repositories/k-diffusion/configs/config_mnist.json

{
    "model": {
        "type": "image_v1",
        "input_channels": 1,
        "input_size": [28, 28],
        "patch_size": 1,
        "mapping_out": 256,
        "depths": [2, 4, 4],
        "channels": [128, 128, 256],
        "self_attn_depths": [false, false, true],
        "has_variance": true,
        "dropout_rate": 0.05,
        "augment_wrapper": true,
        "augment_prob": 0.12,
        "sigma_data": 0.6162,
        "sigma_min": 1e-2,
        "sigma_max": 80,
        "sigma_sample_density": {
            "type": "lognormal",
            "mean": -1.2,
            "std": 1.2
        }
    },
    "dataset": {
        "type": "mnist",
        "location": "data"
    },
    "optimizer": {
        "type": "adamw",
        "lr": 2e-4,
        "betas": [0.95, 0.999],
        "eps": 1e-6,
        "weight_decay": 1e-3
    },
    "lr_sched": {
        "type": "inverse",
        "inv_gamma": 20000.0,
        "power": 1.0,
        "warmup": 0.99
    },
    "ema_sched": {
        "type": "inverse",
        "power": 0.6667,
        "max_value": 0.9999
    }
}