Feature: add tests for meps dataset

.github/workflows/pre-commit.yml

@@ -1,4 +1,4 @@
-name: lint
+name: Linting
 
 on:
   # trigger on pushes to any branch, but not main

.github/workflows/run_tests.yml

@@ -0,0 +1,40 @@
+name: Unit Tests
+
+on:
+  # trigger on pushes to any branch, but not main
+  push:
+    branches-ignore:
+      - main
+  # and also on PRs to main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
+
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          if [ -f requirements.txt ]; then pip install -r requirements.txt; fi
+          pip install torch-geometric>=2.5.2
+      - name: Test with pytest
+        run: |
+          pytest  -v -s
+      - name: Cache data
+        uses: actions/cache@v4
+        with:
+          path: data
+          key: ${{ runner.os }}-meps-reduced-example-data-v0.1.0
+          restore-keys: |
+            ${{ runner.os }}-meps-reduced-example-data-v0.1.0

CHANGELOG.md

@@ -8,6 +8,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [unreleased](https://github.com/joeloskarsson/neural-lam/compare/v0.1.0...HEAD)
 
 ### Added
+- Added tests for loading dataset, creating graph, and training model based on reduced MEPS dataset stored on AWS S3, along with automatic running of tests on push/PR to GitHub. Added caching of test data tp speed up running tests.
+  [/#38](https://github.com/mllam/neural-lam/pull/38)
+  @SimonKamuk
 
 - Replaced `constants.py` with `data_config.yaml` for data configuration management
   [\#31](https://github.com/joeloskarsson/neural-lam/pull/31)

