From 81f64f1fe137f997dc64cebb53034cdbc7667a0c Mon Sep 17 00:00:00 2001
From: shixian.shi <shixian.shi@alibaba-inc.com>
Date: 星期五, 03 三月 2023 19:18:50 +0800
Subject: [PATCH] paraformer_onnx and paraformer_bin batch inference

---
 funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py |   37 ++++++++++++++++++++-----------------
 1 files changed, 20 insertions(+), 17 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..850f007 100644
--- a/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
+++ b/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
@@ -24,7 +24,8 @@
     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,
                  ):
 
         if not Path(model_dir).exists():
@@ -43,14 +44,15 @@
         )
         self.ort_infer = OrtInferSession(model_file, device_id)
         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:
@@ -66,17 +68,20 @@
                 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_cif_peak is None:
+                    for pred in preds:
+                        asr_res.append({'preds': pred})
+                else:
+                    for pred, us_cif_peak_ in zip(preds, us_cif_peak):
+                        text, tokens = pred
+                        timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak_, copy.copy(tokens))
+                        if len(self.plot_timestamp_to):
+                            self.plot_wave_timestamp(waveform_list[0], timestamp_total, self.plot_timestamp_to)
+                        asr_res.append({'preds': text, 'timestamp': timestamp})
         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 +101,7 @@
             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 +176,7 @@
 
         # Change integer-ids to tokens
         token = self.converter.ids2tokens(token_int)
-        # token = token[:valid_token_num-1]
+        token = token[:valid_token_num-self.pred_bias]
         texts = sentence_postprocess(token)
-        text = texts[0]
-        # text = self.tokenizer.tokens2text(token)
-        return text, token
+        return texts
 

--
Gitblit v1.9.1