def get_denoised(prob, ctc_bs=False):
if ctc_bs: # Using ctc beam search before denoising yields only limited improvements a is very slow
text = get_beam_search(prob)
else:
text = get_arg_max(prob)
src_seq, src_valid_length = encode_char(text)
src_seq = mx.nd.array([src_seq], ctx=ctx)
src_valid_length = mx.nd.array(src_valid_length, ctx=ctx)
encoder_outputs, _ = denoiser.encode(src_seq, valid_length=src_valid_length)
states = denoiser.decoder.init_state_from_encoder(encoder_outputs,
encoder_valid_length=src_valid_length)
inputs = mx.nd.full(shape=(1,), ctx=src_seq.context, dtype=np.float32, val=BOS)
output = generator.generate_sequences(inputs, states, text)
return output.strip()
def get_sentence(net, sentence):
scorer = nlp.model.BeamSearchScorer(alpha=0, K=2, from_logits=False)
beam_sampler = nlp.model.BeamSearchSampler(beam_size=5,
decoder=net.decode_logprob,
eos_id=EOS,
scorer=scorer,
max_length=150)
src_seq, src_valid_length = encode_char(sentence)
src_seq = mx.nd.array([src_seq], ctx=ctx)
src_valid_length = mx.nd.array(src_valid_length, ctx=ctx)
encoder_outputs, _ = net.encode(src_seq, valid_length=src_valid_length)
states = net.decoder.init_state_from_encoder(encoder_outputs,
encoder_valid_length=src_valid_length)
inputs = mx.nd.full(shape=(1,), ctx=src_seq.context, dtype=np.float32, val=BOS)
samples, scores, valid_lengths = beam_sampler(inputs, states)
samples = samples[0].asnumpy()
scores = scores[0].asnumpy()
valid_lengths = valid_lengths[0].asnumpy()
return decode_char(samples[0])
src_seq = mx.nd.array([src_seq], ctx=ctx)
src_valid_length = mx.nd.array(src_valid_length, ctx=ctx)
encoder_outputs, _ = denoiser.encode(src_seq,
valid_length=src_valid_length)
states = denoiser.decoder.init_state_from_encoder(
encoder_outputs, encoder_valid_length=src_valid_length)
inputs = mx.nd.full(shape=(1, ),
ctx=src_seq.context,
dtype=np.float32,
val=BOS)
output = generator.generate_sequences(inputs, states, text)
return output.strip()
sentence = "This sentnce has an eror"
src_seq, src_valid_length = encode_char(sentence)
src_seq = mx.nd.array([src_seq], ctx=ctx)
src_valid_length = mx.nd.array(src_valid_length, ctx=ctx)
encoder_outputs, _ = denoiser.encode(src_seq, valid_length=src_valid_length)
states = denoiser.decoder.init_state_from_encoder(
encoder_outputs, encoder_valid_length=src_valid_length)
inputs = mx.nd.full(shape=(1, ),
ctx=src_seq.context,
dtype=np.float32,
val=BOS)
print(sentence)
print("Choice")
print(generator.generate_sequences(inputs, states, sentence))
sentence = "This sentnce has an eror"
src_seq, src_valid_length = encode_char(sentence)
def transform(data, label):
src, src_valid_length = encode_char(data, src=True)
tgt, tgt_valid_length = encode_char(label, src=False)
return src, src_valid_length, tgt, tgt_valid_length, data, label
