def decode_streaming(self,
xs,
params,
idx2token,
exclude_eos=False,
task='ys'):
assert task == 'ys'
assert self.input_type == 'speech'
assert self.ctc_weight > 0
assert self.fwd_weight > 0
assert len(xs) == 1 # batch size
# assert params['recog_length_norm']
global_params = copy.deepcopy(params)
global_params['recog_max_len_ratio'] = 1.0
block_sync = params['recog_block_sync']
streaming = Streaming(xs[0], params, self.enc,
params['recog_block_sync_size'])
hyps = None
best_hyp_id_stream = []
is_reset = True # for the first block
stdout = False
self.eval()
with torch.no_grad():
lm = getattr(self, 'lm_fwd', None)
lm_second = getattr(self, 'lm_second', None)
while True:
# Encode input features block by block
x_chunk, is_last_block, lookback, lookahead = streaming.extract_feature(
)
if is_reset:
self.enc.reset_cache()
eout_chunk_dict = self.encode([x_chunk],
'all',
streaming=True,
lookback=lookback,
lookahead=lookahead)
eout_chunk = eout_chunk_dict[task]['xs']
is_reset = False # detect the first boundary in the same block
# CTC-based VAD
if streaming.is_ctc_vad:
if self.ctc_weight_sub1 > 0:
ctc_probs_chunk = self.dec_fwd_sub1.ctc_probs(
eout_chunk_dict['ys_sub1']['xs'])
# TODO: consider subsampling
else:
ctc_probs_chunk = self.dec_fwd.ctc_probs(eout_chunk)
is_reset = streaming.ctc_vad(ctc_probs_chunk,
stdout=stdout)
# Truncate the most right frames
if is_reset and not is_last_block and streaming.bd_offset >= 0:
eout_chunk = eout_chunk[:, :streaming.bd_offset]
streaming.eout_chunks.append(eout_chunk)
# Chunk-synchronous attention decoding
if isinstance(self.dec_fwd, RNNDecoder) and block_sync:
end_hyps, hyps, aws_seg = self.dec_fwd.beam_search_block_sync(
eout_chunk,
params,
idx2token,
lm,
hyps=hyps,
state_carry_over=False)
merged_hyps = sorted(end_hyps + hyps,
key=lambda x: x['score'],
reverse=True)
best_hyp_id_prefix = np.array(merged_hyps[0]['hyp'][1:])
if len(best_hyp_id_prefix
) > 0 and best_hyp_id_prefix[-1] == self.eos:
# reset beam if <eos> is generated from the best hypothesis
best_hyp_id_prefix = best_hyp_id_prefix[:
-1] # exclude <eos>
# Segmentation strategy 2:
# If <eos> is emitted from the decoder (not CTC),
# the current block is segmented.
if not is_reset:
streaming.bd_offset = eout_chunk.size(1) - 1
is_reset = True
if len(best_hyp_id_prefix) > 0:
# print('\rStreaming (T:%d [frame], offset:%d [frame], blank:%d [frame]): %s' %
# (streaming.offset + eout_chunk.size(1) * streaming.factor,
# self.dec_fwd.n_frames * streaming.factor,
# streaming.n_blanks * streaming.factor,
# idx2token(best_hyp_id_prefix)))
print('\r%s' % (idx2token(best_hyp_id_prefix)))
if is_reset:
# Global decoding over the segmented region
if not block_sync:
eout = torch.cat(streaming.eout_chunks, dim=1)
elens = torch.IntTensor([eout.size(1)])
ctc_log_probs = None
if params['recog_ctc_weight'] > 0:
ctc_log_probs = torch.log(
self.dec_fwd.ctc_probs(eout))
nbest_hyps_id_offline = self.dec_fwd.beam_search(
eout,
elens,
global_params,
idx2token,
lm,
lm_second,
ctc_log_probs=ctc_log_probs)[0]
# print('Offline (T:%d [10ms]): %s' %
# (streaming.offset + eout_chunk.size(1) * streaming.factor,
# idx2token(nbest_hyps_id_offline[0][0])))
# pick up the best hyp from ended and active hypotheses
if not block_sync:
if len(nbest_hyps_id_offline[0][0]) > 0:
best_hyp_id_stream.extend(
nbest_hyps_id_offline[0][0])
else:
if len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
