# -*- coding:utf-8 -*-
# Author: hankcs
# Date: 2020-11-28 21:24
import logging
from typing import Union, List
import torch
from phrasetree.tree import Tree
from torch.utils.data import DataLoader
from hanlp_common.constant import BOS, EOS, IDX
from hanlp.common.dataset import TransformableDataset, SamplerBuilder, PadSequenceDataLoader
from hanlp.common.structure import History
from hanlp.common.torch_component import TorchComponent
from hanlp.common.transform import FieldLength, TransformList
from hanlp.common.vocab import VocabWithNone
from hanlp.components.classifiers.transformer_classifier import TransformerComponent
from hanlp.datasets.parsing.loaders.constituency_dataset import ConstituencyDataset, unpack_tree_to_features, \
build_tree, factorize, remove_subcategory
from hanlp.components.parsers.constituency.crf_constituency_model import CRFConstituencyDecoder, CRFConstituencyModel
from hanlp.metrics.parsing.span import SpanMetric
from hanlp.utils.time_util import CountdownTimer
from hanlp.utils.torch_util import clip_grad_norm
from hanlp_common.util import merge_locals_kwargs, merge_dict, reorder
[docs]class CRFConstituencyParser(TorchComponent):
def __init__(self, **kwargs) -> None:
"""Two-stage CRF Parsing (:cite:`ijcai2020-560`).
Args:
**kwargs: Predefined config.
"""
super().__init__(**kwargs)
self.model: CRFConstituencyModel = self.model
[docs] def build_optimizer(self, trn, **kwargs):
# noinspection PyCallByClass,PyTypeChecker
return TransformerComponent.build_optimizer(self, trn, **kwargs)
[docs] def build_criterion(self, decoder=None, **kwargs):
return decoder
[docs] def build_metric(self, **kwargs):
return SpanMetric()
[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, 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, 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: SpanMetric,
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)
y = batch['chart_id']
loss, span_probs = self.compute_loss(out, y, mask)
if gradient_accumulation and gradient_accumulation > 1:
loss /= gradient_accumulation
loss.backward()
total_loss += loss.item()
if eval_trn:
prediction = self.decode_output(out, mask, batch, span_probs)
self.update_metrics(metric, batch, prediction)
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 out
del mask
def decode_output(self, out, mask, batch, span_probs=None, decoder=None, tokens=None):
s_span, s_label = out
if not decoder:
decoder = self.model.decoder
if mask.any().item():
if span_probs is None:
if self.config.mbr:
s_span = decoder.crf(s_span, mask, mbr=True)
else:
s_span = span_probs
chart_preds = decoder.decode(s_span, s_label, mask)
else:
chart_preds = [[]] * len(tokens)
idx_to_token = self.vocabs.chart.idx_to_token
if tokens is None:
tokens = batch.get('token_', None) # Use the original tokens if any
if tokens is None:
tokens = batch['token']
tokens = [x[1:-1] for x in tokens]
trees = [build_tree(token, [(i, j, idx_to_token[label]) for i, j, label in chart]) for token, chart in
zip(tokens, chart_preds)]
# probs = [prob[:i - 1, 1:i].cpu() for i, prob in zip(lens, s_span.unbind())]
return trees
def update_metrics(self, metric, batch, prediction):
# Add pre-terminals (pos tags) back to prediction for safe factorization (deletion based on pos)
for pred, gold in zip(prediction, batch['constituency']):
pred: Tree = pred
gold: Tree = gold
for p, g in zip(pred.subtrees(lambda t: t.height() == 2), gold.pos()):
token, pos = g
p: Tree = p
assert p.label() == '_'
p.set_label(pos)
metric([factorize(tree, self.config.delete, self.config.equal) for tree in prediction],
[factorize(tree, self.config.delete, self.config.equal) for tree in batch['constituency']])
return metric
def feed_batch(self, batch: dict):
mask = self.compute_mask(batch)
s_span, s_label = self.model(batch)
return (s_span, s_label), mask
def compute_mask(self, batch, offset=1):
lens = batch['token_length'] - offset
seq_len = lens.max()
mask = lens.new_tensor(range(seq_len)) < lens.view(-1, 1, 1)
mask = mask & mask.new_ones(seq_len, seq_len).triu_(1)
return mask
def compute_loss(self, out, y, mask, crf_decoder=None):
if not crf_decoder:
