async def train(self, sources: Sources):
with self.parent.config.strategy.scope():
self.model = TFAutoModelForTokenClassification.from_pretrained(
self.parent.config.model_name_or_path,
from_pt=self.parent.config.from_pt,
config=self.config,
cache_dir=self.parent.config.cache_dir,
)
data_df = await self._preprocess_data(sources)
train_dataset = self.get_dataset(
data_df,
self.tokenizer,
mode="train",
)
trainer = TFTrainer(
model=self.model,
args=self.parent.config,
train_dataset=train_dataset.get_dataset(),
eval_dataset=None,
compute_metrics=self.compute_metrics,
)
trainer.train()
trainer.save_model()
self.tokenizer.save_pretrained(self.parent.config.output_dir)
def __post_init__(self):
self.tokenizer = AutoTokenizer.from_pretrained(
"ckiplab/bert-base-chinese-ner")
self.model = TFAutoModelForTokenClassification.from_pretrained(
"ckiplab/bert-base-chinese-ner",
from_pt=True,
output_hidden_states=True)
test_X_path = self.model_data_path + "test_X.pkl"
test_mapping_path = self.model_data_path + "test_mapping.pkl"
id2tag_path = self.model_data_path + "id2tag.pkl"
test_X, self.test_mapping = GeneralDataPreprocessor.loadTestArrays(
test_X_path, test_mapping_path)
with open(id2tag_path, "rb") as f:
self.id2tag = pickle.load(f)
test_encodings = self.tokenizer(
test_X,
is_split_into_words=True,
padding=True,
truncation=True,
return_token_type_ids=False,
)
self.test_dataset = tf.data.Dataset.from_tensor_slices(
(dict(test_encodings)))
async def accuracy(self, sources: Sources):
if not os.path.isfile(
os.path.join(self.parent.config.output_dir, "tf_model.h5")
):
raise ModelNotTrained("Train model before assessing for accuracy.")
data_df = await self._preprocess_data(sources)
eval_dataset = self.get_dataset(data_df, self.tokenizer, mode="eval",)
with self.parent.config.strategy.scope():
self.model = TFAutoModelForTokenClassification.from_pretrained(
self.parent.config.output_dir,
config=self.config,
cache_dir=self.parent.config.cache_dir,
)
trainer = TFTrainer(
model=self.model,
args=self.parent.config,
train_dataset=None,
eval_dataset=eval_dataset.get_dataset(),
compute_metrics=self.compute_metrics,
)
result = trainer.evaluate()
return Accuracy(result["eval_f1"])
def __post_init__(self):
self.tokenizer = AutoTokenizer.from_pretrained(
"ckiplab/bert-base-chinese-ner")
self.model = TFAutoModelForTokenClassification.from_pretrained(
"ckiplab/bert-base-chinese-ner",
from_pt=True,
output_hidden_states=True)
def load_ner_model(model_name):
"""Load the model"""
try:
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = TFAutoModelForTokenClassification.from_pretrained(model_name)
return model, tokenizer
except:
return [None] * 3
def __init__(self):
self.modified = False
# load models
self.classifier_tokenizer = BertTokenizer.from_pretrained(
CLASSIFIER_PATH)
self.classifier_config = AutoConfig.from_pretrained(CLASSIFIER_PATH)
self.classifier_model = TFBertForSequenceClassification.from_pretrained(
CLASSIFIER_PATH)
self.ner_tokenizer = AutoTokenizer.from_pretrained(PARSBERTNER_PATH)
self.ner_config = AutoConfig.from_pretrained(PARSBERTNER_PATH)
self.ner_model = TFAutoModelForTokenClassification.from_pretrained(
PARSBERTNER_PATH)
self.weather_api = Weather()
self.adhan_api = Adhan()
self.time_api = Time()
self.calender_api = Calender()
async def predict(
self, sources: SourcesContext
) -> AsyncIterator[Tuple[Record, Any, float]]:
if not os.path.isfile(
os.path.join(self.parent.config.output_dir, "tf_model.h5")
):
raise ModelNotTrained("Train model before prediction.")
