提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 funasr/models/emotion2vec/model.py

@@ -38,6 +38,7 @@
    emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
    https://arxiv.org/abs/2312.15185
    """

    def __init__(self, **kwargs):
        super().__init__()
        # import pdb; pdb.set_trace()
@@ -75,7 +76,7 @@
            cfg.get("layer_norm_first"),
            self.alibi_biases,
        )
        self.modality_encoders['AUDIO'] = enc
        self.modality_encoders["AUDIO"] = enc

        self.ema = None

@@ -85,7 +86,9 @@

        self.dropout_input = torch.nn.Dropout(cfg.get("dropout_input"))

        dpr = np.linspace(cfg.get("start_drop_path_rate"), cfg.get("end_drop_path_rate"), cfg.get("depth"))
        dpr = np.linspace(
            cfg.get("start_drop_path_rate"), cfg.get("end_drop_path_rate"), cfg.get("depth")
        )

        self.blocks = torch.nn.ModuleList([make_block(dpr[i]) for i in range(cfg.get("depth"))])

@@ -97,7 +100,6 @@
        self.proj = None
        if vocab_size > 0:
            self.proj = torch.nn.Linear(cfg.get("embed_dim"), vocab_size)


    def forward(
        self,
@@ -114,7 +116,7 @@
        **kwargs,
    ):

        feature_extractor = self.modality_encoders['AUDIO']
        feature_extractor = self.modality_encoders["AUDIO"]

        mask_seeds = None

@@ -146,11 +148,7 @@
            ):
                ab = masked_alibi_bias
                if ab is not None and alibi_scale is not None:
                    scale = (
                        alibi_scale[i]
                        if alibi_scale.size(0) > 1
                        else alibi_scale.squeeze(0)
                    )
                    scale = alibi_scale[i] if alibi_scale.size(0) > 1 else alibi_scale.squeeze(0)
                    ab = ab * scale.type_as(ab)

                x, lr = blk(
@@ -192,15 +190,16 @@
        )
        return res

    def inference(self,
                 data_in,
                 data_lengths=None,
                 key: list = None,
                 tokenizer=None,
                 frontend=None,
                 **kwargs,
                 ):
    
    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):

        # if source_file.endswith('.wav'):
        #     wav, sr = sf.read(source_file)
        #     channel = sf.info(source_file).channels
@@ -213,12 +212,17 @@
        meta_data = {}
        # extract fbank feats
        time1 = time.perf_counter()
        audio_sample_list = load_audio_text_image_video(data_in, fs=16000, audio_fs=kwargs.get("fs", 16000),
                                                        data_type=kwargs.get("data_type", "sound"), tokenizer=tokenizer)
        audio_sample_list = load_audio_text_image_video(
            data_in,
            fs=16000,
            audio_fs=kwargs.get("fs", 16000),
            data_type=kwargs.get("data_type", "sound"),
            tokenizer=tokenizer,
        )
        time2 = time.perf_counter()
        meta_data["load_data"] = f"{time2 - time1:0.3f}"
        meta_data["batch_data_time"] = len(audio_sample_list[0])/kwargs.get("fs", 16000)
        
        meta_data["batch_data_time"] = len(audio_sample_list[0]) / kwargs.get("fs", 16000)

        results = []
        output_dir = kwargs.get("output_dir")
        if output_dir:
@@ -230,13 +234,13 @@
            source = source.view(1, -1)

            feats = self.extract_features(source, padding_mask=None)
            x = feats['x']
            feats = feats['x'].squeeze(0).cpu().numpy()
            if granularity == 'frame':
            x = feats["x"]
            feats = feats["x"].squeeze(0).cpu().numpy()
            if granularity == "frame":
                feats = feats
            elif granularity == 'utterance':
            elif granularity == "utterance":
                feats = np.mean(feats, axis=0)
                

            if output_dir and extract_embedding:
                np.save(os.path.join(output_dir, "{}.npy".format(key[i])), feats)

@@ -245,13 +249,25 @@
            if self.proj:
                x = x.mean(dim=1)
                x = self.proj(x)
                for idx, lab in enumerate(labels):
                    x[:,idx] = -np.inf if lab.startswith("unuse") else x[:,idx]
                x = torch.softmax(x, dim=-1)
                scores = x[0].tolist()

            result_i = {"key": key[i], "labels": labels, "scores": scores}
            select_label = [lb for lb in labels if not lb.startswith("unuse")]
            select_score = [scores[idx] for idx, lb in enumerate(labels) if not lb.startswith("unuse")]

            # result_i = {"key": key[i], "labels": labels, "scores": scores}
            result_i = {"key": key[i], "labels": select_label, "scores": select_score}

            if extract_embedding:
                result_i["feats"] = feats
            results.append(result_i)

            
        return results, meta_data
        return results, meta_data

    def export(self, **kwargs):
        from .export_meta import export_rebuild_model

        models = export_rebuild_model(model=self, **kwargs)
        return models