decoding

游雁

2024-06-14 59bc02b089f7a626fe67907dcfc695eae6883f82

 funasr/models/llm_asr/model.py

@@ -410,19 +410,19 @@
            audio_encoder_output_size = audio_encoder.output_size()
        freeze = audio_encoder_conf.get("freeze", True)
        freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))
        if freeze_layer_num > 0:
            freeze_layer_num = range(freeze_layer_num)
        # if freeze_layer_num > 0:
        #     freeze_layer_num = range(freeze_layer_num)

        if freeze:
            for name, param in audio_encoder.named_parameters():
                if isinstance(freeze_layer_num, (list, tuple)):
                if freeze_layer_num > 0:
                    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:
                        if layer_id < freeze_layer_num:
                            param.requires_grad = False
                    else:
                    elif "ln_post." not in name:
                        param.requires_grad = False
                else:
                    param.requires_grad = False
@@ -449,9 +449,9 @@
            for name, param in model.named_parameters():
                param.requires_grad = False
            model.eval()
        self.llm = model
        llm_dim = model.get_input_embeddings().weight.shape[-1]
        self.llm_dtype = llm_conf.get("llm_dtype", "fp32")
        self.llm = model.to(dtype_map[self.llm_dtype])
        llm_dim = model.get_input_embeddings().weight.shape[-1]

        # adaptor
        adaptor_class = tables.adaptor_classes.get(audio_adaptor)
@@ -496,11 +496,12 @@

        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.encode(speech, 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)
@@ -536,7 +537,9 @@
            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
                inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
                attention_mask=attention_mask,
                labels=labels_ids,
            )
            loss = model_outputs.loss

@@ -560,6 +563,12 @@
            batch_size = int((labels_ids > 0 + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)

        return encoder_out, encoder_out_lens

    def data_template(self, data):
        system, user, assistant = [], [], []
@@ -716,7 +725,8 @@
            speech = speech.to(torch.float16)
        elif kwargs.get("bf16", False):
            speech = speech.to(torch.bfloat16)
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
@@ -801,3 +811,21 @@
            ibest_writer["text_tn"][key[0]] = response_clean

        return results, meta_data


@tables.register("model_classes", "LLMASR3")
class LLMASR3(nn.Module):
    """ """

    def __init__(
        self,
        *args,
        **kwargs,
    ):

        super().__init__(*args, **kwargs)

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech, speech_lengths)
        return encoder_out, encoder_out_lens