# print('Final (T:%d [10ms], offset:%d [10ms]): %s' %
# (streaming.offset + eout_chunk.size(1) * streaming.factor,
# self.dec_fwd.n_frames * streaming.factor,
# idx2token(best_hyp_id_prefix)))
# print('-' * 50)
# for test
# eos_hyp = np.zeros(1, dtype=np.int32)
# eos_hyp[0] = self.eos
# best_hyp_id_stream.extend(eos_hyp)
# reset
streaming.reset(stdout=stdout)
hyps = None
streaming.next_chunk()
# next block will start from the frame next to the boundary
if not is_last_block:
streaming.backoff(x_chunk, self.dec_fwd, stdout=stdout)
if is_last_block:
break
# Global decoding for tail chunks
if not block_sync and len(streaming.eout_chunks) > 0:
eout = torch.cat(streaming.eout_chunks, dim=1)
elens = torch.IntTensor([eout.size(1)])
nbest_hyps_id_offline = self.dec_fwd.beam_search(
eout, elens, global_params, idx2token, lm, lm_second)[0]
# print('MoChA: ' + idx2token(nbest_hyps_id_offline[0][0]))
# print('*' * 50)
if len(nbest_hyps_id_offline[0][0]) > 0:
best_hyp_id_stream.extend(nbest_hyps_id_offline[0][0])
# pick up the best hyp
if not is_reset and block_sync and len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
if len(best_hyp_id_stream) > 0:
return [[np.stack(best_hyp_id_stream, axis=0)]], [None]
else:
return [[[]]], [None]
def decode_streaming(self,
xs,
params,
idx2token,
exclude_eos=False,
task='ys'):
"""Simulate streaming decoding. Both encoding and decoding are performed in the online mode."""
assert task == 'ys'
assert self.input_type == 'speech'
assert self.ctc_weight > 0
assert self.fwd_weight > 0
assert len(xs) == 1 # batch size
# assert params['recog_length_norm']
global_params = copy.deepcopy(params)
global_params['recog_max_len_ratio'] = 1.0
block_sync = params['recog_block_sync']
block_size = params['recog_block_sync_size'] # before subsampling
streaming = Streaming(xs[0], params, self.enc)
factor = self.enc.subsampling_factor
block_size //= factor
hyps = None
best_hyp_id_stream = []
is_reset = True # for the first block
stdout = False
self.eval()
lm = getattr(self, 'lm_fwd', None)
lm_second = getattr(self, 'lm_second', None)
# with torch.no_grad():
while True:
# Encode input features block by block
x_block, is_last_block, cnn_lookback, cnn_lookahead, xlen_block = streaming.extract_feature(
)
if is_reset:
self.enc.reset_cache()
eout_block_dict = self.encode([x_block],
'all',
streaming=True,
cnn_lookback=cnn_lookback,
cnn_lookahead=cnn_lookahead,
xlen_block=xlen_block)
eout_block = eout_block_dict[task]['xs']
is_reset = False # detect the first boundary in the same block
# CTC-based VAD
if streaming.is_ctc_vad:
if self.ctc_weight_sub1 > 0:
ctc_probs_block = self.dec_fwd_sub1.ctc_probs(
eout_block_dict['ys_sub1']['xs'])
# TODO: consider subsampling
else:
ctc_probs_block = self.dec_fwd.ctc_probs(eout_block)
is_reset = streaming.ctc_vad(ctc_probs_block, stdout=stdout)
# Truncate the most right frames
if is_reset and not is_last_block and streaming.bd_offset >= 0:
eout_block = eout_block[:, :streaming.bd_offset]
streaming.cache_eout(eout_block)
# Block-synchronous attention decoding
if isinstance(self.dec_fwd, RNNT):
raise NotImplementedError
elif isinstance(self.dec_fwd, RNNDecoder) and block_sync:
for i in range(math.ceil(eout_block.size(1) / block_size)):
eout_block_i = eout_block[:, i * block_size:(i + 1) *
block_size]
end_hyps, hyps, _ = self.dec_fwd.beam_search_block_sync(
eout_block_i,
params,
idx2token,
hyps,
lm,
state_carry_over=False)
merged_hyps = sorted(end_hyps + hyps,
key=lambda x: x['score'],
reverse=True)
if len(merged_hyps) > 0:
best_hyp_id_prefix = np.array(merged_hyps[0]['hyp'][1:])
if len(best_hyp_id_prefix
) > 0 and best_hyp_id_prefix[-1] == self.eos:
# reset beam if <eos> is generated from the best hypothesis
best_hyp_id_prefix = best_hyp_id_prefix[:
