mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix unk decoding

Browse Source
This commit is contained in:
Desh Raj 2023-06-15 16:08:41 -04:00
parent 323a2993ea
commit 1d2f96326a
2 changed files with 18 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
egs/tedlium3/ASR/zipformer/decode.py
View File

@ -374,6 +374,7 @@ def decode_one_batch(
encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C)
hyps = [] hyps = []
unk = sp.decode(sp.unk_id()).strip()
if params.decoding_method == "fast_beam_search": if params.decoding_method == "fast_beam_search":
hyp_tokens = fast_beam_search_one_best( hyp_tokens = fast_beam_search_one_best(
@ -386,7 +387,8 @@ def decode_one_batch(
max_states=params.max_states, max_states=params.max_states,
) )
for hyp in sp.decode(hyp_tokens): for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split()) hyp = [w for w in hyp.split() if w != unk]
hyps.append(hyp)
elif params.decoding_method == "fast_beam_search_nbest_LG": elif params.decoding_method == "fast_beam_search_nbest_LG":
hyp_tokens = fast_beam_search_nbest_LG( hyp_tokens = fast_beam_search_nbest_LG(
model=model, model=model,
@ -400,7 +402,8 @@ def decode_one_batch(
nbest_scale=params.nbest_scale, nbest_scale=params.nbest_scale,
) )
for hyp in hyp_tokens: for hyp in hyp_tokens:
hyps.append([word_table[i] for i in hyp]) hyp = [word_table[i] for i in hyp if word_table[i] != unk]
hyps.append(hyp)
elif params.decoding_method == "fast_beam_search_nbest": elif params.decoding_method == "fast_beam_search_nbest":
hyp_tokens = fast_beam_search_nbest( hyp_tokens = fast_beam_search_nbest(
model=model, model=model,
@ -414,7 +417,8 @@ def decode_one_batch(
nbest_scale=params.nbest_scale, nbest_scale=params.nbest_scale,
) )
for hyp in sp.decode(hyp_tokens): for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split()) hyp = [w for w in hyp.split() if w != unk]
hyps.append(hyp)
elif params.decoding_method == "fast_beam_search_nbest_oracle": elif params.decoding_method == "fast_beam_search_nbest_oracle":
hyp_tokens = fast_beam_search_nbest_oracle( hyp_tokens = fast_beam_search_nbest_oracle(
model=model, model=model,
@ -429,7 +433,8 @@ def decode_one_batch(
nbest_scale=params.nbest_scale, nbest_scale=params.nbest_scale,
) )
for hyp in sp.decode(hyp_tokens): for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split()) hyp = [w for w in hyp.split() if w != unk]
hyps.append(hyp)
elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
hyp_tokens = greedy_search_batch( hyp_tokens = greedy_search_batch(
model=model, model=model,
@ -437,7 +442,8 @@ def decode_one_batch(
encoder_out_lens=encoder_out_lens, encoder_out_lens=encoder_out_lens,
) )
for hyp in sp.decode(hyp_tokens): for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split()) hyp = [w for w in hyp.split() if w != unk]
hyps.append(hyp)
elif params.decoding_method == "modified_beam_search": elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search( hyp_tokens = modified_beam_search(
model=model, model=model,
@ -446,7 +452,8 @@ def decode_one_batch(
beam=params.beam_size, beam=params.beam_size,
) )
for hyp in sp.decode(hyp_tokens): for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split()) hyp = [w for w in hyp.split() if w != unk]
hyps.append(hyp)
else: else:
batch_size = encoder_out.size(0) batch_size = encoder_out.size(0)
@ -470,7 +477,8 @@ def decode_one_batch(
raise ValueError( raise ValueError(
f"Unsupported decoding method: {params.decoding_method}" f"Unsupported decoding method: {params.decoding_method}"
) )
hyps.append(sp.decode(hyp).split()) hyp = [w for w in sp.decode(hyp).split() if w != unk]
hyps.append(hyp)
if params.decoding_method == "greedy_search": if params.decoding_method == "greedy_search":
return {"greedy_search": hyps} return {"greedy_search": hyps}

5
egs/tedlium3/ASR/zipformer/train.py
View File

@ -67,6 +67,7 @@ from joiner import Joiner
from lhotse.cut import Cut from lhotse.cut import Cut
from lhotse.dataset.sampling.base import CutSampler from lhotse.dataset.sampling.base import CutSampler
from lhotse.utils import fix_random_seed from lhotse.utils import fix_random_seed
from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
from model import Transducer from model import Transducer
from optim import Eden, ScaledAdam from optim import Eden, ScaledAdam
from scaling import ScheduledFloat from scaling import ScheduledFloat
@ -415,7 +416,7 @@ def get_parser():
parser.add_argument( parser.add_argument(
"--keep-last-k", "--keep-last-k",
type=int, type=int,
default=5, default=1,
help="""Only keep this number of checkpoints on disk. help="""Only keep this number of checkpoints on disk.
For instance, if it is 3, there are only 3 checkpoints For instance, if it is 3, there are only 3 checkpoints
in the exp-dir with filenames `checkpoint-xxx.pt`. in the exp-dir with filenames `checkpoint-xxx.pt`.
@ -751,7 +752,7 @@ def compute_loss(
warm_step = params.warm_step warm_step = params.warm_step
texts = batch["supervisions"]["text"] texts = batch["supervisions"]["text"]
y = sp.encode(texts, out_type=int) y = convert_texts_into_ids(texts, sp)
y = k2.RaggedTensor(y).to(device) y = k2.RaggedTensor(y).to(device)
with torch.set_grad_enabled(is_training): with torch.set_grad_enabled(is_training):