Compressing multidimensional weather and climate data into neural networks
Authors: Langwen Huang, Torsten Hoefler
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we test our method by varying model structure and data resolution, apply it to the training dataset of a CNN, and compare the resulting test errors. All the experiments are performed on a single compute node with 3 NVIDIA RTX 3090 GPUs. |
| Researcher Affiliation | Academia | Langwen Huang Department of Computer Science ETH Zurich langwen.huang@inf.ethz.ch Torsten Hoefler Department of Computer Science ETH Zurich torsten.hoefler@inf.ethz.ch |
| Pseudocode | No | The paper includes a diagram of the neural network structure, but no structured pseudocode or algorithm blocks were found. |
| Open Source Code | Yes | The source code is available in https://github.com/huanglangwen/NNCompression . |
| Open Datasets | Yes | The data used for experiments are extracted from the ERA5 dataset (Hersbach et al., 2020)... |
| Dataset Splits | No | The paper states that for their compression method, they use all data available in the training process ('overfit the neural network to the data, there is no need for generalization to unseen inputs. Therefore we use all the data available in the training process.'). For the CNN training, it mentions 'test set' evaluation, but no explicit validation split percentage or count is provided. |
| Hardware Specification | Yes | All the experiments are performed on a single compute node with 3 NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions the Adam optimizer, but no specific software dependencies with version numbers (e.g., Python, PyTorch versions) were provided. |
| Experiment Setup | Yes | Throughout the experiments, we fix the number of FCBlocks d to 12, the number of Fourier features m to 128, and the standard deviation of Fourier features σ to 1.6. We use the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 3e-4. The neural networks are trained for 20 epochs where each has 1,046,437,920 randomly sampled coordinates (matching the total number of grid points in dataset 1) that are split into batches of 389,880 (matching the size of 3 horizontal slices in dataset 1) coordinates. |