# -*- coding:utf-8 -*-
# Author: hankcs
# Date: 2020-12-14 20:34
import logging
from copy import deepcopy
from typing import Union, List, Callable
import torch
from torch.utils.data import DataLoader
from hanlp_common.constant import IDX
from hanlp.common.dataset import PadSequenceDataLoader, SortingSamplerBuilder
from hanlp.common.structure import History
from hanlp.common.torch_component import TorchComponent
from hanlp.common.transform import FieldLength, PunctuationMask
from hanlp.common.vocab import Vocab
from hanlp.components.classifiers.transformer_classifier import TransformerComponent
from hanlp.components.parsers.biaffine.biaffine_dep import BiaffineDependencyParser
from hanlp_common.conll import CoNLLUWord, CoNLLSentence
from hanlp.components.parsers.ud.ud_model import UniversalDependenciesModel
from hanlp.components.parsers.ud.util import generate_lemma_rule, append_bos, sample_form_missing
from hanlp.components.parsers.ud.lemma_edit import apply_lemma_rule
from hanlp.datasets.parsing.loaders.conll_dataset import CoNLLParsingDataset
from hanlp.layers.embeddings.contextual_word_embedding import ContextualWordEmbedding
from hanlp.metrics.accuracy import CategoricalAccuracy
from hanlp.metrics.metric import Metric
from hanlp.metrics.mtl import MetricDict
from hanlp.metrics.parsing.attachmentscore import AttachmentScore
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, merge_dict, reorder
[docs]class UniversalDependenciesParser(TorchComponent):
def __init__(self, **kwargs) -> None:
"""Universal Dependencies Parsing (lemmatization, features, PoS tagging and dependency parsing) implementation
of "75 Languages, 1 Model: Parsing Universal Dependencies Universally" (:cite:`kondratyuk-straka-2019-75`).
Args:
**kwargs: Predefined config.
"""
super().__init__(**kwargs)
self.model: UniversalDependenciesModel = self.model
[docs] def build_dataloader(self,
data,
batch_size,
shuffle=False,
device=None,
logger: logging.Logger = None,
sampler_builder=None,
gradient_accumulation=1,
transformer: ContextualWordEmbedding = None,
**kwargs) -> DataLoader:
transform = [generate_lemma_rule, append_bos, self.vocabs, transformer.transform(), FieldLength('token')]
if not self.config.punct:
transform.append(PunctuationMask('token', 'punct_mask'))
dataset = self.build_dataset(data, transform)
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 = SortingSamplerBuilder(batch_size).build(lens, shuffle, gradient_accumulation)
return PadSequenceDataLoader(dataset, batch_size, shuffle, device=device, batch_sampler=sampler,
pad={'arc': 0}, )
[docs] def build_vocabs(self, trn, logger, **kwargs):
self.vocabs.pos = Vocab(unk_token=None, pad_token=None)
self.vocabs.rel = Vocab(unk_token=None, pad_token=None)
self.vocabs.lemma = Vocab(unk_token=None, pad_token=None)
self.vocabs.feat = Vocab(unk_token=None, pad_token=None)
timer = CountdownTimer(len(trn))
max_seq_len = 0
for each in trn:
max_seq_len = max(max_seq_len, len(each['token']))
timer.log(f'Building vocab [blink][yellow]...[/yellow][/blink] (longest sequence: {max_seq_len})')
for v in self.vocabs.values():
v.set_unk_as_safe_unk()
self.vocabs.lock()
self.vocabs.summary(logger)
def build_dataset(self, data, transform):
dataset = CoNLLParsingDataset(data, transform=transform, prune=sample_form_missing, cache=isinstance(data, str))
return dataset
[docs] def build_optimizer(self, trn, **kwargs):
# noinspection PyCallByClass,PyTypeChecker
return TransformerComponent.build_optimizer(self, trn, **kwargs)
[docs] def build_criterion(self, **kwargs):
pass
[docs] def build_metric(self, **kwargs):
return MetricDict({
'lemmas': CategoricalAccuracy(),
'upos': CategoricalAccuracy(),
'deps': AttachmentScore(),
'feats': CategoricalAccuracy(),
})
[docs] def evaluate_dataloader(self,
data: DataLoader,
criterion: Callable,
metric: MetricDict = None,
output=False,
logger=None,
ratio_width=None,
**kwargs):
metric.reset()
self.model.eval()
timer = CountdownTimer(len(data))
total_loss = 0
for idx, batch in enumerate(data):
out, mask = self.feed_batch(batch)
loss = out['loss']
total_loss += loss.item()
self.decode_output(out, mask, batch)
self.update_metrics(metric, batch, out, mask)
report = f'loss: {total_loss / (idx + 1):.4f} {metric.cstr()}'
timer.log(report, logger=logger, ratio_percentage=False, ratio_width=ratio_width)
del loss
del out
del mask
return total_loss / len(data), metric
# noinspection PyMethodOverriding
[docs] def build_model(self,
transformer: ContextualWordEmbedding,
n_mlp_arc,
n_mlp_rel,
mlp_dropout,
mix_embedding,
layer_dropout,
training=True,
**kwargs) -> torch.nn.Module:
assert bool(transformer.scalar_mix) == bool(mix_embedding), 'transformer.scalar_mix has to be 1 ' \
'when mix_embedding is non-zero.'
