# -*- coding:utf-8 -*-
# Author: hankcs
# Date: 2020-06-22 20:54
import logging
from copy import copy
from typing import Union, List, Callable, Dict, Any
from bisect import bisect
import torch
import torch.nn.functional as F
from torch import nn
from torch.utils.data import DataLoader
from hanlp_common.constant import IDX, PRED
from hanlp.common.dataset import PadSequenceDataLoader, SamplerBuilder, TransformableDataset
from hanlp.common.structure import History
from hanlp.common.torch_component import TorchComponent
from hanlp.common.transform import FieldLength
from hanlp.common.vocab import Vocab
from hanlp.components.srl.span_bio.baffine_tagging import SpanBIOSemanticRoleLabelingModel
from hanlp.datasets.srl.loaders.conll2012 import CoNLL2012SRLBIODataset
from hanlp.layers.crf.crf import CRF
from hanlp.layers.embeddings.contextual_word_embedding import find_transformer
from hanlp.layers.embeddings.embedding import Embedding
from hanlp.layers.transformers.utils import build_optimizer_scheduler_with_transformer
from hanlp.metrics.chunking.sequence_labeling import get_entities
from hanlp.metrics.f1 import F1
from hanlp.utils.string_util import guess_delimiter
from hanlp.utils.time_util import CountdownTimer
from hanlp.utils.torch_util import clip_grad_norm, lengths_to_mask
from hanlp_common.util import merge_locals_kwargs, reorder
[docs]class SpanBIOSemanticRoleLabeler(TorchComponent):
def __init__(self, **kwargs) -> None:
"""A span based Semantic Role Labeling task using BIO scheme for tagging the role of each token. Given a
predicate and a token, it uses biaffine (:cite:`dozat:17a`) to predict their relations as one of BIO-ROLE.
Args:
**kwargs: Predefined config.
"""
super().__init__(**kwargs)
self.model: SpanBIOSemanticRoleLabelingModel = None
[docs] def build_optimizer(self,
trn,
epochs,
lr,
adam_epsilon,
weight_decay,
warmup_steps,
transformer_lr=None,
gradient_accumulation=1,
**kwargs):
num_training_steps = len(trn) * epochs // gradient_accumulation
if transformer_lr is None:
transformer_lr = lr
transformer = find_transformer(self.model.embed)
optimizer, scheduler = build_optimizer_scheduler_with_transformer(self.model, transformer,
lr, transformer_lr,
num_training_steps, warmup_steps,
weight_decay, adam_epsilon)
return optimizer, scheduler
[docs] def build_criterion(self, decoder=None, **kwargs):
if self.config.crf:
if not decoder:
decoder = self.model.decoder
if isinstance(decoder, torch.nn.DataParallel):
decoder = decoder.module
return decoder.crf
else:
return nn.CrossEntropyLoss(reduction=self.config.loss_reduction)
[docs] def build_metric(self, **kwargs):
return F1()
[docs] def execute_training_loop(self,
trn: DataLoader,
dev: DataLoader,
epochs,
criterion,
optimizer,
metric,
save_dir,
logger: logging.Logger,
devices,
ratio_width=None,
patience=0.5,
**kwargs):
if isinstance(patience, float):
patience = int(patience * epochs)
best_epoch, best_metric = 0, -1
timer = CountdownTimer(epochs)
history = History()
for epoch in range(1, epochs + 1):
logger.info(f"[yellow]Epoch {epoch} / {epochs}:[/yellow]")
self.fit_dataloader(trn, criterion, optimizer, metric, logger, history=history, ratio_width=ratio_width,
**self.config)
loss, dev_metric = self.evaluate_dataloader(dev, criterion, metric, logger=logger, ratio_width=ratio_width)
timer.update()
report = f"{timer.elapsed_human} / {timer.total_time_human} ETA: {timer.eta_human}"
if dev_metric > best_metric:
best_epoch, best_metric = epoch, copy(dev_metric)
self.save_weights(save_dir)
report += ' [red](saved)[/red]'
else:
report += f' ({epoch - best_epoch})'
if epoch - best_epoch >= patience:
report += ' early stop'
logger.info(report)
if epoch - best_epoch >= patience:
break
if not best_epoch:
self.save_weights(save_dir)
elif best_epoch != epoch:
self.load_weights(save_dir)
logger.info(f"Max score of dev is {best_metric} at epoch {best_epoch}")
logger.info(f"Average time of each epoch is {timer.elapsed_average_human}")
logger.info(f"{timer.elapsed_human} elapsed")
