem Tor beim 1:1 gegen Sunderland verletzt ausgewechselt"
}, {
"text":
"Ein 31-jähriger Polizist soll einer 42-Jährigen den Knöchel gebrochen haben"
}, {
"text":
"18 Menschen verschleppt. Kabul – Nach einem Hubschrauber-Absturz im Norden Afghanistans haben Sicherheitskräfte am Mittwoch versucht"
}]
#TODO enable loading here again after we have finished migration towards "processor.tasks"
#inf = Inferencer.load(save_dir)
inf = Inferencer(model=model, processor=processor)
result = inf.run_inference(dicts=basic_texts)
assert result[0]["predictions"][0]["label"] == "OTHER"
assert abs(result[0]["predictions"][0]["probability"] - 0.7) <= 0.1
loaded_processor = TextClassificationProcessor.load_from_dir(save_dir)
inf2 = Inferencer(model=model, processor=loaded_processor)
result_2 = inf2.run_inference(dicts=basic_texts)
pprint(list(zip(result, result_2)))
for r1, r2 in list(zip(result, result_2)):
assert r1 == r2
# if(__name__=="__main__"):
# test_doc_classification()
# b) and a prediction head on top that is suited for our task => Question Answering
prediction_head = QuestionAnsweringHead(layer_dims=[768, len(label_list)])
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
# 5. Create an optimizer
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=3e-5,
warmup_proportion=0.1,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
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
model = trainer.train(model)
prediction_head = QuestionAnsweringHead(layer_dims=[768, len(label_list)])
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=1e-5,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": 0.2
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer(
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
def train_from_scratch():
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="")
ml_logger.init_experiment(experiment_name="from_scratch", run_name="debug")
#########################
######## Settings
########################
set_all_seeds(seed=39)
device, n_gpu = initialize_device_settings(use_cuda=True)
evaluate_every = 5000
vocab_size = 30522
# dev_filename = None
save_dir = Path("saved_models/train_from_scratch")
n_epochs = 10
learning_rate = 1e-4
warmup_proportion = 0.05
batch_size = 16 # (probably only possible via gradient accumulation steps)
max_seq_len = 64
# 1.Create a tokenizer
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = BertStyleLMProcessor(
data_dir=Path("data/lm_finetune_nips"),
tokenizer=tokenizer,
max_seq_len=max_seq_len,
train_filename="train.txt",
dev_split=2000 / 8_000_000,
dev_filename=None,
test_filename=None,
)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and
# calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
distributed=False)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.from_scratch("bert", vocab_size)
# b) and *two* prediction heads on top that are suited for our task => Language Model finetuning
lm_prediction_head = BertLMHead(768, vocab_size)
next_sentence_head = NextSentenceHead([768, 2], task_name="nextsentence")
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=learning_rate,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": warmup_proportion
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device,
grad_acc_steps=8,
)
# 6. Feed everything to the Trainer, which keeps care of growing our model and evaluates it from time to time
trainer = Trainer.create_or_load_checkpoint(
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,
grad_acc_steps=8,
checkpoint_root_dir=Path(
"saved_models/train_from_scratch/checkpoints"),
)
# 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:
model.save(save_dir)
processor.save(save_dir)
def train_evaluation_single(seed=42):
##########################
########## Settings
##########################
set_all_seeds(seed=seed)
device, n_gpu = initialize_device_settings(use_cuda=True)
# GPU utilization on 4x V100
# 40*4, 14.3/16GB on master, 12.6/16 on others
batch_size = 40 * 4
n_epochs = 2
evaluate_every = 2000000 # disabling dev eval
lang_model = "roberta-base"
do_lower_case = False # roberta is a cased model
train_filename = "train-v2.0.json"
dev_filename = "dev-v2.0.json"
# Load model and train
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
processor = SquadProcessor(
tokenizer=tokenizer,
max_seq_len=256,
label_list=["start_token", "end_token"],
metric="squad",
train_filename=train_filename,
dev_filename=dev_filename,
test_filename=None,
data_dir=Path("testsave/data/squad20"),
)
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
distributed=False)
language_model = LanguageModel.load(lang_model)
prediction_head = QuestionAnsweringHead(n_best=5)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token"],
device=device,
)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=3e-5,
schedule_opts={
"name": "LinearWarmup",
"warmup_proportion": 0.2
},
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device)
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,
)
starttime = time()
trainer.train()
elapsed = time() - starttime
save_dir = Path("testsave/roberta-qa-dev")
model.save(save_dir)
processor.save(save_dir)
# Create Evaluator
evaluator = Evaluator(data_loader=data_silo.get_data_loader("dev"),
tasks=data_silo.processor.tasks,
device=device)
results = evaluator.eval(model)
f1_score = results[0]["f1"] * 100
em_score = results[0]["EM"] * 100
tnacc = results[0]["top_n_accuracy"] * 100
print(results)
print(elapsed)
gold_f1 = 82.155
gold_EM = 77.714
gold_tnrecall = 97.3721 #
gold_elapsed = 1135
np.testing.assert_allclose(
f1_score,
gold_f1,
rtol=0.01,
err_msg=f"FARM Training changed for f1 score by: {f1_score - gold_f1}")
np.testing.assert_allclose(
em_score,
gold_EM,
rtol=0.01,
err_msg=f"FARM Training changed for EM by: {em_score - gold_EM}")
np.testing.assert_allclose(
tnacc,
gold_tnrecall,
rtol=0.01,
err_msg=
f"FARM Training changed for top 5 accuracy by: {em_score - gold_EM}")
np.testing.assert_allclose(
elapsed,
gold_elapsed,
rtol=0.1,
err_msg=
f"FARM Training speed changed significantly by: {elapsed - gold_elapsed} seconds"
)
def test_doc_regression(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 1
evaluate_every = 2
lang_model = "bert-base-cased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
processor = RegressionProcessor(tokenizer=tokenizer,
max_seq_len=8,
data_dir="samples/doc_regr",
train_filename="train-sample.tsv",
dev_filename="test-sample.tsv",
test_filename=None,
label_column_name="label")
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = LanguageModel.load(lang_model)
prediction_head = RegressionHead(layer_dims=[768, 1])
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence_continuous"],
device=device)
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=2e-5,
warmup_proportion=0.1,
n_batches=len(data_silo.loaders["train"]),
n_epochs=1)
trainer = Trainer(optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
evaluate_every=evaluate_every,
device=device)
model = trainer.train(model)
save_dir = "testsave/doc_regr"
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{
"text":
"The dress is just fabulous and it totally fits my size. The fabric is of great quality and the seams are really well hidden. I am super happy with this purchase and I am looking forward to trying some more from the same brand."
