From 149063ced4d2d5269f0472677228eadfcb4a4d8a Mon Sep 17 00:00:00 2001
From: 维石 <shixian.shi@alibaba-inc.com>
Date: 星期三, 17 四月 2024 14:33:24 +0800
Subject: [PATCH] update seaco finetune

---
 funasr/models/paraformer_streaming/model.py |   24 ++++++++++++++----------
 1 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/funasr/models/paraformer_streaming/model.py b/funasr/models/paraformer_streaming/model.py
index 45875a2..499b487 100644
--- a/funasr/models/paraformer_streaming/model.py
+++ b/funasr/models/paraformer_streaming/model.py
@@ -235,8 +235,7 @@
         decoder_out_1st = None
         pre_loss_att = None
         if self.sampling_ratio > 0.0:
-            if self.step_cur < 2:
-                logging.info("enable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
+
             if self.use_1st_decoder_loss:
                 sematic_embeds, decoder_out_1st, pre_loss_att = \
                     self.sampler_with_grad(encoder_out, encoder_out_lens, ys_pad,
@@ -246,8 +245,6 @@
                     self.sampler(encoder_out, encoder_out_lens, ys_pad,
                                  ys_pad_lens, pre_acoustic_embeds, scama_mask)
         else:
-            if self.step_cur < 2:
-                logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
             sematic_embeds = pre_acoustic_embeds
         
         # 1. Forward decoder
@@ -534,10 +531,14 @@
         for i in range(n):
             kwargs["is_final"] = _is_final and i == n -1
             audio_sample_i = audio_sample[i*chunk_stride_samples:(i+1)*chunk_stride_samples]
-
-            # extract fbank feats
-            speech, speech_lengths = extract_fbank([audio_sample_i], data_type=kwargs.get("data_type", "sound"),
-                                                   frontend=frontend, cache=cache["frontend"], is_final=kwargs["is_final"])
+            if kwargs["is_final"] and len(audio_sample_i) < 960:
+                cache["encoder"]["tail_chunk"] = True
+                speech = cache["encoder"]["feats"]
+                speech_lengths = torch.tensor([speech.shape[1]], dtype=torch.int64).to(speech.device)
+            else:
+                # extract fbank feats
+                speech, speech_lengths = extract_fbank([audio_sample_i], data_type=kwargs.get("data_type", "sound"),
+                                                       frontend=frontend, cache=cache["frontend"], is_final=kwargs["is_final"])
             time3 = time.perf_counter()
             meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
             meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
@@ -563,5 +564,8 @@
             ibest_writer["text"][key[0]] = text_postprocessed
 
         return result, meta_data
-
-
+    
+    def export(self, **kwargs):
+        from .export_meta import export_rebuild_model
+        models = export_rebuild_model(model=self, **kwargs)
+        return models
\ No newline at end of file

--
Gitblit v1.9.1