# noinspection PyMethodOverriding
[docs] def fit_dataloader(self,
trn: DataLoader,
criterion,
optimizer,
metric,
logger: logging.Logger,
history: History,
gradient_accumulation=1,
grad_norm=None,
ratio_width=None,
eval_trn=False,
**kwargs):
optimizer, scheduler = optimizer
self.model.train()
timer = CountdownTimer(history.num_training_steps(len(trn), gradient_accumulation=gradient_accumulation))
total_loss = 0
for idx, batch in enumerate(trn):
pred, mask = self.feed_batch(batch)
loss = self.compute_loss(criterion, pred, batch['srl_id'], mask)
if gradient_accumulation and gradient_accumulation > 1:
loss /= gradient_accumulation
loss.backward()
total_loss += loss.item()
if eval_trn:
prediction = self.decode_output(pred, mask, batch)
self.update_metrics(metric, prediction, batch)
if history.step(gradient_accumulation):
self._step(optimizer, scheduler, grad_norm)
report = f'loss: {total_loss / (idx + 1):.4f} {metric}' if eval_trn else f'loss: {total_loss / (idx + 1):.4f}'
timer.log(report, logger=logger, ratio_percentage=False, ratio_width=ratio_width)
del loss
del pred
del mask
def naive_decode(self, pred, mask, batch, decoder=None):
vocab = self.vocabs['srl'].idx_to_token
results = []
for sent, matrix in zip(batch['token'], pred.argmax(-1).tolist()):
results.append([])
for token, tags_per_token in zip(sent, matrix):
tags_per_token = [vocab[x] for x in tags_per_token][:len(sent)]
srl_per_token = get_entities(tags_per_token)
results[-1].append(srl_per_token)
return results
def decode_output(self, pred, mask, batch, decoder=None):
# naive = self.naive_decode(pred, mask, batch, decoder)
vocab = self.vocabs['srl'].idx_to_token
if mask is not None:
if self.config.crf:
if not decoder:
decoder = self.model.decoder
crf: CRF = decoder.crf
token_index, mask = mask
pred = crf.decode(pred, mask)
pred = sum(pred, [])
else:
pred = pred[mask].argmax(-1)
pred = pred.tolist()
pred = [vocab[x] for x in pred]
results = []
offset = 0
for sent in batch['token']:
results.append([])
for token in sent:
tags_per_token = pred[offset:offset + len(sent)]
srl_per_token = get_entities(tags_per_token)
results[-1].append(srl_per_token)
offset += len(sent)
assert offset == len(pred)
# assert results == naive
return results
def update_metrics(self, metric, prediction, batch):
for p, g in zip(prediction, batch['srl_set']):
srl = set()
for i, args in enumerate(p):
srl.update((i, start, end, label) for (label, start, end) in args)
metric(srl, g)
return metric
def feed_batch(self, batch: dict):
lens = batch['token_length']
mask2d = lengths_to_mask(lens)
pred = self.model(batch, mask=mask2d)
mask3d = self.compute_mask(mask2d)
if self.config.crf:
token_index = mask3d[0]
pred = pred.flatten(end_dim=1)[token_index]
pred = F.log_softmax(pred, dim=-1)
return pred, mask3d
def compute_mask(self, mask2d):
mask3d = mask2d.unsqueeze_(-1).expand(-1, -1, mask2d.size(1))
mask3d = mask3d & mask3d.transpose(1, 2)
if self.config.crf:
mask3d = mask3d.flatten(end_dim=1)
token_index = mask3d[:, 0]
mask3d = mask3d[token_index]
return token_index, mask3d
else:
return mask3d
def _step(self, optimizer, scheduler, grad_norm):
clip_grad_norm(self.model, grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
# noinspection PyMethodOverriding
[docs] def build_model(self, embed: Embedding, encoder, training, **kwargs) -> torch.nn.Module:
# noinspection PyCallByClass
model = SpanBIOSemanticRoleLabelingModel(
embed.module(training=training, vocabs=self.vocabs),
encoder,
len(self.vocabs.srl),
self.config.n_mlp_rel,
self.config.mlp_dropout,
self.config.crf,
)
return model
# noinspection PyMethodOverriding
[docs] def build_dataloader(self, data, batch_size,
sampler_builder: SamplerBuilder = None,
gradient_accumulation=1,
shuffle=False, device=None, logger: logging.Logger = None,
transform=None,
**kwargs) -> DataLoader:
if isinstance(data, TransformableDataset):
dataset = data
else:
transforms = [self.config.embed.transform(vocabs=self.vocabs), self.vocabs, FieldLength('token')]
if transform:
transforms.insert(0, transform)