with self.parent.config.strategy.scope():
self.model = TFAutoModelForTokenClassification.from_pretrained(
self.parent.config.output_dir,
config=self.config,
cache_dir=self.parent.config.cache_dir,
)
async for record in sources.with_features(
[self.parent.config.words.name]
):
sentence = record.features([self.parent.config.words.name])
df = self.pd.DataFrame(sentence, index=[0])
test_dataset = self.get_dataset(df, self.tokenizer, mode="test",)
trainer = TFTrainer(
model=self.model,
args=self.parent.config,
train_dataset=None,
eval_dataset=None,
compute_metrics=self.compute_metrics,
)
predictions, label_ids, _ = trainer.predict(
test_dataset.get_dataset()
)
preds_list, labels_list = self.align_predictions(
predictions, label_ids
)
preds = [
{word: preds_list[0][i]}
for i, word in enumerate(
sentence[self.parent.config.words.name].split()
)
]
record.predicted(self.parent.config.predict.name, preds, "Nan")
yield record
def main(_):
logging.set_verbosity(logging.INFO)
args = flags.FLAGS.flag_values_dict()
if (os.path.exists(args["output_dir"]) and os.listdir(args["output_dir"])
and args["do_train"] and not args["overwrite_output_dir"]):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome."
.format(args["output_dir"]))
if args["fp16"]:
tf.config.optimizer.set_experimental_options(
{"auto_mixed_precision": True})
if args["tpu"]:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
tpu=args["tpu"])
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
args["n_device"] = args["num_tpu_cores"]
elif len(args["gpus"].split(",")) > 1:
args["n_device"] = len(
[f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
strategy = tf.distribute.MirroredStrategy(
devices=[f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
elif args["no_cuda"]:
args["n_device"] = 1
strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
else:
args["n_device"] = len(args["gpus"].split(","))
strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" +
args["gpus"].split(",")[0])
logging.warning(
"n_device: %s, distributed training: %s, 16-bits training: %s",
args["n_device"],
bool(args["n_device"] > 1),
args["fp16"],
)
labels = get_labels(args["labels"])
num_labels = len(labels)
pad_token_label_id = -1
# IBO
print(args["config_name"]
if args["config_name"] else args["model_name_or_path"])
config = AutoConfig.from_pretrained(
args["config_name"]
if args["config_name"] else args["model_name_or_path"],
num_labels=num_labels,
cache_dir=args["cache_dir"],
)
logging.info("Training/evaluation parameters %s", args)
args["model_type"] = config.model_type
# Training
if args["do_train"]:
tokenizer = AutoTokenizer.from_pretrained(
args["tokenizer_name"]
if args["tokenizer_name"] else args["model_name_or_path"],
do_lower_case=args["do_lower_case"],
cache_dir=args["cache_dir"],
)
with strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
args["model_name_or_path"],
from_pt=bool(".bin" in args["model_name_or_path"]),
config=config,
cache_dir=args["cache_dir"],
)
train_batch_size = args["per_device_train_batch_size"] * args[
"n_device"]
train_dataset, num_train_examples = load_and_cache_examples(
args,
tokenizer,
labels,
pad_token_label_id,
train_batch_size,
mode="train")
train_dataset = strategy.experimental_distribute_dataset(train_dataset)
train(
args,
strategy,
train_dataset,
tokenizer,
model,
num_train_examples,
labels,
train_batch_size,
pad_token_label_id,
)
os.makedirs(args["output_dir"], exist_ok=True)
logging.info("Saving model to %s", args["output_dir"])
model.save_pretrained(args["output_dir"])
tokenizer.save_pretrained(args["output_dir"])
# Evaluation
if args["do_eval"]:
tokenizer = AutoTokenizer.from_pretrained(
args["output_dir"], do_lower_case=args["do_lower_case"])
checkpoints = []
results = []
if args["eval_all_checkpoints"]:
checkpoints = list(
os.path.dirname(c) for c in sorted(
glob.glob(args["output_dir"] + "/**/" + TF2_WEIGHTS_NAME,
recursive=True),
key=lambda f: int("".join(filter(str.isdigit, f)) or -1),
))
