configs/inference.json

{
    "imports": [
        "$import torch",
        "$from datetime import datetime",
        "$from pathlib import Path"
    ],
    "bundle_root": ".",
    "model_dir": "$@bundle_root + '/models'",
    "output_dir": "$@bundle_root + '/output'",
    "create_output_dir": "$Path(@output_dir).mkdir(exist_ok=True)",
    "device": "$torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')",
    "output_postfix": "$datetime.now().strftime('sample_%Y%m%d_%H%M%S')",
    "spatial_dims": 3,
    "image_channels": 1,
    "latent_channels": 8,
    "latent_shape": [
        8,
        36,
        44,
        28
    ],
    "autoencoder_def": {
        "_target_": "monai.networks.nets.autoencoderkl.AutoencoderKL",
        "spatial_dims": "@spatial_dims",
        "in_channels": "@image_channels",
        "out_channels": "@image_channels",
        "latent_channels": "@latent_channels",
        "channels": [
            64,
            128,
            256
        ],
        "num_res_blocks": 2,
        "norm_num_groups": 32,
        "norm_eps": 1e-06,
        "attention_levels": [
            false,
            false,
            false
        ],
        "with_encoder_nonlocal_attn": false,
        "with_decoder_nonlocal_attn": false,
        "include_fc": false
    },
    "network_def": {
        "_target_": "monai.networks.nets.diffusion_model_unet.DiffusionModelUNet",
        "spatial_dims": "@spatial_dims",
        "in_channels": "@latent_channels",
        "out_channels": "@latent_channels",
        "channels": [
            256,
            256,
            512
        ],
        "attention_levels": [
            false,
            true,
            true
        ],
        "num_head_channels": [
            0,
            64,
            64
        ],
        "num_res_blocks": 2,
        "include_fc": false,
        "use_combined_linear": false
    },
    "load_autoencoder_path": "$@bundle_root + '/models/model_autoencoder.pt'",
    "load_autoencoder": "$@autoencoder_def.load_old_state_dict(torch.load(@load_autoencoder_path))",
    "autoencoder": "$@autoencoder_def.to(@device)",
    "load_diffusion_path": "$@model_dir + '/model.pt'",
    "load_diffusion": "$@network_def.load_old_state_dict(torch.load(@load_diffusion_path))",
    "diffusion": "$@network_def.to(@device)",
    "noise_scheduler": {
        "_target_": "monai.networks.schedulers.ddim.DDIMScheduler",
        "_requires_": [
            "@load_diffusion",
            "@load_autoencoder"
        ],
        "num_train_timesteps": 1000,
        "beta_start": 0.0015,
        "beta_end": 0.0195,
        "schedule": "scaled_linear_beta",
        "clip_sample": false
    },
    "noise": "$torch.randn([1]+@latent_shape).to(@device)",
    "set_timesteps": "$@noise_scheduler.set_timesteps(num_inference_steps=50)",
    "inferer": {
        "_target_": "scripts.ldm_sampler.LDMSampler",
        "_requires_": "@set_timesteps"
    },
    "saver": {
        "_target_": "SaveImage",
        "_requires_": "@create_output_dir",
        "output_dir": "@output_dir",
        "output_postfix": "@output_postfix"
    },
    "run": [
        "$@inferer.run(@noise, @autoencoder, @diffusion, @noise_scheduler, @saver)"
    ]
}