dataset = self.build_dataset(data, transforms)
if self.vocabs.mutable:
# noinspection PyTypeChecker
self.build_vocabs(dataset, logger)
lens = [len(x['token_input_ids']) for x in dataset]
if sampler_builder:
sampler = sampler_builder.build(lens, shuffle, gradient_accumulation)
else:
sampler = None
return PadSequenceDataLoader(dataset, batch_size, shuffle, device=device, batch_sampler=sampler)
def build_dataset(self, data, transform):
dataset = CoNLL2012SRLBIODataset(data,
transform=transform,
doc_level_offset=self.config.get('doc_level_offset', True),
cache=isinstance(data, str))
return dataset
[docs] def build_vocabs(self, dataset, logger, **kwargs):
self.vocabs.srl = Vocab(pad_token=None, unk_token=None)
timer = CountdownTimer(len(dataset))
max_seq_len = 0
for sample in dataset:
max_seq_len = max(max_seq_len, len(sample['token_input_ids']))
timer.log(f'Building vocab [blink][yellow]...[/yellow][/blink] (longest sequence: {max_seq_len})')
self.vocabs['srl'].set_unk_as_safe_unk() # C-ARGM-FRQ appears only in test set
self.vocabs.lock()
self.vocabs.summary(logger)
if self.config.get('delimiter') is None:
tokens = dataset[0]['token']
self.config.delimiter = guess_delimiter(tokens)
logger.info(f'Guess the delimiter between tokens could be [blue]"{self.config.delimiter}"[/blue]. '
f'If not, specify `delimiter` in `fit()`')
[docs] def predict(self, data: Union[str, List[str]], batch_size: int = None, **kwargs):
if not data:
return []
flat = self.input_is_flat(data)
if flat:
data = [data]
dataloader = self.build_dataloader(self.build_samples(data), batch_size, device=self.device, **kwargs)
results = []
order = []
for batch in dataloader:
pred, mask = self.feed_batch(batch)
prediction = self.decode_output(pred, mask, batch)
results.extend(self.prediction_to_result(prediction, batch))
order.extend(batch[IDX])
results = reorder(results, order)
if flat:
return results[0]
return results
def build_samples(self, data):
return [{'token': token} for token in data]
# noinspection PyMethodOverriding
[docs] def fit(self,
trn_data,
dev_data,
save_dir,
embed,
encoder=None,
lr=1e-3,
transformer_lr=1e-4,
adam_epsilon=1e-8,
warmup_steps=0.1,
weight_decay=0,
crf=False,
n_mlp_rel=300,
mlp_dropout=0.2,
batch_size=32,
gradient_accumulation=1,
grad_norm=1,
loss_reduction='mean',
epochs=30,
delimiter=None,
doc_level_offset=True,
eval_trn=False,
logger=None,
devices: Union[float, int, List[int]] = None,
transform=None,
**kwargs):
return super().fit(**merge_locals_kwargs(locals(), kwargs))
def compute_loss(self, criterion, pred, srl, mask):
if self.config.crf:
token_index, mask = mask
criterion: CRF = criterion
loss = -criterion.forward(pred, srl.flatten(end_dim=1)[token_index], mask,
reduction=self.config.loss_reduction)
else:
loss = criterion(pred[mask], srl[mask])
return loss
# noinspection PyMethodOverriding
[docs] @torch.no_grad()
def evaluate_dataloader(self, data: DataLoader, criterion: Callable, metric, logger, ratio_width=None,
filename=None, **kwargs):
self.model.eval()
timer = CountdownTimer(len(data))
total_loss = 0
metric.reset()
for idx, batch in enumerate(data):
pred, mask = self.feed_batch(batch)
loss = self.compute_loss(criterion, pred, batch['srl_id'], mask)
total_loss += loss.item()
prediction = self.decode_output(pred, mask, batch)
self.update_metrics(metric, prediction, batch)
report = f'loss: {total_loss / (idx + 1):.4f} {metric}'
timer.log(report, logger=logger, ratio_percentage=False, ratio_width=ratio_width)
return total_loss / timer.total, metric
def input_is_flat(self, data) -> bool:
return isinstance(data[0], str)
def prediction_to_result(self, prediction: List, batch: Dict[str, Any], delimiter=None) -> List:
if delimiter is None:
delimiter = self.config.delimiter
for matrix, tokens in zip(prediction, batch['token']):
result = []
for i, arguments in enumerate(matrix):
if arguments:
pas = [(delimiter.join(tokens[x[1]:x[2]]),) + x for x in arguments]
pas.insert(bisect([a[1] for a in arguments], i), (tokens[i], PRED, i, i + 1))
result.append(pas)
yield result
