potsdam_segmentation.json

{
    "name": "potsdam_Deform",
    "phase": "train", // train or test

    "path": { //path of score-based models
        "log": "logs",
        "tb_logger": "tb_logger",
        "results": "results",
        "checkpoint": "checkpoints",
        "resume_state": "pretrain_score/I190000_E97"
    },

    "path_Seg": { //path of segmentation modules
        "log": "logs",
        "tb_logger": "tb_logger",
        "results": "results/potsdam_Deform", // visualization
        "checkpoint": "checkpoints/potsdam_Deform",
        "resume_state": null
    },

    "datasets": {
        "train": {
            "name": "potsdam",
            "dataroot": "../data/potsdam/train_1",
            "img_dir": "images_256",
            "mask_dir": "masks_256",
            "batch_size": 1,
            "num_workers": 8,
            "use_shuffle": true,
            "mosaic_ratio": 0.25
        },
        "val": {
            "name": "potsdam",
            "dataroot": "../data/potsdam/fs_test",
            "img_dir": "images_256",
            "mask_dir": "masks_256",
            "batch_size": 2,
            "num_workers": 8,
            "use_shuffle": true,
            "mosaic_ratio": 0.0
        },
        "test": {
            "name": "potsdam",
            "dataroot": "../data/potsdam/fs_test",
            "img_dir": "images_256",
            "mask_dir": "masks_256",
            "batch_size": 2,
            "num_workers": 8,
            "use_shuffle": false,
            "mosaic_ratio": 0.0
        }
    },

    "model_Seg": {
        "type": "Deform_concat",
        "downsample_times": 0,
        "hidden_dim": 256,
        "feat_scales": [2, 5, 8, 11, 14],
        "out_channels": 6,
        "loss_type": "JointLoss",
        "output_size": 256,
        "feat_type": "dec", //Options: 'enc'/'dec'
        "t": [50, 150, 300]
    },

    "model": {
        "which_model_G": "sr3", // use the ddpm or sr3 network structure
        "finetune_norm": false,
        "unet": {
            "in_channel": 3,
            "out_channel": 3,
            "inner_channel": 128,
            "channel_multiplier": [
                1,
                2,
                4,
                8,
                8
            ],
            "attn_res": [
                16
            ],
            "res_blocks": 2,
            "dropout": 0.2
        },
        "beta_schedule": {
            "train": {
                "schedule": "linear",
                "n_timestep": 2000,
                "linear_start": 1e-6,
                "linear_end": 1e-2
            },
            "val": {
                "schedule": "linear",
                "n_timestep": 2000,
                "linear_start": 1e-6,
                "linear_end": 1e-2
            },
            "test": {
                "schedule": "linear",
                "n_timestep": 2000,
                "linear_start": 1e-6,
                "linear_end": 1e-2
            }
        },
        "diffusion": {
            "image_size": 256,
            "channels": 3, //sample channel
            "loss": "l2", //options "l1" or "l2"
            "conditional": false // unconditional generation or unconditional generation(super_resolution)
        }
    },

    "train": {
        "n_epoch": 200,
        "train_print_freq": 50,
        "val_freq": 5,
        "save_freq": 5,
        "val_print_freq": 50,
        "optimizer": {
            "type": "adamw",
            "lr": 1e-4
        },
        "sheduler": {
            "lr_policy": "linear",  // options: linear, poly, step
            "n_steps": 3,           //this is aplicable only if step sheduler using
            "gamma": 0.1            //Multiplicative factor of learning rate decay. Default: 0.1.
        }
    },
    "wandb": {
        "project": "potsdam_Deform"
    }
}