Function _process_chunk

infer_flowsep.py:143–179

(chunk_wav)

        input_len = full_wav.shape[0]

        def _process_chunk(chunk_wav):
            chunk_wav = preprocessor.preprocess_chunk(chunk_wav)
            if len(chunk_wav) < FLOWSEP_CHUNK_IN:
                pad = np.zeros(FLOWSEP_CHUNK_IN - len(chunk_wav), dtype=np.float32)
                chunk_wav = np.concatenate([chunk_wav, pad])
            chunk_wav = chunk_wav[:FLOWSEP_CHUNK_IN]
            mixed_mel, _ = preprocessor.wav_feature_extraction(chunk_wav.reshape(1, -1))
            batch = {
                "fname": [str(audio_file)],
                "text": [args.text],
                "caption": [args.text],
                "waveform": torch.rand(1, 1, FLOWSEP_CHUNK_IN).to(device),
                "log_mel_spec": torch.rand(1, 1024, 64).to(device),
                "sampling_rate": torch.tensor([FLOWSEP_SR]).to(device),
                "label_vector": torch.rand(1, 527).to(device),
                "stft": torch.rand(1, 1024, 512).to(device),
                "mixed_waveform": torch.from_numpy(
                    chunk_wav.reshape(1, 1, FLOWSEP_CHUNK_IN)
                ).to(device),
                "mixed_mel": mixed_mel.reshape(
                    1, mixed_mel.shape[0], mixed_mel.shape[1]
                ).to(device),
            }
            result = latent_diffusion.generate_sample(
                [batch],
                name="temp_result",
                unconditional_guidance_scale=1.0,
                ddim_steps=20,
                n_gen=1,
                save=False,
                save_mixed=False,
            )
            if isinstance(result, np.ndarray):
                out = result.squeeze()
            else:
                out = result.squeeze().cpu().numpy()
            return out[:FLOWSEP_CHUNK_OUT]

        if input_len <= FLOWSEP_CHUNK_IN:
            sep_audio = _process_chunk(full_wav.copy())