modelMusicSep.yaml

project: "[Project] SepReformer" ### Dont't change
notes: "SepReformer final version" ### Insert schanges(plz write details !!!)
# ------------------------------------------------------------------------------------------------------------------------------ #
config:
    # ------------------------------------------------------------ #
    dataset:
        max_len: 96000
        sampling_rate: 24000
        type: "ja_capella_power"
        train: "train"
        val: "test"
        test: "test"
    # ------------------------------------------------------------ #
    dataloader:
        batch_size: 2
        pin_memory: false
        num_workers: 4
        drop_last: false
        shuffle: true
        prefetch_factor: 10
    # ------------------------------------------------------------ #
    model:
        num_stages: &var_model_num_stages 3 # R
        num_spks: &var_model_num_spks 7
        module_audio_enc:
            in_channels: 1
            out_channels: &var_model_audio_enc_out_channels 256
            kernel_size: &var_model_audio_enc_kernel_size 32 # L
            stride: &var_model_audio_enc_stride 8 # S
            groups: 1
            bias: false
        module_feature_projector:
            num_channels: *var_model_audio_enc_out_channels
            in_channels: *var_model_audio_enc_out_channels
            out_channels: &feature_projector_out_channels 128 # F
            kernel_size: 1
            bias: false
        module_separator:
            num_stages: *var_model_num_stages
            relative_positional_encoding:
                in_channels: *feature_projector_out_channels
                num_heads: 8
                maxlen: 2000
                embed_v: false
            enc_stage:
                num_patterns: 2
                global_blocks:
                    in_channels: *feature_projector_out_channels
                    num_mha_heads: 8
                    dropout_rate: 0.05
                local_blocks:
                    in_channels: *feature_projector_out_channels
                    kernel_size: 65
                    dropout_rate: 0.05
                down_conv_layer:
                    in_channels: *feature_projector_out_channels
                    samp_kernel_size: &var_model_samp_kernel_size 5
            spk_split_stage:
                in_channels: *feature_projector_out_channels
                num_spks: *var_model_num_spks
            simple_fusion:
                out_channels: *feature_projector_out_channels
            dec_stage:
                num_spks: *var_model_num_spks
                num_patterns: 3
                global_blocks:
                    in_channels: *feature_projector_out_channels
                    num_mha_heads: 8
                    dropout_rate: 0.05
                local_blocks:
                    in_channels: *feature_projector_out_channels
                    kernel_size: 65
                    dropout_rate: 0.05
                spk_attention:
                    in_channels: *feature_projector_out_channels
                    num_mha_heads: 8
                    dropout_rate: 0.05
        module_output_layer:
            in_channels: *var_model_audio_enc_out_channels
            out_channels: *feature_projector_out_channels
            num_spks: *var_model_num_spks

        module_audio_dec:
            in_channels: *var_model_audio_enc_out_channels
            out_channels: 1
            kernel_size: *var_model_audio_enc_kernel_size
            stride: *var_model_audio_enc_stride
            bias: false
    losses:
        PIT_SPECTRAL:
            lambda: 0.3
            weights: [1, 1, 1]
            window_lengths: [256, 512, 1024]
            hop_lengths: [64, 128, 256]
        PIT_MEL:
            # overall loss weight
            lambda: 0.7

            # equal weight for each of the 7 scales
            weights: [1, 1, 1, 1, 1, 1, 1]

            # number of mel bins at each scale (very coarse → very fine)
            mels: [5, 10, 20, 40, 80, 160, 320]

            # analysis window lengths (in samples)
            window_lengths: [32, 64, 128, 256, 512, 1024, 2048]

            # hop lengths (here 25% overlap)
            hop_lengths: [8, 16, 32, 64, 128, 256, 512]

            # all start from 0 Hz
            mel_fmin: [0, 0, 0, 0, 0, 0, 0]

            # all go up to Nyquist (16 kHz / 2 = 8 kHz)
            mel_fmax: [8000, 8000, 8000, 8000, 8000, 8000, 8000]
        PIT_L1:
            lambda: 1
        loss_g:
            lambda: 1
        loss_f: 
            lambda: 1
    # ------------------------------------------------------------ #
    discriminator:
        rates: [2, 3, 5, 7, 11]
        periods: [2, 3, 5, 7, 11]
        fft_sizes: [2048, 1024, 512]
    # ------------------------------------------------------------ #
    engine:
        ckpt_path_model: ""
        max_epochs: 200
        accum_steps: 10
        gpuid: "0" ### "0"(single-gpu) or "0, 1" (multi-gpu)
        gpu_ids: [0,1,2,3]
        mvn: false
        clip_norm: 5
        start_scheduling: 50
        test_epochs: [100, 120, 150, 170]
        learning_rate: 5e-4
        learning_rate_disc: 2e-4
        weight_decay: 1.0e-2
        log_interval: 20
        ckpt_interval: 1000
        seed: 42
        losses:
            PIT_MEL: 0.7
            PIT_STFT: 0.3
            PIT_L1: 1
            loss_g: 1
            loss_f: 1
            PIT_SDR: 1
    evaluation_metrics:
        PESQ:
            mode: "nb"
            sr: 8000