def fit(self, train_sents, dev_sents=None):
for epoch_i in range(1, self.num_epochs + 1):
random.shuffle(train_sents)
num_correct, num_total = self._fit_epoch(train_sents)
print("Epoch {}: {}/{} = {}".format(epoch_i,
num_correct, num_total,
num_correct / num_total))
if dev_sents:
performance_dict = measure_performance(dev_sents, self.parse(dev_sents))
perf_measure = performance_dict['uas']
if perf_measure > self.best_epoch_score:
self.best_epoch = epoch_i
self.best_epoch_score = perf_measure
else:
# We didn't get better: should we stop?
no_improvement_epochs = epoch_i - self.best_epoch
if no_improvement_epochs == self.early_stop_after:
print("No improvements for {} epochs. Stopping".format(self.early_stop_after))
return
print("Dev: {}".format(performance_dict))
def _fit_epoch(self, train_sents):
trainer = ParallelTrain(train_sents, batch_size=self.train_batch_size, parser=self)
parses = trainer.parse()
performance_dict = measure_performance(train_sents, parses)
return performance_dict['ua'], performance_dict['total']