# noinspection PyTypeChecker
return UniversalDependenciesModel(transformer.module(training=training),
n_mlp_arc,
n_mlp_rel,
mlp_dropout,
len(self.vocabs.rel),
len(self.vocabs.lemma),
len(self.vocabs.pos),
len(self.vocabs.feat),
mix_embedding,
layer_dropout)
[docs] def predict(self, data: Union[List[str], List[List[str]]], batch_size: int = None, **kwargs):
if not data:
return []
flat = self.input_is_flat(data)
if flat:
data = [data]
samples = self.build_samples(data)
if not batch_size:
batch_size = self.config.batch_size
dataloader = self.build_dataloader(samples,
device=self.devices[0], shuffle=False,
**merge_dict(self.config,
batch_size=batch_size,
overwrite=True,
**kwargs))
order = []
outputs = []
for batch in dataloader:
out, mask = self.feed_batch(batch)
self.decode_output(out, mask, batch)
outputs.extend(self.prediction_to_human(out, batch))
order.extend(batch[IDX])
outputs = reorder(outputs, order)
if flat:
return outputs[0]
return outputs
def build_samples(self, data: List[List[str]]):
return [{'FORM': x} for x in data]
[docs] def fit(self,
trn_data,
dev_data,
save_dir,
transformer: ContextualWordEmbedding,
sampler_builder=None,
mix_embedding: int = 13,
layer_dropout: int = 0.1,
n_mlp_arc=768,
n_mlp_rel=256,
mlp_dropout=.33,
lr=1e-3,
transformer_lr=2.5e-5,
patience=0.1,
batch_size=32,
epochs=30,
gradient_accumulation=1,
adam_epsilon=1e-8,
weight_decay=0,
warmup_steps=0.1,
grad_norm=1.0,
tree=False,
proj=False,
punct=False,
logger=None,
verbose=True,
devices: Union[float, int, List[int]] = None, **kwargs):
return super().fit(**merge_locals_kwargs(locals(), kwargs))
[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, eval_trn=True, **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,
eval_trn=eval_trn, **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, deepcopy(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.cstr()} 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: MetricDict,
logger: logging.Logger,
history: History,
gradient_accumulation=1,
grad_norm=None,
ratio_width=None,
eval_trn=True,
**kwargs):
optimizer, scheduler = optimizer
metric.reset()
self.model.train()
timer = CountdownTimer(history.num_training_steps(len(trn), gradient_accumulation=gradient_accumulation))
total_loss = 0
for idx, batch in enumerate(trn):
out, mask = self.feed_batch(batch)
loss = out['loss']
if gradient_accumulation and gradient_accumulation > 1:
loss /= gradient_accumulation
loss.backward()
total_loss += loss.item()
if eval_trn:
self.decode_output(out, mask, batch)
self.update_metrics(metric, batch, out, mask)
if history.step(gradient_accumulation):
self._step(optimizer, scheduler, grad_norm)
report = f'loss: {total_loss / (idx + 1):.4f} {metric.cstr()}' 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 out
del mask
def decode_output(self, outputs, mask, batch):
arc_scores, rel_scores = outputs['class_probabilities']['deps']['s_arc'], \
outputs['class_probabilities']['deps']['s_rel']
arc_preds, rel_preds = BiaffineDependencyParser.decode(self, arc_scores, rel_scores, mask, batch)
outputs['arc_preds'], outputs['rel_preds'] = arc_preds, rel_preds
return outputs
def update_metrics(self, metrics, batch, outputs, mask):
arc_preds, rel_preds, puncts = outputs['arc_preds'], outputs['rel_preds'], batch.get('punct_mask', None)
BiaffineDependencyParser.update_metric(self, arc_preds, rel_preds, batch['arc'], batch['rel_id'], mask, puncts,
metrics['deps'], batch)
for task, key in zip(['lemmas', 'upos', 'feats'], ['lemma_id', 'pos_id', 'feat_id']):
metric: Metric = metrics[task]
pred = outputs['class_probabilities'][task]
gold = batch[key]
metric(pred.detach(), gold, mask=mask)
return metrics
def feed_batch(self, batch: dict):
mask = self.compute_mask(batch)
output_dict = self.model(batch, mask)
if self.model.training:
mask = mask.clone()
mask[:, 0] = 0
return output_dict, mask
def compute_mask(self, batch):
lens = batch['token_length']
mask = lengths_to_mask(lens)
return mask
def _step(self, optimizer, scheduler, grad_norm):
clip_grad_norm(self.model, grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
def input_is_flat(self, data):
# noinspection PyCallByClass,PyTypeChecker
return BiaffineDependencyParser.input_is_flat(self, data, False)
def prediction_to_human(self, outputs: dict, batch):
arcs, rels = outputs['arc_preds'], outputs['rel_preds']
upos = outputs['class_probabilities']['upos'][:, 1:, :].argmax(-1).tolist()
feats = outputs['class_probabilities']['feats'][:, 1:, :].argmax(-1).tolist()
lemmas = outputs['class_probabilities']['lemmas'][:, 1:, :].argmax(-1).tolist()
lem_vocab = self.vocabs['lemma'].idx_to_token
pos_vocab = self.vocabs['pos'].idx_to_token
feat_vocab = self.vocabs['feat'].idx_to_token
# noinspection PyCallByClass,PyTypeChecker
for tree, form, lemma, pos, feat in zip(BiaffineDependencyParser.prediction_to_head_rel(
self, arcs, rels, batch), batch['token'], lemmas, upos, feats):
form = form[1:]
assert len(form) == len(tree)
lemma = [apply_lemma_rule(t, lem_vocab[r]) for t, r in zip(form, lemma)]
pos = [pos_vocab[x] for x in pos]
feat = [feat_vocab[x] for x in feat]
yield CoNLLSentence(
[CoNLLUWord(id=i + 1, form=fo, lemma=l, upos=p, feats=fe, head=a, deprel=r) for
i, (fo, (a, r), l, p, fe) in enumerate(zip(form, tree, lemma, pos, feat))])
def __call__(self, data, batch_size=None, **kwargs) -> Union[CoNLLSentence, List[CoNLLSentence]]:
return super().__call__(data, batch_size, **kwargs)