-1] # exclude <eos>
# Segmentation strategy 2:
# If <eos> is emitted from the decoder (not CTC),
# the current block is segmented.
if not is_reset:
streaming._bd_offset = eout_block.size(1) - 1
# TODO: fix later
is_reset = True
if len(best_hyp_id_prefix) > 0:
print(
'\rStreaming (T:%d [10ms], offset:%d [10ms], blank:%d [10ms]): %s'
% (streaming.offset + eout_block.size(1) * factor,
self.dec_fwd.n_frames * factor,
streaming.n_blanks * factor,
idx2token(best_hyp_id_prefix)))
elif isinstance(self.dec_fwd, TransformerDecoder):
best_hyp_id_prefix = []
raise NotImplementedError
if is_reset:
# Global decoding over the segmented region
if not block_sync:
eout = streaming.pop_eouts()
elens = torch.IntTensor([eout.size(1)])
ctc_log_probs = None
if params['recog_ctc_weight'] > 0:
ctc_log_probs = torch.log(self.dec_fwd.ctc_probs(eout))
nbest_hyps_id_offline = self.dec_fwd.beam_search(
eout,
elens,
global_params,
idx2token,
lm,
lm_second,
ctc_log_probs=ctc_log_probs,
exclude_eos=exclude_eos)[0]
# pick up the best hyp from ended and active hypotheses
if block_sync:
if len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
else:
if len(nbest_hyps_id_offline[0][0]) > 0:
best_hyp_id_stream.extend(nbest_hyps_id_offline[0][0])
# reset
streaming.reset(stdout=stdout)
hyps = None
streaming.next_block()
if is_last_block:
break
# next block will start from the frame next to the boundary
streaming.backoff(x_block, self.dec_fwd, stdout=stdout)
# Global decoding for tail blocks
if not block_sync and streaming.n_cache_block > 0:
eout = streaming.pop_eouts()
elens = torch.IntTensor([eout.size(1)])
nbest_hyps_id_offline = self.dec_fwd.beam_search(
eout,
elens,
global_params,
idx2token,
lm,
lm_second,
exclude_eos=exclude_eos)[0]
if len(nbest_hyps_id_offline[0][0]) > 0:
best_hyp_id_stream.extend(nbest_hyps_id_offline[0][0])
# pick up the best hyp
if not is_reset and block_sync and len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
if len(best_hyp_id_stream) > 0:
return [[np.stack(best_hyp_id_stream, axis=0)]], [None]
else:
return [[[]]], [None]
def decode_streaming(self,
xs,
params,
idx2token,
exclude_eos=False,
task='ys'):
"""Simulate streaming encoding+decoding. Both encoding and decoding are performed in the online mode."""
block_size = params.get('recog_block_sync_size') # before subsampling
cache_emb = params.get('recog_cache_embedding')
ctc_weight = params.get('recog_ctc_weight')
assert task == 'ys'
assert self.input_type == 'speech'
assert self.ctc_weight > 0
assert self.fwd_weight > 0 or self.ctc_weight == 1.0
assert len(xs) == 1 # batch size
assert params.get('recog_block_sync')
# assert params.get('recog_length_norm')
streaming = Streaming(xs[0], params, self.enc)
factor = self.enc.subsampling_factor
block_size //= factor
assert block_size >= 1, "block_size is too small."
