release

processing_minicpmo.py

@@ -102,6 +102,31 @@ class MiniCPMOProcessor(ProcessorMixin):
 
         return MiniCPMOBatchFeature(data={**model_inputs})
 
+    def get_audio_placeholder(self, audio_lens, chunk_input, chunk_length):
+        pool_step = 2
+        feature_lens = math.ceil(audio_lens / self.feature_extractor.hop_length)
+
+        feature_lens = (feature_lens - 1) // 2 + 1
+        output_lens = (feature_lens - pool_step) // pool_step + 1
+
+        if chunk_input:
+            fbank_feat_in_chunk = int(chunk_length * 100)
+            cnn_feat_in_chunk = (fbank_feat_in_chunk - 1) // 2 + 1
+            audio_embeds_in_chunk = (cnn_feat_in_chunk - pool_step) // pool_step + 1
+            num_audio_chunks = (output_lens + audio_embeds_in_chunk - 1) // audio_embeds_in_chunk
+
+            place_holders = ""
+            total_unk_len = 0
+            for _ in range(num_audio_chunks):
+                unk_len = min(audio_embeds_in_chunk, output_lens - total_unk_len)
+                place_holders += self.tokenizer.audio_start + "<unk>" * unk_len + self.tokenizer.audio_end
+                total_unk_len += unk_len
+            audio_placeholder = place_holders
+        else:
+            audio_placeholder = self.tokenizer.audio_start + "<unk>" * output_lens + self.tokenizer.audio_end
+
+        return audio_placeholder
+
     def audio_feature_extract(
         self,
         audios: Union[np.ndarray, List[np.ndarray], List[List[np.ndarray]]],
@@ -111,31 +136,6 @@ class MiniCPMOProcessor(ProcessorMixin):
         chunk_length: Optional[int] = 1,
         **kwargs,
     ):
-        def get_audio_placeholder(audio_lens, chunk_input):
-            pool_step = 2
-            feature_lens = math.ceil(audio_lens / self.feature_extractor.hop_length)
-
-            feature_lens = (feature_lens - 1) // 2 + 1
-            output_lens = (feature_lens - pool_step) // pool_step + 1
-
-            if chunk_input:
-                fbank_feat_in_chunk = int(chunk_length * 100)
-                cnn_feat_in_chunk = (fbank_feat_in_chunk - 1) // 2 + 1
-                audio_embeds_in_chunk = (cnn_feat_in_chunk - pool_step) // pool_step + 1
-                num_audio_chunks = (output_lens + audio_embeds_in_chunk - 1) // audio_embeds_in_chunk
-
-                place_holders = ""
-                total_unk_len = 0
-                for _ in range(num_audio_chunks):
-                    unk_len = min(audio_embeds_in_chunk, output_lens - total_unk_len)
-                    place_holders += self.tokenizer.audio_start + "<unk>" * unk_len + self.tokenizer.audio_end
-                    total_unk_len += unk_len
-                audio_placeholder = place_holders
-            else:
-                audio_placeholder = self.tokenizer.audio_start + "<unk>" * output_lens + self.tokenizer.audio_end
-
-            return audio_placeholder
-
         if isinstance(audios, np.ndarray):
             audios_list = [[audios]]
         elif isinstance(audios[0], np.ndarray):
@@ -156,7 +156,7 @@ class MiniCPMOProcessor(ProcessorMixin):
         # audio placeholder not dependent on audio_parts
         for audios in audios_list:
             if audios:
-                audio_ph_list.append([get_audio_placeholder(len(a), chunk_input) for a in audios])
+                audio_ph_list.append([self.get_audio_placeholder(len(a), chunk_input, chunk_length) for a in audios])
             else:
                 audio_ph_list.append([])