README.md

@@ -1,3 +1,6 @@
+![Linting](https://github.com/mllam/neural-lam/actions/workflows/pre-commit.yml/badge.svg)
+![Automatic tests](https://github.com/mllam/neural-lam/actions/workflows/run_tests.yml/badge.svg)
+
 <p align="middle">
     <img src="figures/neural_lam_header.png" width="700">
 </p>
@@ -279,6 +282,8 @@ pre-commit run --all-files
 ```
 from the root directory of the repository.
 
+Furthermore, all tests in the ```tests``` directory will be run upon pushing changes by a github action. Failure in any of the tests will also reject the push/PR.
+
 # Contact
 If you are interested in machine learning models for LAM, have questions about our implementation or ideas for extending it, feel free to get in touch.
 You can open a github issue on this page, or (if more suitable) send an email to [joel.oskarsson@liu.se](mailto:joel.oskarsson@liu.se).

create_mesh.py

@@ -153,7 +153,7 @@ def prepend_node_index(graph, new_index):
     return networkx.relabel_nodes(graph, to_mapping, copy=True)
 
 
-def main():
+def main(input_args=None):
     parser = ArgumentParser(description="Graph generation arguments")
     parser.add_argument(
         "--data_config",
@@ -186,7 +186,7 @@ def main():
         default=0,
         help="Generate hierarchical mesh graph (default: 0, no)",
     )
-    args = parser.parse_args()
+    args = parser.parse_args(input_args)
 
     # Load grid positions
     config_loader = config.Config.from_file(args.data_config)

docs/notebooks/create_reduced_meps_dataset.ipynb

@@ -0,0 +1,239 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Creating meps_example_reduced\n",
+    "This notebook outlines how the small-size test dataset ```meps_example_reduced``` was created based on the slightly larger dataset ```meps_example```. The zipped up datasets are 263 MB and 2.6 GB, respectively. See [README.md](../../README.md) for info on how to download ```meps_example```.\n",
+    "\n",
+    "The dataset was reduced in size by reducing the number of grid points and variables.\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Standard library\n",
+    "import os\n",
+    "\n",
+    "# Third-party\n",
+    "import numpy as np\n",
+    "import torch"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "\n",
+    "The number of grid points was reduced to 1/4 by halving the number of coordinates in both the x and y direction. This was done by removing a quarter of the grid points along each outer edge, so the center grid points would stay centered in the new set.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load existing grid\n",
+    "grid_xy = np.load('data/meps_example/static/nwp_xy.npy')\n",
+    "# Get slices in each dimension by cutting off a quarter along each edge\n",
+    "num_x, num_y = grid_xy.shape[1:]\n",
+    "x_slice = slice(num_x//4, 3*num_x//4)\n",
+    "y_slice = slice(num_y//4, 3*num_y//4)\n",
+    "# Index and save reduced grid\n",
+    "grid_xy_reduced = grid_xy[:, x_slice, y_slice]\n",
+    "np.save('data/meps_example_reduced/static/nwp_xy.npy', grid_xy_reduced)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "\n",
+    "This cut out the border, so a new perimeter of 10 grid points was established as border (10 was also the border size in the original \"meps_example\").\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Outer 10 grid points are border\n",
+    "old_border_mask = np.load('data/meps_example/static/border_mask.npy')\n",
+    "assert np.all(old_border_mask[10:-10, 10:-10] == False)\n",
+    "assert np.all(old_border_mask[:10, :] == True)\n",
+    "assert np.all(old_border_mask[:, :10] == True)\n",
+    "assert np.all(old_border_mask[-10:,:] == True)\n",
+    "assert np.all(old_border_mask[:,-10:] == True)\n",
+    "\n",
+    "# Create new array with False everywhere but the outer 10 grid points\n",
+    "border_mask = np.zeros_like(grid_xy_reduced[0,:,:], dtype=bool)\n",
+    "border_mask[:10] = True\n",
+    "border_mask[:,:10] = True\n",
+    "border_mask[-10:] = True\n",
+    "border_mask[:,-10:] = True\n",
+    "np.save('data/meps_example_reduced/static/border_mask.npy', border_mask)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "A few other files also needed to be copied using only the new reduced grid"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load surface_geopotential.npy, index only values from the reduced grid, and save to new file\n",
+    "surface_geopotential = np.load('data/meps_example/static/surface_geopotential.npy')\n",
+    "surface_geopotential_reduced = surface_geopotential[x_slice, y_slice]\n",
+    "np.save('data/meps_example_reduced/static/surface_geopotential.npy', surface_geopotential_reduced)\n",
+    "\n",
+    "# Load pytorch file grid_features.pt\n",
+    "grid_features = torch.load('data/meps_example/static/grid_features.pt')\n",
+    "# Index only values from the reduced grid. \n",
+    "# First reshape from (num_grid_points_total, 4) to (num_grid_points_x, num_grid_points_y, 4), \n",
+    "# then index, then reshape back to new total number of grid points\n",
+    "print(grid_features.shape)\n",
+    "grid_features_new = grid_features.reshape(num_x, num_y, 4)[x_slice,y_slice,:].reshape((-1, 4))\n",
+    "# Save to new file\n",
+    "torch.save(grid_features_new, 'data/meps_example_reduced/static/grid_features.pt')\n",
+    "\n",
+    "# flux_stats.pt is just a vector of length 2, so the grid shape and variable changes does not change this file\n",
+    "torch.save(torch.load('data/meps_example/static/flux_stats.pt'), 'data/meps_example_reduced/static/flux_stats.pt')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "\n",
+    "The number of variables was reduced by truncating the variable list to the first 8."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "num_vars = 8\n",
+    "\n",
+    "# Load parameter_weights.npy, truncate to first 8 variables, and save to new file\n",
+    "parameter_weights = np.load('data/meps_example/static/parameter_weights.npy')\n",
+    "parameter_weights_reduced = parameter_weights[:num_vars]\n",
+    "np.save('data/meps_example_reduced/static/parameter_weights.npy', parameter_weights_reduced)\n",
+    "\n",
+    "# Do the same for following 4 pytorch files\n",
+    "for file in ['diff_mean', 'diff_std', 'parameter_mean', 'parameter_std']:\n",
+    "    old_file = torch.load(f'data/meps_example/static/{file}.pt')\n",
+    "    new_file = old_file[:num_vars]\n",
+    "    torch.save(new_file, f'data/meps_example_reduced/static/{file}.pt')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Lastly the files in each of the directories train, test, and val have to be reduced. The folders all have the same structure with files of the following types:\n",
+    "```\n",
+    "nwp_YYYYMMDDHH_mbrXXX.npy\n",
+    "wtr_YYYYMMDDHH.npy\n",
+    "nwp_toa_downwelling_shortwave_flux_YYYYMMDDHH.npy\n",
+    "```\n",
+    "with ```YYYYMMDDHH``` being some date with hours, and ```XXX``` being some 3-digit integer.\n",
+    "\n",
+    "The first type of file has x and y in dimensions 1 and 2, and variable index in dimension 3. Dimension 0 is unchanged.\n",
+    "The second type has has x and y in dimensions 1 and 2. Dimension 0 is unchanged.\n",
+    "The last type has just x and y as the only 2 dimensions.\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "(65, 268, 238, 18)\n",
+      "(65, 268, 238)\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(np.load('data/meps_example/samples/train/nwp_2022040100_mbr000.npy').shape)\n",
+    "print(np.load('data/meps_example/samples/train/nwp_toa_downwelling_shortwave_flux_2022040112.npy').shape)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "The following loop goes through each file in each sample folder and indexes them according to the dimensions given by the file name."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for sample in ['train', 'test', 'val']:\n",
+    "    files = os.listdir(f'data/meps_example/samples/{sample}')\n",
+    "\n",
+    "    for f in files:\n",
+    "        data = np.load(f'data/meps_example/samples/{sample}/{f}')\n",
+    "        if 'mbr' in f:\n",
+    "            data = data[:,x_slice,y_slice,:num_vars]\n",
+    "        elif 'wtr' in f:\n",
+    "            data = data[x_slice, y_slice]\n",
+    "        else:\n",
+    "            data = data[:,x_slice,y_slice]\n",
+    "        np.save(f'data/meps_example_reduced/samples/{sample}/{f}', data)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Lastly, the file ```data_config.yaml``` is modified manually by truncating the variable units, long and short names, and setting the new grid shape. Also the unit descriptions containing ```^``` was automatically parsed using latex, and to avoid having to install latex in the GitHub CI/CD pipeline, this was changed to ```**```. \n",
+    "\n",
+    "This new config file was placed in ```data/meps_example_reduced```, and that directory was then zipped and placed in a European Weather Cloud S3 bucket."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

neural_lam/utils.py

@@ -1,5 +1,6 @@
 # Standard library
 import os
+import shutil
 
 # Third-party
 import numpy as np
@@ -250,7 +251,11 @@ def fractional_plot_bundle(fraction):
     Get the tueplots bundle, but with figure width as a fraction of
     the page width.
     """
-    bundle = bundles.neurips2023(usetex=True, family="serif")
+    # If latex is not available, some visualizations might not render correctly,
+    # but will at least not raise an error.
+    # Alternatively, use unicode raised numbers.
+    usetex = True if shutil.which("latex") else False
+    bundle = bundles.neurips2023(usetex=usetex, family="serif")
     bundle.update(figsizes.neurips2023())
     original_figsize = bundle["figure.figsize"]
     bundle["figure.figsize"] = (

requirements.txt

@@ -13,3 +13,5 @@ plotly>=5.15.0
 
 # for dev
 pre-commit>=2.15.0
+pytest>=8.1.1
+pooch>=1.8.1