def valid_on_epoch(self, epoch, filename, memory):
self.logger.info("Epoch {:02} {} begins validing ...................".format(epoch, self.tag))
self.model.eval()
start_time = time.time()
score = Fscore(self.tag)
datas = ActDataset.read_file(filename, memory)
for pair in datas:
cnet = pair['cnet']
class_string = pair['label']
gold_classes = ActDataset.class_info(class_string)
pred_classes = decode_act(self.model, cnet, memory, self.cuda)
score.update_tp_fp_fn(pred_classes, gold_classes)
fscore = score.output_fscore(self.logger, epoch)
elapsed_time = time.time() - start_time
self.logger.info("Epoch {:02} {} end validing elapsed_time: {:6.0f}s".format(
epoch, self.tag, elapsed_time)
)
self.logger.info('*****************************************************')
return fscore
def valid_on_epoch(self, epoch, filename, memory):
self.logger.info(
"Epoch {:02} {} begins validing ...................".format(
epoch, self.tag))
self.model.eval()
start_time = time.time()
score = Fscore(self.tag)
lines = SLUDataset.read_file(filename)
for (utterance, class_string) in lines:
gold_classes = SLUDataset.class_info(class_string)
pred_classes = decode_slu(self.model, utterance, memory, self.cuda)
score.update_tp_fp_fn(pred_classes, gold_classes)
fscore = score.output_fscore(self.logger, epoch)
elapsed_time = time.time() - start_time
self.logger.info(
"Epoch {:02} {} ends validing elapsed_time: {:6.0f}s".format(
epoch, self.tag, elapsed_time))
self.logger.info(
'*****************************************************')
return fscore
def valid_on_epoch(self, data_iter, epoch):
self.logger.info("Epoch {:02} begins validation .......................".format(epoch))
self.model.eval()
valid_score = Fscore('Valid')
start_time = time.time()
for (i, data_batch) in enumerate(data_iter):
datas, lengths, labels, raw_classes, _ = data_batch
probs = self.model(datas, lengths)
valid_score = self.update_fscore(valid_score, probs.data, raw_classes, data_iter.idx2class)
elapsed_time = time.time() - start_time
self.logger.info("Epoch {:02} Valid, elapsed time: {:6.0f}s".format(
epoch, elapsed_time
))
fscore = valid_score.output_fscore(self.logger, epoch)
return fscore
def train_on_epoch(self, data_iter, epoch):
self.logger.info("Epoch {:02} begins training .......................".format(epoch))
self.model.train()
train_score = Fscore('Train')
total_loss = 0.
total_sent = 0.
start_time = time.time()
for (i, data_batch) in enumerate(data_iter):
# data batch definition
datas, lengths, labels, raw_classes, _ = data_batch
probs = self.model(datas, lengths)
# loss calculation
loss = self.criterion(probs, labels)
# statistics
loss_data = loss.data.clone()
total_loss += loss_data.item()
total_sent += datas.size(0)
train_score = self.update_fscore(train_score, probs.data, raw_classes, data_iter.idx2class)
# update parameters
self.model.zero_grad()
loss.backward()
self.optimizer.step()
elapsed_time = time.time() - start_time
avg_loss = total_loss / total_sent
self.logger.info("Epoch {:02} Train, loss: {:6.2f}; elapsed time: {:6.0f}s".format(
epoch, avg_loss, elapsed_time
))
fscore = train_score.output_fscore(self.logger, epoch)
return fscore