hyps = None
best_hyp_id_stream = []
is_reset = True # for the first block
stdout = False
self.eval()
helper = BeamSearch(params.get('recog_beam_width'), self.eos,
params.get('recog_ctc_weight'),
params.get('recog_lm_weight'), self.device)
lm = getattr(self, 'lm_fwd', None)
lm_second = getattr(self, 'lm_second', None)
lm = helper.verify_lm_eval_mode(lm, params.get('recog_lm_weight'),
cache_emb)
if lm is not None:
assert isinstance(lm, RNNLM)
lm_second = helper.verify_lm_eval_mode(
lm_second, params.get('recog_lm_second_weight'), cache_emb)
# cache token embeddings
if cache_emb and self.fwd_weight > 0:
self.dec_fwd.cache_embedding(self.device)
while True:
# Encode input features block by block
x_block, is_last_block, cnn_lookback, cnn_lookahead, xlen_block = streaming.extract_feature(
)
if is_reset:
self.enc.reset_cache()
eout_block_dict = self.encode([x_block],
'all',
streaming=True,
cnn_lookback=cnn_lookback,
cnn_lookahead=cnn_lookahead,
xlen_block=xlen_block)
eout_block = eout_block_dict[task]['xs']
is_reset = False # detect the first boundary in the same block
# CTC-based VAD
if streaming.is_ctc_vad:
if self.ctc_weight_sub1 > 0:
ctc_probs_block = self.dec_fwd_sub1.ctc.probs(
eout_block_dict['ys_sub1']['xs'])
# TODO: consider subsampling
else:
ctc_probs_block = self.dec_fwd.ctc.probs(eout_block)
is_reset = streaming.ctc_vad(ctc_probs_block, stdout=stdout)
# Truncate the most right frames
if is_reset and not is_last_block and streaming.bd_offset >= 0:
eout_block = eout_block[:, :streaming.bd_offset]
streaming.cache_eout(eout_block)
# Block-synchronous decoding
if ctc_weight == 1 or self.ctc_weight == 1:
end_hyps, hyps = self.dec_fwd.ctc.beam_search_block_sync(
eout_block, params, helper, idx2token, hyps, lm)
elif isinstance(self.dec_fwd, RNNT):
end_hyps, hyps = self.dec_fwd.beam_search_block_sync(
eout_block, params, helper, idx2token, hyps, lm)
elif isinstance(self.dec_fwd, RNNDecoder):
for i in range(math.ceil(eout_block.size(1) / block_size)):
eout_block_i = eout_block[:, i * block_size:(i + 1) *
block_size]
end_hyps, hyps, _ = self.dec_fwd.beam_search_block_sync(
eout_block_i, params, helper, idx2token, hyps, lm)
elif isinstance(self.dec_fwd, TransformerDecoder):
raise NotImplementedError
else:
raise NotImplementedError(self.dec_fwd)
merged_hyps = sorted(end_hyps + hyps,
key=lambda x: x['score'],
reverse=True)
if len(merged_hyps) > 0:
best_hyp_id_prefix = np.array(merged_hyps[0]['hyp'][1:])
if len(hyps) == 0 or (len(best_hyp_id_prefix) > 0
and best_hyp_id_prefix[-1] == self.eos):
# reset beam if <eos> is generated from the best hypothesis
best_hyp_id_prefix = best_hyp_id_prefix[:
-1] # exclude <eos>
# Segmentation strategy 2:
# If <eos> is emitted from the decoder (not CTC),
# the current block is segmented.
if not is_reset:
streaming._bd_offset = eout_block.size(
1) - 1 # TODO: fix later
is_reset = True
if len(best_hyp_id_prefix) > 0:
n_frames = self.dec_fwd.ctc.n_frames if ctc_weight == 1 or self.ctc_weight == 1 else self.dec_fwd.n_frames
print(
'\rStreaming (T:%d [10ms], offset:%d [10ms], blank:%d [10ms]): %s'
% (streaming.offset + eout_block.size(1) * factor,
n_frames * factor, streaming.n_blanks * factor,
idx2token(best_hyp_id_prefix)))
if is_reset:
# pick up the best hyp from ended and active hypotheses
if len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
# reset
streaming.reset(stdout=stdout)
hyps = None
streaming.next_block()
if is_last_block:
break
# next block will start from the frame next to the boundary
streaming.backoff(x_block, self.dec_fwd, stdout=stdout)
# pick up the best hyp
if not is_reset and len(best_hyp_id_prefix) > 0:
best_hyp_id_stream.extend(best_hyp_id_prefix)
if len(best_hyp_id_stream) > 0:
return [[np.stack(best_hyp_id_stream, axis=0)]], [None]
else:
return [[[]]], [None]