decoding

游雁

2024-06-13 5079d545956097629ae692c0e3c23f2005008dcd

 funasr/models/llm_asr/model.py

@@ -413,15 +413,16 @@

        if freeze:
            for name, param in audio_encoder.named_parameters():
                idx = re.search(r"\.\d+\.", name)
                if idx is not None:
                    beg, end = idx.regs[0]
                    layer_id = int(name[beg + 1 : end - 1])
                    if isinstance(freeze_layer_num, (list, tuple)):
                if isinstance(freeze_layer_num, (list, tuple)):
                    idx = re.search(r"\.\d+\.", name)
                    if idx is not None:
                        beg, end = idx.regs[0]
                        layer_id = int(name[beg + 1 : end - 1])
                        if layer_id in freeze_layer_num:
                            param.requires_grad = False
                    else:
                        param.requires_grad = False
                else:
                    param.requires_grad = False

            audio_encoder.eval()

        self.audio_encoder = audio_encoder
@@ -525,7 +526,7 @@
                ]

        labels_ids[labels_ids == -1] = -100

        attention_mask[attention_mask < 0] = 0
        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
        )