def validation(self):
self.model.eval()
all_prediction, all_ys = [], []
gold_transcripts = []
total_loss = 0.
for step, data in enumerate(self.dev_loader):
bos = self.vocab['<BOS>']
eos = self.vocab['<EOS>']
pad = self.vocab['<PAD>']
xs, ilens, ys, ys_in, ys_out, _, _, trans = to_gpu(
data, bos, eos, pad)
log_probs , prediction, attns, _, _ =\
self.model(xs, ilens, None,
max_dec_timesteps=self.config['max_dec_timesteps'])
seq_len = [y.size(0) + 1 for y in ys]
mask = cc(_seq_mask(seq_len=seq_len, max_len=log_probs.size(1)))
loss = (-torch.sum(log_probs * mask)) / sum(seq_len)
total_loss += loss.item()
all_prediction = all_prediction + prediction.cpu().numpy().tolist()
all_ys = all_ys + [y.cpu().numpy().tolist() for y in ys]
gold_transcripts += trans
avg_loss = total_loss / len(self.dev_loader)
cer, prediction_sents, ground_truth_sents = self.ind2sent(
all_prediction, all_ys)
self.model.train()
return avg_loss, cer, prediction_sents, ground_truth_sents
def mask_and_cal_sum(self, log_probs, ys, mask=None):
if mask is None:
seq_len = [y.size(0) + 1 + 4 for y in ys]
mask = cc(_seq_mask(seq_len=seq_len, max_len=log_probs.size(1)))
else:
seq_len = [y.size(0) for y in ys]
# divide by total length
loss = torch.sum(log_probs * mask) / sum(seq_len)
return loss
def mask_and_cal_loss(self, log_probs, ys, mask=None):
# mask is batch x max_len
# add 1 to EOS
if mask is None:
seq_len = [y.size(0) + 1 for y in ys]
mask = cc(_seq_mask(seq_len=seq_len, max_len=log_probs.size(1)))
else:
seq_len = [y.size(0) for y in ys]
# divide by total length
loss = -torch.sum(log_probs * mask) / sum(seq_len)
return loss
def validation(self):
self.encoder.eval()
self.encoder2.eval()
self.decoder.eval()
self.attention.eval()
self.disen_clean.eval()
self.disen_nuisance.eval()
self.reconstructor.eval()
all_prediction, all_ys = [], []
gold_transcripts = []
total_loss = 0.
for step, data in enumerate(self.dev_loader):
bos = self.vocab['<BOS>']
eos = self.vocab['<EOS>']
pad = self.vocab['<PAD>']
xs, ilens, ys, ys_in, ys_out, _, _, trans = to_gpu(
data, bos, eos, pad)
# input the encoder
clean_repre, enc_lens, _, _ = self.encoder(xs, ilens)
# feeding
logits, log_probs, prediction, attns = \
self.decoder(clean_repre, enc_lens, None,
max_dec_timesteps=self.config['max_dec_timesteps'])
seq_len = [y.size(0) + 1 for y in ys]
mask = cc(_seq_mask(seq_len=seq_len, max_len=log_probs.size(1)))
loss = (-torch.sum(log_probs * mask)) / sum(seq_len)
total_loss += loss.item()
all_prediction = all_prediction + prediction.cpu().numpy().tolist()
all_ys = all_ys + [y.cpu().numpy().tolist() for y in ys]
gold_transcripts += trans
# calculate loss
avg_loss = total_loss / len(self.dev_loader)
cer, prediction_sents, ground_truth_sents = self.ind2sent(
all_prediction, all_ys)
self.encoder.train()
self.encoder2.train()
self.decoder.train()
self.attention.train()
self.disen_clean.train()
self.disen_nuisance.train()
self.reconstructor.train()
return avg_loss, cer, prediction_sents, ground_truth_sents