logging.info("Evaluate the following checkpoints: %s", checkpoints)
if len(checkpoints) == 0:
checkpoints.append(args["output_dir"])
for checkpoint in checkpoints:
global_step = checkpoint.split("-")[-1] if re.match(
".*checkpoint-[0-9]", checkpoint) else "final"
with strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
checkpoint)
y_true, y_pred, eval_loss = evaluate(args,
strategy,
model,
tokenizer,
labels,
pad_token_label_id,
mode="dev")
report = metrics.classification_report(y_true, y_pred, digits=4)
if global_step:
results.append({
global_step + "_report": report,
global_step + "_loss": eval_loss
})
output_eval_file = os.path.join(args["output_dir"], "eval_results.txt")
with tf.io.gfile.GFile(output_eval_file, "w") as writer:
for res in results:
for key, val in res.items():
if "loss" in key:
logging.info(key + " = " + str(val))
writer.write(key + " = " + str(val))
writer.write("\n")
else:
logging.info(key)
logging.info("\n" + report)
writer.write(key + "\n")
writer.write(report)
writer.write("\n")
if args["do_predict"]:
tokenizer = AutoTokenizer.from_pretrained(
args["output_dir"], do_lower_case=args["do_lower_case"])
model = TFAutoModelForTokenClassification.from_pretrained(
args["output_dir"])
eval_batch_size = args["per_device_eval_batch_size"] * args["n_device"]
predict_dataset, _ = load_and_cache_examples(args,
tokenizer,
labels,
pad_token_label_id,
eval_batch_size,
mode="test")
y_true, y_pred, pred_loss = evaluate(args,
strategy,
model,
tokenizer,
labels,
pad_token_label_id,
mode="test")
output_test_results_file = os.path.join(args["output_dir"],
"test_results.txt")
output_test_predictions_file = os.path.join(args["output_dir"],
"test_predictions.txt")
report = metrics.classification_report(y_true, y_pred, digits=4)
with tf.io.gfile.GFile(output_test_results_file, "w") as writer:
report = metrics.classification_report(y_true, y_pred, digits=4)
logging.info("\n" + report)
writer.write(report)
writer.write("\n\nloss = " + str(pred_loss))
with tf.io.gfile.GFile(output_test_predictions_file, "w") as writer:
with tf.io.gfile.GFile(os.path.join(args["data_dir"], "test.txt"),
"r") as f:
example_id = 0
for line in f:
if line.startswith(
"-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not y_pred[example_id]:
example_id += 1
elif y_pred[example_id]:
output_line = line.split(
)[0] + " " + y_pred[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logging.warning(
"Maximum sequence length exceeded: No prediction for '%s'.",
line.split()[0])
def main():
# region Argument Parsing
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_ner",
model_args,
data_args,
framework="tensorflow")
# endregion
# region Setup logging
# we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# If passed along, set the training seed now.
if training_args.seed is not None:
set_seed(training_args.seed)
# endregion
# region Loading datasets
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
if raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if data_args.text_column_name is not None:
text_column_name = data_args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if data_args.label_column_name is not None:
label_column_name = data_args.label_column_name
elif f"{data_args.task_name}_tags" in column_names:
label_column_name = f"{data_args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# endregion
# region Load config and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
num_labels=num_labels)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
num_labels=num_labels)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
if not tokenizer_name_or_path:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True,
add_prefix_space=True)
else:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True)
# endregion
# region Preprocessing the raw datasets
# First we tokenize all the texts.
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
max_length=data_args.max_length,
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.
