def cal_metrics(self, output, input):
output_id = output.topk(1)[1].squeeze(-1)
pack = Pack()
loss = calculate_loss(output, input.tgt_for_metric)
assert not t.isinf(loss)
output_str = [self.vocab.convert_id2str(i) for i in output_id]
tgt_str = [self.vocab.convert_id2str(i) for i in input.tgt_for_metric]
cer = sum(
[calculate_cer(i[0], i[1])
for i in zip(output_str, tgt_str)]) * 100 / len(output_str)
pack.add(loss=loss, cer=t.Tensor([cer]))
return pack
def cal_metrics(self, output, input):
pack = Pack()
pred, gold = output.pred, output.gold
loss, n_correct = cal_performance(pred, gold)
output_id = pred.topk(1)[1].squeeze(-1)
assert not t.isinf(loss)
output_str = [self.vocab.convert_id2str(i) for i in output_id]
tgt_str = [self.vocab.convert_id2str(i) for i in gold]
cer = sum(
[calculate_cer(i[0], i[1])
for i in zip(output_str, tgt_str)]) * 100 / len(output_str)
pack.add(loss=loss, cer=t.Tensor([cer]))
#pack.add(loss=loss)
return pack
def forward(self, input):
wave, wave_len, text_for_input, text_len = input.wave, input.wave_len, input.tgt_for_input, input.tgt_len
"""
# Args:
# padded_input: N x Ti x D
# input_lengths: N
# padded_targets: N x To
# """
encoder_padded_outputs, *_ = self.encoder(wave, wave_len)
# pred is score before softmax
pred, gold, *_ = self.decoder(text_for_input, encoder_padded_outputs,
text_len)
pack = Pack()
pack.add(pred=pred, gold=gold)
return pack
def __call__(self, batch):
pack = Pack()
wave = [i[0] for i in batch]
tgt_for_input = [i[1] for i in batch]
tgt_for_metric = [i[2] for i in batch]
wave, wave_len = Padder.pad_tri(wave, 0)
tgt_for_input, tgt_for_input_len = Padder.pad_two(tgt_for_input, 0)
tgt_for_metric, tgt_for_metric_len = Padder.pad_two(tgt_for_metric, 0)
pack.add(wave=wave,
tgt_for_input=tgt_for_input.long(),
wave_len=t.Tensor(wave_len).long(),
tgt_len=t.Tensor(tgt_for_input_len).long())
pack.add(tgt_for_metric=tgt_for_metric.long())
if self.use_cuda:
pack = pack.cuda()
return pack
def cal_metric(self, model_output):
metrics = Pack()
metrics.add(loss=model_output.wave.sum())
return metrics, None
def forward(self, pack):
output = Pack()
wave, wave_length = pack.wave, pack.wave_len
output.add(wave=wave)
return output