},
{
"text":
"it just did not fit right. The top is very thin showing everything."
},
]
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
assert isinstance(result[0]["predictions"][0]["pred"], np.float32)
def test_doc_classification(caplog=None):
if caplog:
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 1
batch_size = 1
evaluate_every = 2
lang_model = "distilbert-base-german-cased"
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=False)
processor = TextClassificationProcessor(
tokenizer=tokenizer,
max_seq_len=8,
data_dir=Path("samples/doc_class"),
train_filename=Path("train-sample.tsv"),
label_list=["OTHER", "OFFENSE"],
metric="f1_macro",
dev_filename="test-sample.tsv",
test_filename=None,
dev_split=0.0,
label_column_name="coarse_label")
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = DistilBert.load(lang_model)
prediction_head = TextClassificationHead()
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
n_batches=len(data_silo.loaders["train"]),
n_epochs=1,
device=device,
schedule_opts=None)
trainer = Trainer(optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
device=device)
model = trainer.train(model)
save_dir = Path("testsave/doc_class")
model.save(save_dir)
processor.save(save_dir)
basic_texts = [{
"text": "Malte liebt Berlin."
}, {
"text":
"Schartau sagte dem Tagesspiegel, dass Fischer ein Idiot sei."
}]
inf = Inferencer.load(save_dir, batch_size=2)
result = inf.inference_from_dicts(dicts=basic_texts)
assert isinstance(result[0]["predictions"][0]["probability"], np.float32)
def execML(self, job):
start_time = time.time()
if job.task == 'analyse':
basic_texts = []
# Will donwload and store dataset...
sample = self.downloadAndConvertText(job, job.data_sample)
for text in sample.encode('utf-8').splitlines():
basic_texts.append({'text': text.decode('utf-8')})
# Will donwload and store model...
self.downloadAndStoreZIPModel(job, job.model)
self.updateJobStatus(job, 'analysing')
save_dir = 'tmp/' + job.model['id']
model = Inferencer.load(save_dir)
result = model.inference_from_dicts(dicts=basic_texts)
self.persistResult(job, result)
model.close_multiprocessing_pool()
self.updateJobStatus(job, 'completed')
elif job.task == 'train':
self.updateJobStatus(job, 'training')
# Will donwload and store dataset...
self.downloadAndStoreZIPDataset(job, job.data_source)
# Will donwload and store model...
self.downloadAndStoreZIPModel(job, job.model)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 4
evaluate_every = 400
do_lower_case = False
batch_size = 32
lang_model = os.path.join(Path.cwd(), 'tmp', job.model['id'])
ner_labels = [
"[PAD]", "X", "O", "B-MISC", "I-MISC", "B-PER", "I-PER",
"B-ORG", "I-ORG", "B-LOC", "I-LOC", "B-OTH", "I-OTH"
]
# 1. Create a tokenizer
tokenizer = Tokenizer.load(
pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case,
tokenizer_class='BertTokenizer'
) #tokenizer_class='BertTokenizer'
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = NERProcessor(tokenizer=tokenizer,
max_seq_len=128,
data_dir=str(
os.path.join(Path.cwd(), 'tmp',
job.data_source['id'])),
delimiter=' ',
metric='seq_f1',
label_list=ner_labels)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_processes=1)
# 4. Create an AdaptiveModel
# 4.1 which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# 4.2 and a prediction head on top that is suited for our task => NER
prediction_head = TokenClassificationHead(
num_labels=len(ner_labels))
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=['per_token'],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=1e-5,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
device=device,
)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant 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
trainer.train()
# 8. Hooray! You have a model. Store it:
newModelId = str(uuid.uuid4())
save_dir = 'tmp/' + newModelId
model.save(save_dir)
processor.save(save_dir)
model.close_multiprocessing_pool()
self.persistZIPModel(newModelId, job)
self.updateJobStatus(job, 'completed')
elapsed_time = time.time() - start_time
print('Execution time max: ',
elapsed_time,
'for job.id:',
job.id,
flush=True)
return {'status': True, 'code': 'ok', 'msg': 'success'}
data_silo = DataSilo(processor=processor, batch_size=batch_size)
# loading the pretrained BERT base cased model
language_model = LanguageModel.load(lang_model)
# prediction head for our model that is suited for classifying news article genres
prediction_head = MultiLabelTextClassificationHead(num_labels=len(label_list))
model = AdaptiveModel(language_model=language_model,
prediction_heads=[prediction_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_sequence"],
device=device)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs)
trainer = Trainer(model=model,
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
device=device)
trainer.train()
save_dir = "saved_models/my_model_xlm"
model.save(save_dir)
