README.md

UnetFixUp

Unet Implementation using residual block and fixup initialization

Please cite this paper if you are using this code into your project as it was developped for it.

@article{griffiths2022outcast,
    title={OutCast: Single Image Relighting with Cast Shadows},
    author={Griffiths, David and Ritschel, Tobias and Philip, Julien},
    journal={Computer Graphics Forum},
    volume={43},
    year={2022},
    organization={Wiley Online Library}
}

Unet using residual blocks and residual chains without any normalization layer. Example of cfg to instanciate the network:

    from omegaconf import DictConfig
    cfg = DictConfig(
        {
            "feat": 32, #Number of features at the highest resolution
            "in_feat": 3, #Number of input channels
            "out_feat": 3, #Number of output channels
            "down_layers": 5, #Number of downsamplings
            "identity_layers": 3, #Number of residual blocks before and after bottleneck. Meaning for a value of 3, we have 6 residual blocks at each level with two convolutions each
            "bottleneck_layers": 6, #Number of residuals blocks for bottleneck
            "skips": True, #Skip connections
            "act_fn": "relu",
            "out_act_fn": "none", #Activation after the final layer, usually none
            "max_feat": 256, #We doubles features when downsampling but cap it to this value
            "script_submodules": True, #Scripting for faster more efficient network
            "dim": 2, #Should be 2 or 3, for a 2D or 3D Unet
        }
    )