GenSDF Supplementary Information

That is, the length of each side on our cube is 2 instead of 1. We experimented with a larger number for generating shape features (e.g., Model Hyperparameters For Eqs. (3) and (9) (main paper) we the set the loss coefficients to For our meta-learning stage, we define one epoch as sampling one object from one class. This takes up roughly 4.5 GB of GPU memory and we trained to 40,000 epochs in 2 days on an NVIDIA A100 For our semi-supervised stage, one epoch is defined as sampling all objects once; i.e., each epoch Following DeepSDF [4] with a number of changes, see text. "lin" represents a fully-connected layer. "conv(a,b,c)" represents a 2D convolutional layer with kernel "pool(a,b,c)" represents a 2D max Note that as mentioned previously, we train on all 130,000 query points for each unlabeled object.