label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
if data_args.max_train_samples is not None:
max_train_samples = min(len(train_dataset),
data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
if data_args.max_eval_samples is not None:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# endregion
with training_args.strategy.scope():
# region Initialize model
if model_args.model_name_or_path:
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
config=config,
)
else:
logger.info("Training new model from scratch")
model = TFAutoModelForTokenClassification.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# endregion
# region Create TF datasets
# We need the DataCollatorForTokenClassification here, as we need to correctly pad labels as
# well as inputs.
collate_fn = DataCollatorForTokenClassification(tokenizer=tokenizer,
return_tensors="tf")
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
dataset_options = tf.data.Options()
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
# model.prepare_tf_dataset() wraps a Hugging Face dataset in a tf.data.Dataset which is ready to use in
# training. This is the recommended way to use a Hugging Face dataset when training with Keras. You can also
# use the lower-level dataset.to_tf_dataset() method, but you will have to specify things like column names
# yourself if you use this method, whereas they are automatically inferred from the model input names when
# using model.prepare_tf_dataset()
# For more info see the docs:
# https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset
# https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.to_tf_dataset
tf_train_dataset = model.prepare_tf_dataset(
train_dataset,
collate_fn=collate_fn,
batch_size=total_train_batch_size,
shuffle=True,
).with_options(dataset_options)
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
tf_eval_dataset = model.prepare_tf_dataset(
eval_dataset,
collate_fn=collate_fn,
batch_size=total_eval_batch_size,
shuffle=False,
).with_options(dataset_options)
# endregion
# region Optimizer, loss and compilation
num_train_steps = int(
len(tf_train_dataset) * training_args.num_train_epochs)
if training_args.warmup_steps > 0:
num_warmup_steps = training_args.warmup_steps
elif training_args.warmup_ratio > 0:
num_warmup_steps = int(num_train_steps *
training_args.warmup_ratio)
else:
num_warmup_steps = 0
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
adam_global_clipnorm=training_args.max_grad_norm,
)
model.compile(optimizer=optimizer, jit_compile=training_args.xla)
# endregion
# Metrics
metric = evaluate.load("seqeval")
def get_labels(y_pred, y_true):
# Transform predictions and references tensos to numpy arrays
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
true_labels = [[
label_list[l] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# endregion
# region Preparing push_to_hub and model card
push_to_hub_model_id = training_args.push_to_hub_model_id
model_name = model_args.model_name_or_path.split("/")[-1]
if not push_to_hub_model_id:
if data_args.dataset_name is not None:
push_to_hub_model_id = f"{model_name}-finetuned-{data_args.dataset_name}"
else:
push_to_hub_model_id = f"{model_name}-finetuned-token-classification"
model_card_kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "token-classification"
}
if data_args.dataset_name is not None:
model_card_kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
model_card_kwargs[
"dataset_args"] = data_args.dataset_config_name
model_card_kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
model_card_kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
callbacks = [
PushToHubCallback(
output_dir=training_args.output_dir,
model_id=push_to_hub_model_id,
organization=training_args.push_to_hub_organization,
token=training_args.push_to_hub_token,
tokenizer=tokenizer,
**model_card_kwargs,
)
]
else:
callbacks = []
# endregion
# region Training
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(f" Total train batch size = {total_train_batch_size}")
# Only show the progress bar once on each machine.
model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks,
)
# endregion
# region Predictions
# If you have variable batch sizes (i.e. not using pad_to_max_length), then
# this bit might fail on TF < 2.8 because TF can't concatenate outputs of varying seq
# length from predict().
try:
predictions = model.predict(
tf_eval_dataset,
batch_size=training_args.per_device_eval_batch_size)["logits"]
except tf.python.framework.errors_impl.InvalidArgumentError:
raise ValueError(
"Concatenating predictions failed! If your version of TensorFlow is 2.8.0 or older "
"then you will need to use --pad_to_max_length to generate predictions, as older "
"versions of TensorFlow cannot concatenate variable-length predictions as RaggedTensor."
