From bf4b3ef9cb95acaa2b92b98f236c4f3228cdbc2d Mon Sep 17 00:00:00 2001
From: hnluo <haoneng.lhn@alibaba-inc.com>
Date: 星期四, 21 九月 2023 16:30:43 +0800
Subject: [PATCH] Merge pull request #976 from alibaba-damo-academy/dev_lhn
---
funasr/bin/asr_inference_launch.py | 84 ++++++++++++++++++++++++++++++++---------
1 files changed, 65 insertions(+), 19 deletions(-)
diff --git a/funasr/bin/asr_inference_launch.py b/funasr/bin/asr_inference_launch.py
index ea0f221..50b9886 100644
--- a/funasr/bin/asr_inference_launch.py
+++ b/funasr/bin/asr_inference_launch.py
@@ -427,7 +427,7 @@
else:
text_postprocessed, word_lists = postprocessed_result[0], postprocessed_result[1]
item = {'key': key, 'value': text_postprocessed}
- if timestamp_postprocessed != "" or len(timestamp) == 0:
+ if timestamp_postprocessed != "":
item['timestamp'] = timestamp_postprocessed
asr_result_list.append(item)
finish_count += 1
@@ -652,7 +652,7 @@
batch_size_token_ms_cum = 0
beg_idx = 0
beg_asr_total = time.time()
- for j, _ in tqdm(enumerate(range(0, n))):
+ for j, _ in enumerate(tqdm(range(0, n))):
batch_size_token_ms_cum += (sorted_data[j][0][1] - sorted_data[j][0][0])
if j < n - 1 and (batch_size_token_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size_token_ms and (sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size_token_threshold_s:
continue
@@ -719,7 +719,7 @@
item = {'key': key, 'value': text_postprocessed_punc}
if text_postprocessed != "":
item['text_postprocessed'] = text_postprocessed
- if time_stamp_postprocessed != "" or len(time_stamp) == 0:
+ if time_stamp_postprocessed != "":
item['time_stamp'] = time_stamp_postprocessed
item['sentences'] = time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed)
@@ -842,37 +842,72 @@
data = yaml.load(f, Loader=yaml.Loader)
return data
- def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+ def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], encoder_chunk_look_back=0,
+ decoder_chunk_look_back=0, batch_size=1):
if len(cache) > 0:
return cache
config = _read_yaml(asr_train_config)
enc_output_size = config["encoder_conf"]["output_size"]
feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
- "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+ "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size,
+ "encoder_chunk_look_back": encoder_chunk_look_back, "last_chunk": False, "opt": None,
"feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
cache["encoder"] = cache_en
- cache_de = {"decode_fsmn": None}
+ cache_de = {"decode_fsmn": None, "decoder_chunk_look_back": decoder_chunk_look_back, "opt": None, "chunk_size": chunk_size}
cache["decoder"] = cache_de
return cache
- def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+ def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], encoder_chunk_look_back=0,
+ decoder_chunk_look_back=0, batch_size=1):
if len(cache) > 0:
config = _read_yaml(asr_train_config)
enc_output_size = config["encoder_conf"]["output_size"]
feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
- "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
- "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)),
- "tail_chunk": False}
+ "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size,
+ "encoder_chunk_look_back": encoder_chunk_look_back, "last_chunk": False, "opt": None,
+ "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
cache["encoder"] = cache_en
- cache_de = {"decode_fsmn": None}
+ cache_de = {"decode_fsmn": None, "decoder_chunk_look_back": decoder_chunk_look_back, "opt": None, "chunk_size": chunk_size}
cache["decoder"] = cache_de
return cache
+
+ #def _prepare_cache(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+ # if len(cache) > 0:
+ # return cache
+ # config = _read_yaml(asr_train_config)
+ # enc_output_size = config["encoder_conf"]["output_size"]
+ # feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
+ # cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
+ # "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+ # "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)), "tail_chunk": False}
+ # cache["encoder"] = cache_en
+
+ # cache_de = {"decode_fsmn": None}
+ # cache["decoder"] = cache_de
+