crf_decoder = self.model.decoder
loss, span_probs = crf_decoder.loss(out[0], out[1], y, mask, self.config.mbr)
if loss < 0: # wired negative loss
loss *= 0
return loss, span_probs
def _step(self, optimizer, scheduler, grad_norm):
clip_grad_norm(self.model, grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
[docs] @torch.no_grad()
def evaluate_dataloader(self, data, criterion, logger=None, ratio_width=None, metric=None, output=None, **kwargs):
self.model.eval()
total_loss = 0
if not metric:
metric = self.build_metric()
else:
metric.reset()
timer = CountdownTimer(len(data))
for idx, batch in enumerate(data):
out, mask = self.feed_batch(batch)
y = batch['chart_id']
loss, span_probs = self.compute_loss(out, y, mask)
total_loss += loss.item()
prediction = self.decode_output(out, mask, batch, span_probs)
self.update_metrics(metric, batch, prediction)
timer.log(f'loss: {total_loss / (idx + 1):.4f} {metric}', ratio_percentage=False, logger=logger,
ratio_width=ratio_width)
total_loss /= len(data)
if output:
output.close()
return total_loss, metric
# noinspection PyMethodOverriding
[docs] def build_model(self, encoder, training=True, **kwargs) -> torch.nn.Module:
decoder = CRFConstituencyDecoder(n_labels=len(self.vocabs.chart), n_hidden=encoder.get_output_dim(), **kwargs)
encoder = encoder.module(vocabs=self.vocabs, training=training)
return CRFConstituencyModel(encoder, decoder)
[docs] def build_dataloader(self,
data,
batch_size,
sampler_builder: SamplerBuilder = None,
gradient_accumulation=1,
shuffle=False,
device=None,
logger: logging.Logger = None,
**kwargs) -> DataLoader:
if isinstance(data, TransformableDataset):
dataset = data
else:
transform = self.config.encoder.transform()
if self.config.get('transform', None):
transform = TransformList(self.config.transform, transform)
dataset = self.build_dataset(data, transform, logger)
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)
[docs] def predict(self, data: Union[str, List[str]], **kwargs):
if not data:
return []
flat = self.input_is_flat(data)
if flat:
data = [data]
samples = self.build_samples(data)
dataloader = self.build_dataloader(samples, device=self.device, **kwargs)
outputs = []
orders = []
for idx, batch in enumerate(dataloader):
out, mask = self.feed_batch(batch)
prediction = self.decode_output(out, mask, batch, span_probs=None)
# prediction = [x[0] for x in prediction]
outputs.extend(prediction)
orders.extend(batch[IDX])
outputs = reorder(outputs, orders)
if flat:
return outputs[0]
return outputs
def input_is_flat(self, data):
return isinstance(data[0], str)
def build_samples(self, data):
return [{'token': [BOS] + token + [EOS]} for token in data]
# noinspection PyMethodOverriding
[docs] def fit(self,
trn_data,
dev_data,
save_dir,
encoder,
lr=5e-5,
transformer_lr=None,
adam_epsilon=1e-8,
weight_decay=0,
warmup_steps=0.1,
grad_norm=1.0,
n_mlp_span=500,
n_mlp_label=100,
mlp_dropout=.33,
batch_size=None,
batch_max_tokens=5000,
gradient_accumulation=1,
epochs=30,
patience=0.5,
mbr=True,
sampler_builder=None,
delete=('', ':', '``', "''", '.', '?', '!', '-NONE-', 'TOP', ',', 'S1'),
equal=(('ADVP', 'PRT'),),
no_subcategory=True,
eval_trn=True,
transform=None,
devices=None,
logger=None,
seed=None,
**kwargs):
if isinstance(equal, tuple):
equal = dict(equal)
return super().fit(**merge_locals_kwargs(locals(), kwargs))
def build_dataset(self, data, transform, logger=None):
_transform = [
unpack_tree_to_features,
self.vocabs,
FieldLength('token'),
transform
]
if self.config.get('no_subcategory', True):
_transform.insert(0, remove_subcategory)
dataset = ConstituencyDataset(data,
transform=_transform,
cache=isinstance(data, str))
return dataset
[docs] def build_vocabs(self, trn, logger, **kwargs):
self.vocabs.chart = VocabWithNone(pad_token=None, unk_token=None)
timer = CountdownTimer(len(trn))
max_seq_len = 0
for each in trn:
max_seq_len = max(max_seq_len, len(each['token_input_ids']))
timer.log(f'Building vocab [blink][yellow]...[/yellow][/blink] (longest sequence: {max_seq_len})')
self.vocabs.chart.set_unk_as_safe_unk()
self.vocabs.lock()
self.vocabs.summary(logger)