)
if isinstance(predictions, tf.RaggedTensor):
predictions = predictions.to_tensor(default_value=-100)
predictions = tf.math.argmax(predictions, axis=-1).numpy()
if "label" in eval_dataset:
labels = eval_dataset.with_format("tf")["label"]
else:
labels = eval_dataset.with_format("tf")["labels"]
if isinstance(labels, tf.RaggedTensor):
labels = labels.to_tensor(default_value=-100)
labels = labels.numpy()
attention_mask = eval_dataset.with_format("tf")["attention_mask"]
if isinstance(attention_mask, tf.RaggedTensor):
attention_mask = attention_mask.to_tensor(default_value=-100)
attention_mask = attention_mask.numpy()
labels[attention_mask == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
eval_metric = compute_metrics()
logger.info("Evaluation metrics:")
for key, val in eval_metric.items():
logger.info(f"{key}: {val:.4f}")
if training_args.output_dir is not None:
output_eval_file = os.path.join(training_args.output_dir,
"all_results.json")
with open(output_eval_file, "w") as writer:
writer.write(json.dumps(eval_metric))
# endregion
if training_args.output_dir is not None and not training_args.push_to_hub:
# If we're not pushing to hub, at least save a local copy when we're done
model.save_pretrained(training_args.output_dir)
# from transformers import pipeline
# nlp = pipeline("sentiment-analysis")
# print(nlp("I hate you"))
# print(nlp("I love you"))
# nlp = pipeline("ner")
# sequence = "Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, therefore very" \ "close to the Manhattan Bridge which is visible from the window."
# print(nlp(sequence))
from transformers import TFAutoModelForTokenClassification, AutoTokenizer
import tensorflow as tf
model = TFAutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
label_list = ["O","B-MISC", "I-MISC","B-PER","I-PER","B-ORG","I-ORG","B-LOC","I-LOC"]
sequence = "Hugging Face Inc. is a company based in New York City. Its headquarters are in DUMBO, therefore very" \
"close to the Manhattan Bridge."
tokens = tokenizer.tokenize(tokenizer.decode(tokenizer.encode(sequence)))
inputs = tokenizer.encode(sequence, return_tensors="tf")
outputs = model(inputs)[0]
predictions = tf.argmax(outputs, axis=2)
print(tokens)
print(inputs)
print(outputs)
print(predictions)
def main():
# region Argument Parsing
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# endregion
# region Setup logging
# we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# If passed along, set the training seed now.
if training_args.seed is not None:
set_seed(training_args.seed)
# endregion
# region Loading datasets
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
if raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if data_args.text_column_name is not None:
text_column_name = data_args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if data_args.label_column_name is not None:
label_column_name = data_args.label_column_name
elif f"{data_args.task_name}_tags" in column_names:
label_column_name = f"{data_args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# endregion
# region Load config and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
num_labels=num_labels)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
num_labels=num_labels)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
if not tokenizer_name_or_path:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True,
add_prefix_space=True)
else:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True)
# endregion
# region Preprocessing the raw datasets
# First we tokenize all the texts.
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
max_length=data_args.max_length,
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.
label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# endregion
with training_args.strategy.scope():
# region Initialize model
if model_args.model_name_or_path:
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
config=config,
)
else:
logger.info("Training new model from scratch")
model = TFAutoModelForTokenClassification.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# endregion
# region Create TF datasets
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
train_batches_per_epoch = len(train_dataset) // total_train_batch_size
tf_train_dataset = dataset_to_tf(
train_dataset,
tokenizer,
total_batch_size=total_train_batch_size,
num_epochs=training_args.num_train_epochs,
shuffle=True,
)
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
eval_batches_per_epoch = len(eval_dataset) // total_eval_batch_size
tf_eval_dataset = dataset_to_tf(
eval_dataset,
tokenizer,
total_batch_size=total_eval_batch_size,
num_epochs=training_args.num_train_epochs,
shuffle=False,
)
# endregion
# region Optimizer, loss and compilation
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=int(training_args.num_train_epochs *
train_batches_per_epoch),
num_warmup_steps=training_args.warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
)
def dummy_loss(y_true, y_pred):
return tf.reduce_mean(y_pred)
model.compile(loss={"loss": dummy_loss}, optimizer=optimizer)
# endregion
# Metrics
metric = load_metric("seqeval")
def get_labels(y_pred, y_true):
# Transform predictions and references tensos to numpy arrays
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
true_labels = [[
label_list[l] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# endregion
# region Training
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(f" Total train batch size = {total_train_batch_size}")
