Method __call__

evaluation/evaluate_clotho.py:45–102

r"""Evalute.

(
        self,
        pl_model: pl.LightningModule
    )

        self.audio_dir = 'evaluation/data/clotho'

    def __call__(
        self,
        pl_model: pl.LightningModule
    ) -> Dict:
        r"""Evalute."""

        print(f'Evaluation on Clotho Evaluation with [caption] queries.')
        
        pl_model.eval()
        device = pl_model.device

        sisdrs_list = []
        sdris_list = []

        with torch.no_grad():
            for eval_data in tqdm(self.eval_list):

                idx, caption, _, _, _ = eval_data

                source_path = os.path.join(self.audio_dir, f'segment-{idx}.wav')
                mixture_path = os.path.join(self.audio_dir, f'mixture-{idx}.wav')

                source, fs = librosa.load(source_path, sr=self.sampling_rate, mono=True)
                mixture, fs = librosa.load(mixture_path, sr=self.sampling_rate, mono=True)

                sdr_no_sep = calculate_sdr(ref=source, est=mixture)
                                
                text = [caption]

                conditions = pl_model.query_encoder.get_query_embed(
                    modality='text',
                    text=text,
                    device=device 
                )

                input_dict = {
                    "mixture": torch.Tensor(mixture)[None, None, :].to(device),
                    "condition": conditions,
                }

                sep_segment = pl_model.ss_model(input_dict)["waveform"]
                # sep_segment: (batch_size=1, channels_num=1, segment_samples)

                sep_segment = sep_segment.squeeze(0).squeeze(0).data.cpu().numpy()
                # sep_segment: (segment_samples,)

                sdr = calculate_sdr(ref=source, est=sep_segment)
                sdri = sdr - sdr_no_sep
                sisdr = calculate_sisdr(ref=source, est=sep_segment)


                sisdrs_list.append(sisdr)
                sdris_list.append(sdri)

        mean_sisdr = np.mean(sisdrs_list)
        mean_sdri = np.mean(sdris_list)
        
        return mean_sisdr, mean_sdri