kitti_eb5.json

{
    "generic": {
        "seed": 13
    },
    "training": {
        "n_iters": 40000,
        "batch_size": 8,
        "validation_interval": 200,
        "nsteps_accumulation_gradient": 2,
        "lr": 0.0002,
        "wd": 0.05,
        "div_factor": 1,
        "final_div_factor": 10,
        "loss": {
            "name": "SILog",
            "weight": 10.0
        }
    },
    "data": {
        "crop": "garg",
        "train_dataset": "KITTIDataset",
        "val_dataset": "KITTIDataset",
        "data_root": "datasets",
        "benchmark": false,
        "augmentations": {
            "horizontal_flip": 0.5,
            "random_rotation": 2.5,
            "random_scale": 0.05,
            "random_translation": 0.05,
            "random_brightness": 0.1,
            "random_contrast": 0.1,
            "random_saturation": 0.1,
            "random_gamma": 0.2,
            "random_hue": 0.1,
            "random_sharpness": 0.1,
            "random_posterize": 4,
            "random_solarize": 0.2,
            "rotation_p": 1,
            "scale_p": 1,
            "translation_p": 0,
            "brightness_p": 1,
            "contrast_p": 1,
            "saturation_p": 1,
            "gamma_p": 1,
            "hue_p": 1,
            "sharpness_p": 0,
            "posterize_p": 0,
            "solarize_p": 0,
            "equalize_p": 0,
            "autocontrast_p": 0
        }
    },
    "model": {
        "num_heads": 1,
        "attn_dec": true,
        "activation": "gelu",
        "expansion": 4,
        "output_dim": 1,
        "isd": {
            "latent_dim": 128,
            "num_resolutions": 3,
            "depths": 2
        },
        "pixel_decoder": {
            "heads": 4,
            "depths": 6,
            "hidden_dim": 256,
            "anchor_points": 8
        },
        "pixel_encoder": {
            "img_size": [
                1216,
                352
            ],
            "name": "efficientnet_b5",
            "lr_dedicated": 5e-05
        },
        "afp": {
            "context_low_resolutions_skip": 1,
            "depths": 2,
            "latent_dim": 128,
            "num_latents": 32
        }
    }
}