# Only show the progress bar once on each machine.
model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
epochs=int(training_args.num_train_epochs),
steps_per_epoch=train_batches_per_epoch,
validation_steps=eval_batches_per_epoch,
)
# endregion
# region Predictions
# For predictions, we preload the entire validation set - note that if you have a really giant validation
# set, you might need to change this!
eval_inputs = {
key: tf.ragged.constant(eval_dataset[key]).to_tensor()
for key in eval_dataset.features
}
predictions = model.predict(
eval_inputs,
batch_size=training_args.per_device_eval_batch_size)["logits"]
predictions = tf.math.argmax(predictions, axis=-1)
labels = np.array(eval_inputs["labels"])
labels[np.array(eval_inputs["attention_mask"]) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
eval_metric = compute_metrics()
logger.info("Evaluation metrics:")
for key, val in eval_metric.items():
logger.info(f"{key}: {val:.4f}")
# endregion
# We don't do predictions in the strategy scope because there are some issues in there right now.
# They'll get fixed eventually, promise!
if training_args.output_dir is not None:
model.save_pretrained(training_args.output_dir)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = get_labels(data_args.labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (
TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
)
if training_args.do_train
else None
)
eval_dataset = (
TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
)
if training_args.do_eval
else None
)
def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -1:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset.get_dataset() if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
return results
# # save recognized organizations in new column
# news['entities'] = news.Article.apply(get_entities)
# huggingface NER pipeline
from transformers import TokenClassificationPipeline, TFAutoModelForTokenClassification, AutoTokenizer
# from transformers import pipeline
# nlp = pipeline('ner')
# print(pd.DataFrame(nlp(sentence))) # gives out tokens and labes
sentence = 'Apple and Microsoft plan to form a joint venture for the development of cloud-based computing ' \
'infrastrucutre.'
## BERT tokenizer and token classification
nlp = TokenClassificationPipeline(model=TFAutoModelForTokenClassification.from_pretrained(
'distilbert-base-cased'), tokenizer=AutoTokenizer.from_pretrained('distilbert-base-cased'),
framework='tf')
print(pd.DataFrame(nlp(sentence)))
from transformers import DistilBertTokenizer, TFDistilBertForTokenClassification
import tensorflow as tf
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-cased')
model = TFDistilBertForTokenClassification.from_pretrained('distilbert-base-cased')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] # Batch size 1
labels = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids))) # Batch size 1
print(model(input_ids))
import numpy as np
from transformers import AutoTokenizer, pipeline, TFDistilBertModel
tokenizer = AutoTokenizer.from_pretrained('distilbert-base-cased')
def main(_):
logging.set_verbosity(logging.INFO)
args = flags.FLAGS.flag_values_dict()
if args["fp16"]:
tf.config.optimizer.set_experimental_options(
{"auto_mixed_precision": True})
if args["tpu"]:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
tpu=args["tpu"])
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
args["n_device"] = args["num_tpu_cores"]
elif len(args["gpus"].split(",")) > 1:
args["n_device"] = len(
[f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
strategy = tf.distribute.MirroredStrategy(
devices=[f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
elif args["no_cuda"]:
args["n_device"] = 1
strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
else:
args["n_device"] = len(args["gpus"].split(","))
strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" +
args["gpus"].split(",")[0])
logging.warning(
"n_device: %s, distributed training: %s, 16-bits training: %s",
args["n_device"],
bool(args["n_device"] > 1),
args["fp16"],
)
labels = get_labels(args["labels"])
pad_token_label_id = -1
logging.info("predict parameters %s", args)
tokenizer = AutoTokenizer.from_pretrained(
