default_config.json

{
    "model": "src.training.dcc_tf",
    "model_params":
    {
        "label_len": 41,
        "L": 32,
        "enc_dim": 512,
        "num_enc_layers": 10,
        "dec_dim": 256,
        "num_dec_layers": 1,
        "dec_buf_len": 13,
        "dec_chunk_size": 13,
        "out_buf_len": 4,
        "use_pos_enc": "true"
    },
    "train_data":
    {
        "input_dir": "data/FSDSoundScapes",
        "dset": "train",
        "sr": 44100,
        "resample_rate": null,
	"max_num_targets":3
    },
    "val_data":
    {
        "input_dir": "data/FSDSoundScapes",
        "dset": "val",
        "sr": 44100,
        "resample_rate": null,
	"max_num_targets":3
    },
    "test_data":
    {
        "input_dir": "data/FSDSoundScapes",
        "dset": "test",
        "sr": 44100,
        "resample_rate": null,
	"max_num_targets":3
    },
    "optim":
    {
        "lr": 5e-4,
        "weight_decay": 0.0
    },
    "lr_sched":
    {
        "mode": "max",
        "factor": 0.1,
        "patience": 5,
        "min_lr": 5e-6,
        "threshold": 0.1,
        "threshold_mode": "abs"
    },
    "base_metric": "scale_invariant_signal_noise_ratio",
    "fix_lr_epochs": 50,
    "epochs": 150,
    "batch_size": 16,
    "eval_batch_size": 64,
    "n_workers": 16
}