for i, (x, y, x_lens, y_lens) in tqdm(enumerate(dataloader)):
# x = x.to("cuda")
x = x.to(device)
outs, out_lens = model(x, x_lens)
outs = F.softmax(outs, 1)
outs = outs.transpose(1, 2)
ys = []
offset = 0
for y_len in y_lens:
ys.append(y[offset : offset + y_len])
offset += y_len
out_strings, out_offsets = decoder.decode(outs, out_lens)
y_strings = decoder.convert_to_strings(ys)
for pred, truth in zip(out_strings, y_strings):
trans, ref = pred[0], truth[0]
cer += decoder.cer(trans, ref) / float(len(ref))
cer /= len(dataloader.dataset)
model.train()
return cer
if __name__ == "__main__":
with open("./labels.json") as f:
vocabulary = json.load(f)
vocabulary = "".join(vocabulary)
model = GatedConv(vocabulary)
if(not os.path.exists(save_path)):
os.mkdir(save_path)
model.to(device)
train(model)
cer = 0
print("decoding")
with torch.no_grad():
for i, (x, y, x_lens, y_lens) in tqdm(enumerate(dataloader)):
x = x.to("cuda")
outs, out_lens = model(x, x_lens)
outs = F.softmax(outs, 1)
outs = outs.transpose(1, 2)
ys = []
offset = 0
for y_len in y_lens:
ys.append(y[offset : offset + y_len])
offset += y_len
out_strings, out_offsets = decoder.decode(outs, out_lens)
y_strings = decoder.convert_to_strings(ys)
for pred, truth in zip(out_strings, y_strings):
trans, ref = pred[0], truth[0]
cer += decoder.cer(trans, ref) / float(len(ref))
cer /= len(dataloader.dataset)
model.train()
return cer
if __name__ == "__main__":
with open("data_aishell/labels.json") as f:
vocabulary = json.load(f)
vocabulary = "".join(vocabulary)
model = GatedConv(vocabulary)
model.to("cuda")
train(model)