args["output_dir"], do_lower_case=args["do_lower_case"])
model = TFAutoModelForTokenClassification.from_pretrained(
args["output_dir"])
while True:
print('Input chinese sentence:')
line = str(input())
if line == 'quit':
break
if len(line) < 1:
print(
'Please input a chinese sentence or "quit" to break this loop:'
)
continue
examples = read_examples_from_line(line)
features = convert_examples_to_features(
examples,
labels,
args["max_seq_length"],
tokenizer,
cls_token_at_end=bool(args["model_type"] in ["xlnet"]),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if args["model_type"] in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=bool(args["model_type"] in ["roberta"]),
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
pad_on_left=bool(args["model_type"] in ["xlnet"]),
# pad on the left for xlnet
pad_token=tokenizer.pad_token_id,
pad_token_segment_id=tokenizer.pad_token_type_id,
pad_token_label_id=pad_token_label_id,
)
feature = features[0]
X = collections.OrderedDict()
X["input_ids"] = tf.train.Feature(int64_list=tf.train.Int64List(
value=list(feature.input_ids)))
X["input_mask"] = tf.train.Feature(int64_list=tf.train.Int64List(
value=list(feature.input_mask)))
X["segment_ids"] = tf.train.Feature(int64_list=tf.train.Int64List(
value=list(feature.segment_ids)))
X["label_ids"] = tf.train.Feature(int64_list=tf.train.Int64List(
value=list(feature.label_ids)))
tf_example = tf.train.Example(features=tf.train.Features(feature=X))
tf_example = tf_example.SerializeToString()
max_seq_length = args["max_seq_length"]
name_to_features = {
"input_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
"input_mask": tf.io.FixedLenFeature([max_seq_length], tf.int64),
"segment_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
"label_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
}
def _decode_record(record):
example = tf.io.parse_single_example(record, name_to_features)
features = {}
features["input_ids"] = example["input_ids"]
features["input_mask"] = example["input_mask"]
features["segment_ids"] = example["segment_ids"]
return features, example["label_ids"]
dataset = []
dataset.append(tf_example)
dataset = tf.data.Dataset.from_tensor_slices(dataset)
dataset = dataset.map(_decode_record)
batch_size = 1
dataset = dataset.batch(batch_size)
eval_features, eval_labels = iter(dataset).next()
inputs = {
"attention_mask": eval_features["input_mask"],
"training": False
}
if args["model_type"] != "distilbert":
inputs["token_type_ids"] = (eval_features["segment_ids"]
if args["model_type"]
in ["bert", "xlnet"] else None)
with strategy.scope():
logits = model(eval_features["input_ids"], **inputs)[0]
active_loss = tf.reshape(eval_labels, (-1, )) != pad_token_label_id
preds = logits.numpy()
label_ids = eval_labels.numpy()
preds = np.argmax(preds, axis=2)
y_pred = [[] for _ in range(label_ids.shape[0])]
for i in range(label_ids.shape[0]):
for j in range(label_ids.shape[1]):
if label_ids[i, j] != pad_token_label_id:
y_pred[i].append(labels[preds[i, j]])
tokens = tokenizer.tokenize(line)
print('## tokens = %s' % tokens)
print('## y_pred = %s' % y_pred)
print('## %s = %s' % (len(tokens), len(y_pred[0])))
word_group = []
subword = {}
def _add_word(subword):
word_group.append(subword['token'] + '/' + subword['flag'])
subword.clear()
for i, token in enumerate(tokens):
flag = y_pred[0][i]
print('## %s = %s' % (token, flag))
if flag.startswith('B'):
if len(subword) > 0:
_add_word(subword)
subword['token'] = token
subword['flag'] = flag
elif flag.startswith('I'):
if (len(subword) > 0 and (y_pred[0][i - 1].startswith('I')
or y_pred[0][i - 1].startswith('B'))
and (y_pred[0][i - 1][1:] == flag[1:])):
subword['token'] = subword['token'] + token
continue
elif len(subword) > 0:
_add_word(subword)
subword['token'] = token
subword['flag'] = flag
else:
if len(subword) > 0:
_add_word(subword)
subword['token'] = token
subword['flag'] = flag
_add_word(subword)
if len(subword) > 0:
_add_word(subword)
print('## word_group = %s' % word_group)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
module = import_module("tasks")
try:
token_classification_task_clazz = getattr(module, model_args.task_type)
