From 129cfcd9f283dea0d64f2e20b77662febc2d802c Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 23 三月 2023 10:01:32 +0800
Subject: [PATCH] cer tool
---
funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py | 52 +++++++++++++++++++++++++++++++---------------------
1 files changed, 31 insertions(+), 21 deletions(-)
diff --git a/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py b/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
index 4a55bdf..5567940 100644
--- a/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
+++ b/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
@@ -24,13 +24,18 @@
def __init__(self, model_dir: Union[str, Path] = None,
batch_size: int = 1,
device_id: Union[str, int] = "-1",
- plot_timestamp: bool = False,
+ plot_timestamp_to: str = "",
+ pred_bias: int = 1,
+ quantize: bool = False,
+ intra_op_num_threads: int = 4,
):
if not Path(model_dir).exists():
raise FileNotFoundError(f'{model_dir} does not exist.')
model_file = os.path.join(model_dir, 'model.onnx')
+ if quantize:
+ model_file = os.path.join(model_dir, 'model_quant.onnx')
config_file = os.path.join(model_dir, 'config.yaml')
cmvn_file = os.path.join(model_dir, 'am.mvn')
config = read_yaml(config_file)
@@ -41,16 +46,17 @@
cmvn_file=cmvn_file,
**config['frontend_conf']
)
- self.ort_infer = OrtInferSession(model_file, device_id)
+ self.ort_infer = OrtInferSession(model_file, device_id, intra_op_num_threads=intra_op_num_threads)
self.batch_size = batch_size
- self.plot = plot_timestamp
+ self.plot_timestamp_to = plot_timestamp_to
+ self.pred_bias = pred_bias
def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
waveform_nums = len(waveform_list)
asr_res = []
for beg_idx in range(0, waveform_nums, self.batch_size):
- res = {}
+
end_idx = min(waveform_nums, beg_idx + self.batch_size)
feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
try:
@@ -58,25 +64,31 @@
am_scores, valid_token_lens = outputs[0], outputs[1]
if len(outputs) == 4:
# for BiCifParaformer Inference
- us_alphas, us_cif_peak = outputs[2], outputs[3]
+ us_alphas, us_peaks = outputs[2], outputs[3]
else:
- us_alphas, us_cif_peak = None, None
+ us_alphas, us_peaks = None, None
except ONNXRuntimeError:
#logging.warning(traceback.format_exc())
logging.warning("input wav is silence or noise")
preds = ['']
else:
- preds, raw_token = self.decode(am_scores, valid_token_lens)[0]
- res['preds'] = preds
- if us_cif_peak is not None:
- timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak, copy.copy(raw_token))
- res['timestamp'] = timestamp
- if self.plot:
- self.plot_wave_timestamp(waveform_list[0], timestamp_total)
- asr_res.append(res)
+ preds = self.decode(am_scores, valid_token_lens)
+ if us_peaks is None:
+ for pred in preds:
+ pred = sentence_postprocess(pred)
+ asr_res.append({'preds': pred})
+ else:
+ for pred, us_peaks_ in zip(preds, us_peaks):
+ raw_tokens = pred
+ timestamp, timestamp_raw = time_stamp_lfr6_onnx(us_peaks_, copy.copy(raw_tokens))
+ text_proc, timestamp_proc, _ = sentence_postprocess(raw_tokens, timestamp_raw)
+ # logging.warning(timestamp)
+ if len(self.plot_timestamp_to):
+ self.plot_wave_timestamp(waveform_list[0], timestamp, self.plot_timestamp_to)
+ asr_res.append({'preds': text_proc, 'timestamp': timestamp_proc, "raw_tokens": raw_tokens})
return asr_res
- def plot_wave_timestamp(self, wav, text_timestamp):
+ def plot_wave_timestamp(self, wav, text_timestamp, dest):
# TODO: Plot the wav and timestamp results with matplotlib
import matplotlib
matplotlib.use('Agg')
@@ -96,7 +108,7 @@
x_adj = 0.045 if char != '<sil>' else 0.12
ax1.text((start + end) * 0.5 - x_adj, 0, char)
# plt.legend()
- plotname = "funasr/runtime/python/onnxruntime/debug.png"
+ plotname = "{}/timestamp.png".format(dest)
plt.savefig(plotname, bbox_inches='tight')
def load_data(self,
@@ -171,9 +183,7 @@
# Change integer-ids to tokens
token = self.converter.ids2tokens(token_int)
- # token = token[:valid_token_num-1]
- texts = sentence_postprocess(token)
- text = texts[0]
- # text = self.tokenizer.tokens2text(token)
- return text, token
+ token = token[:valid_token_num-self.pred_bias]
+ # texts = sentence_postprocess(token)
+ return token
--
Gitblit v1.9.1