Pretrained models available in solaris

solaris provides access to a number of pre-trained models from the SpaceNet challenges. See the table below for a summary. Note that the model name in the first column should be used as the "model_name" argument in the config file if you wish to use that model with solaris. Note that we re-trained the

competitors’ models for compatibility with solaris and the training parameters, inputs, and performance may vary slightly from their original models.

Model details

Model name	Model type	Model details	# Parameters	Input shape	Output shape	Config file	Weights file
xdxd_spacenet4	Segmentation UNet	Encoder: VGG16	29.3M	3x512x512	1x512x512	link	link (117 MB)
selimsef_spacenet4_resnet34unet	Segmentation UNet	Encoder: ResNet-34	30.0M	4x416x416	3x416x416	link	link (120 MB)
selimsef_spacenet4_densenet121unet	Segmentation UNet	Encoder: DenseNet-121	15.6M	3x384x384	3x384x384	link	link (63 MB)
selimsef_spacenet4_densenet161unet	Segmentation UNet	Encoder: DenseNet-161	41.1M	3x384x384	3x384x384	link	link (158 MB)

Training details

Below is a summary of the training hyperparameters for each model. For image pre-processing and augmentation pipelines see the config files linked above. Note that our hyperparameters may differ from the competitors’ original values. See their solution descriptions for more on their implementations.

Model name	Loss function	Optimizer	Learning Rate	Training input	Training mask	Batch size	Training Epochs	Pre-trained weights
xdxd_spacenet4	BCE + Jaccard (4:1)	Adam default params	1e-4 with decay	SpaceNet 4 Pan-sharpened RGB	Footprints only	12	60	None
selimsef_spacenet4_resnet34unet	Focal + Dice (1:1)	AdamW 1e-3 weight decay	2e-4 with decay	SpaceNet 4 Pan-sharpened RGB+NIR	3-channel (FP, (edge, contact)	42	70	ImageNet (encoder only)
selimsef_spacenet4_densenet121unet	Focal + Dice (1:1)	AdamW 1e-3 weight decay	2e-4 with decay	SpaceNet 4 Pan-sharpened RGB	3-channel (FP, (edge, contact)	32	70	ImageNet (encoder only)
selimsef_spacenet4_densenet161unet	Focal + Dice (1:1)	AdamW 1e-3 weight decay	2e-4 with decay	SpaceNet 4 Pan-sharpened RGB	3-channel (FP, (edge, contact)	20	60	ImageNet (encoder only)