token_classification_task: TokenClassificationTask = token_classification_task_clazz()
except AttributeError:
raise ValueError(
f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.DEBUG,
)
"""
logging.basicConfig(
filename="/scratch/project_2001426/harttu/july-2020/transformers-ner/test.log",
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.DEBUG,
filemode='w'
)
"""
logger.info("FROM run_tf_ner.py")
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = token_classification_task.get_labels(data_args.labels)
print("LABELS:")
print(labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
print("LABEL_MAP:")
print(label_map)
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
print("CONFIG")
print(config)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=True,#bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
print("SETTING DATASETS")
# Get datasets
train_dataset = (
TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
)
if training_args.do_train
else None
)
eval_dataset = (
TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
)
if training_args.do_eval
else None
)
def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -100:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
"""
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
"""
import tensorflow as tf
initial_learning_rate=0.001
decay_steps=10000
print("INITIALIZING TRAINER")
# Initialize our Trainer
trainer = TFTrainer(
model=model,
#optimizers=(tf.keras.optimizers.Adam(
#learning_rate=initial_learning_rate, beta_1=0.9,
#beta_2=0.999, epsilon=1e-07, amsgrad=False,
#name='Adam'),tf.keras.optimizers.schedules.PolynomialDecay(
#initial_learning_rate, decay_steps,
#end_learning_rate=0.0001, power=1.0,
#cycle=False, name=None
#)),
args=training_args,
train_dataset=train_dataset.get_dataset() if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
print("TRAINING")
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
print("EVALUATING")
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
print("PREDICTING")
if training_args.do_predict:
test_dataset = TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
return results
def main(self):
model_args = ModelArguments(model_name_or_path=self.model_name_or_path)
data_args = DataTrainingArguments(data_dir=self.data_dir,
labels='./labels.txt',
max_seq_length=self.max_seq_length)
training_args = TFTrainingArguments(
output_dir=self.output_dir,
do_eval=self.do_eval,
do_predict=self.do_predict,
do_train=self.do_train,
per_device_train_batch_size=self.per_device_train_batch_size,
save_steps=self.save_steps,
seed=self.seed)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
module = import_module("tasks")
try:
token_classification_task_clazz = getattr(module,
model_args.task_type)
token_classification_task: TokenClassificationTask = token_classification_task_clazz(
)
except AttributeError:
raise ValueError(
f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = token_classification_task.get_labels(data_args.labels)
# JQ
labels.extend('PAD')
label_map: Dict[int,
str] = {i: label
for i, label in enumerate(labels)}
# JQ
label_map[9] = 'PAD'
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
# START HERE
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i
for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
) if training_args.do_train else None)
eval_dataset = (TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
) if training_args.do_eval else None)
def align_predictions(
predictions: np.ndarray,
label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -100:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(
p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset.get_dataset()
if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(
test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir,
"test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(
training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"),
"r") as f:
example_id = 0
for line in f:
if line.startswith(
"-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split(
)[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning(
"Maximum sequence length exceeded: No prediction for '%s'.",
line.split()[0])
return results