Pretrained weights for the method presented in "Amortized Inference for Gaussian Process Hyperparameters of Structured Kernels" (UAI 2023). The model main_state_dict_paper.pth is trained on the base symbols SE,LIN and PER and its two-gram multiplications and main_state_dict_with_matern.pth is trained on SE,LIN,PER and Matern52 and its two-gram multiplications. The first .pth file can be use with the PaperAmortizedStructuredConfig config in the main repo and the second file with the AmortizedStructuredWithMaternConfig config.
These weights correspond to the main repo. Make sure that you have installed git lfs. Clone the repo and then pull in order to download the files
git lfs pull
Training was conducted as outlined in the paper. Both files are checkpoints after the second training phase (after noise-variance fine-tuning), whereas for the first model the second phase consists of 200.000 datasets and for the second model of 40.000 datasets. The other training parameters were identical except we used a batch size of 32 for the second model instead of 128.
Amor-Struct-GP (including these weights) is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.