Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Generative Adversarial Neural Operators
Authors: Md Ashiqur Rahman, Manuel A Florez, Anima Anandkumar, Zachary E Ross, Kamyar Azizzadenesheli
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically study GANO in controlled cases where both input and output functions are samples from GRFs and compare its performance to the finite-dimensional counterpart GAN. We empirically study the efficacy of GANO on real-world function data of volcanic activities and show its superior performance over GAN. |
| Researcher Affiliation | Collaboration | Md Ashiqur Rahman EMAIL Department of Computer Science Purdue University Manuel A. Florez EMAIL Seismological Laboratory California Institute of Technology Anima Anandkumar EMAIL Computing + Mathematical Sciences California Institute of Technology Zachary E. Ross EMAIL Seismological Laboratory California Institute of Technology Kamyar Azizzadenesheli EMAIL NVIDIA Corporation |
| Pseudocode | Yes | Algorithm 1 GANO 1: Input: Gradient penalty weight λ, number of discriminator updates per iteration nd, number of generator updates per iteration n G, number of samples per update m. 2: Init: Initialize generator parameters θG, discriminator parameters θd, and optimizers Optd, Opt G 3: for each iteration t = 1, . . . do 4: for τ = 1, . . . , nd do 5: Sample {ai}m i from PA, {ui}m i from PU, and {γi}m i from U[0, 1] 6: Compute loss L := 1m Pm i d(ui) d(G(ai)) + λ( d(u)|u=λG(ai)+(1 λ)ui U 1)2 7: Update θd via a call to Optd(L, θd) 8: end for 9: for τ = 1, . . . , n G do 10: Sample {ai}m i from PA 11: Compute loss L := 1m Pm i d(G(ai)) 12: Update θG via a call to Opt G(L, θG) 13: end for 14: end for |
| Open Source Code | Yes | We release the code to generate the data sets in the first part of the empirical study. For the purpose of bench-marking, we also release the processed volcano dataset, which is ready to be deployed in future studies. We also release the implementation code along with the training procedure. |
| Open Datasets | Yes | We release the code to generate the data sets in the first part of the empirical study. For the purpose of bench-marking, we also release the processed volcano dataset, which is ready to be deployed in future studies. |
| Dataset Splits | No | The paper states, "We train GANO on the entire dataset of 4096 inteferograms." It does not provide specific training, validation, or test splits for reproduction. |
| Hardware Specification | No | Please note that, for this empirical study, we use smaller models in GANO in order to fit the high-resolution data to the present GPU machines. |
| Software Dependencies | No | For training, we use Adam optimizer (Kingma & Ba, 2014) and choose a 2D domain of [0, 1]2 to be the domain where both input and output functions are defined on. |
| Experiment Setup | Yes | The models in GANO consist of eight-layer neural operators following the architecture in (Rahman et al., 2022). The initial lifting dimension, i.e., co-dimension is set to 16 and the number of modes is set to 20. To implement the GAN baseline model, we deploy convolutional neural networks, consisting of ten layers for the generator and half the size discriminator, and use Wasserstein loss for the training. For both models, we kept the number of parameters of the generative models roughly the same (20M). To train GAN models, we use GANO loss with the gradient penalty provided in the prior section. This choice is made to avoid otherwise required parameter turning for GAN loss for any resolution. We use the same grid representation of the input and output functions for the GAN and GANO studies. For training, we use Adam optimizer (Kingma & Ba, 2014) and choose a 2D domain of [0, 1]2 to be the domain where both input and output functions are defined on. ... Gradient penalty weight λ, number of discriminator updates per iteration nd, number of generator updates per iteration n G, number of samples per update m. |