decoding

游雁

2024-06-14 47209220c653e2faa2c6c83801874b9f1df7d7dc

 funasr/models/llm_asr/model.py

@@ -422,7 +422,7 @@
                        layer_id = int(name[beg + 1 : end - 1])
                        if layer_id < freeze_layer_num:
                            param.requires_grad = False
                    elif not name.startswith("audio_encoder.ln_post"):
                    elif "ln_post." not in name:
                        param.requires_grad = False
                else:
                    param.requires_grad = False
@@ -496,11 +496,14 @@

        batch_size, frames, _ = speech.shape

        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        with torch.cuda.amp.autocast(enabled=False):
            # audio encoder
            encoder_out, encoder_out_lens = self.audio_encoder(
                speech.permute(0, 2, 1), speech_lengths
            )

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
            # audio_adaptor
            encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)