+ # return cache
+
+ #def _cache_reset(cache: dict = {}, chunk_size=[5, 10, 5], batch_size=1):
+ # if len(cache) > 0:
+ # config = _read_yaml(asr_train_config)
+ # enc_output_size = config["encoder_conf"]["output_size"]
+ # feats_dims = config["frontend_conf"]["n_mels"] * config["frontend_conf"]["lfr_m"]
+ # cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, enc_output_size)),
+ # "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
+ # "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], feats_dims)),
+ # "tail_chunk": False}
+ # cache["encoder"] = cache_en
+
+ # cache_de = {"decode_fsmn": None}
+ # cache["decoder"] = cache_de
+
+ # return cache
def _forward(
data_path_and_name_and_type,
@@ -901,24 +936,34 @@
is_final = False
cache = {}
chunk_size = [5, 10, 5]
+ encoder_chunk_look_back = 0
+ decoder_chunk_look_back = 0
if param_dict is not None and "cache" in param_dict:
cache = param_dict["cache"]
if param_dict is not None and "is_final" in param_dict:
is_final = param_dict["is_final"]
if param_dict is not None and "chunk_size" in param_dict:
chunk_size = param_dict["chunk_size"]
+ if param_dict is not None and "encoder_chunk_look_back" in param_dict:
+ encoder_chunk_look_back = param_dict["encoder_chunk_look_back"]
+ if encoder_chunk_look_back > 0:
+ chunk_size[0] = 0
+ if param_dict is not None and "decoder_chunk_look_back" in param_dict:
+ decoder_chunk_look_back = param_dict["decoder_chunk_look_back"]
# 7 .Start for-loop
# FIXME(kamo): The output format should be discussed about
raw_inputs = torch.unsqueeze(raw_inputs, axis=0)
asr_result_list = []
- cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1)
+ cache = _prepare_cache(cache, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back,
+ decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
item = {}
if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "sound":
sample_offset = 0
speech_length = raw_inputs.shape[1]
stride_size = chunk_size[1] * 960
- cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1)
+ cache = _prepare_cache(cache, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back,
+ decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
final_result = ""
for sample_offset in range(0, speech_length, min(stride_size, speech_length - sample_offset)):
if sample_offset + stride_size >= speech_length - 1:
@@ -939,7 +984,8 @@
asr_result_list.append(item)
if is_final:
- cache = _cache_reset(cache, chunk_size=chunk_size, batch_size=1)
+ cache = _cache_reset(cache, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back,
+ decoder_chunk_look_back=decoder_chunk_look_back, batch_size=1)
return asr_result_list
return _forward
@@ -1297,7 +1343,7 @@
quantize_modules: Optional[List[str]] = None,
quantize_dtype: Optional[str] = "float16",
streaming: Optional[bool] = False,
- simu_streaming: Optional[bool] = False,
+ fake_streaming: Optional[bool] = False,
full_utt: Optional[bool] = False,
chunk_size: Optional[int] = 16,
left_context: Optional[int] = 16,
@@ -1374,7 +1420,7 @@
quantize_modules=quantize_modules,
quantize_dtype=quantize_dtype,
streaming=streaming,
- simu_streaming=simu_streaming,
+ fake_streaming=fake_streaming,
full_utt=full_utt,
chunk_size=chunk_size,
left_context=left_context,
@@ -1432,8 +1478,8 @@
final_hyps = speech2text.streaming_decode(
speech[_end: len(speech)], is_final=True
)
- elif speech2text.simu_streaming:
- final_hyps = speech2text.simu_streaming_decode(**batch)
+ elif speech2text.fake_streaming:
+ final_hyps = speech2text.fake_streaming_decode(**batch)
elif speech2text.full_utt:
final_hyps = speech2text.full_utt_decode(**batch)
else:
@@ -1823,7 +1869,7 @@
group.add_argument("--lm_weight", type=float, default=1.0, help="RNNLM weight")
group.add_argument("--ngram_weight", type=float, default=0.9, help="ngram weight")
group.add_argument("--streaming", type=str2bool, default=False)
- group.add_argument("--simu_streaming", type=str2bool, default=False)
+ group.add_argument("--fake_streaming", type=str2bool, default=False)
group.add_argument("--full_utt", type=str2bool, default=False)
group.add_argument("--chunk_size", type=int, default=16)
group.add_argument("--left_context", type=int, default=16)
--
Gitblit v1.9.1