revised

.gitignore

@@ -0,0 +1,14 @@
+.dev/
+.snakemake/
+.vscode/
+logs/
+results/data_partition/
+results/preprocess/
+results/train/
+results/tune
+results/figures
+results/experiments
+slurm/
+local/
+
+**/__pycache__

LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2022, Francesco Zanetta
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

README.md

@@ -0,0 +1,18 @@
+# Physics-constrained deep learning postprocessing of temperature and humidity
+
+Workflow for the physics-constrained deep learning postprocessing of temperature and humidity (paper under review). This work investigates the effect of enforcing dependencies between variables by constraining the optimization of neural networks with thermodynamic state equations. 
+
+Pre-print: [https://arxiv.org/abs/2212.04487](https://arxiv.org/abs/2212.04487)
+
+__Installation:__
+If using conda, simply replace mamba with conda. We reccommend you set the two environment variables that indicate where the workflow environments and data will be located. By default, these will be located in a `.snakemake/conda` and `data/` respectively.
+```
+mamba env create -f environment.yaml
+mamba env config vars set SNAKEMAKE_CONDA_PREFIX=<path> SNAKEMAKE_DATA_DIR=<path>
+```
+
+Visualize the workflow:
+```
+snakemake all_results --dag | dot -Tpdf > dag.pdf
+snakemake all_results --rulegraph | dot -Tpdf > rulegraph.pdf
+```

Snakefile

@@ -0,0 +1,84 @@
+import pandas as pd
+import os
+from functools import partial 
+
+configfile: "config/config.yaml"
+configfile: "config/experiments.yaml"
+configfile: "config/extras.yaml"
+configfile: "config/tuning.yaml"
+
+APPROACHES = [
+    "unconstrained",
+    "loss_constrained",
+    "architecture_constrained",
+    "offline_constrained"
+]
+
+TASKS = [
+    "air_temperature",
+    "dew_point_temperature",
+    "surface_air_pressure",
+    "relative_humidity",
+    "water_vapor_mixing_ratio",
+]
+
+N_SPLITS = config["data_partitioning"]["forecast_reference_time"]["n_splits"]
+DATA_DIR = Path(os.getenv("SNAKEMAKE_DATA_DIR", "data/"))
+
+
+include: "rules/common.smk"
+include: "rules/eda.smk"
+include: "rules/main.smk"
+include: "rules/extras.smk"
+
+# rule all_eda:
+#     "results/eda/stations/", "results/eda/model"
+
+rule all_results:
+    input:
+        expand(
+            "results/experiments/{experiment}/{partition}/{focus}",
+            partition=["test"], 
+            experiment=["default","time_generalization"],
+            focus=["performance","physical_consistency"]
+        ),
+        expand(
+            "results/experiments/{experiment}/{partition}/analysis",
+            partition=["test"],
+            experiment=["loss_alpha", "data_efficiency", "time_generalization"]
+        ),
+        expand(
+            "results/experiments/{experiment}/{partition}/physical_consistency",
+            partition=["test"], 
+            experiment=["data_reduction_consistency"]
+        )
+
+
+rule all_performance:
+    input:
+        expand(
+            "results/experiments/{experiment}/{partition}/performance",
+            partition=["test", "train"], 
+            experiment=["default","time_generalization"]
+        )
+
+
+rule all_physical_consistency:
+    input:
+        expand(
+            "results/experiments/{experiment}/{partition}/physical_consistency",
+            partition=["test"], 
+            experiment=["default","time_generalization","data_reduction_consistency"]
+        )
+
+
+rule all_analysis:
+    input:
+        expand(
+            "results/experiments/{experiment}/{partition}/analysis",
+            partition=["test"],
+            experiment=["loss_alpha", "data_efficiency", "time_generalization"]
+        )
+
+
+

config/config.yaml

@@ -0,0 +1,78 @@
+logging:
+  level: DEBUG
+  format: "%(asctime)s %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s"
+  datefmt: "%Y-%m-%d %H:%M:%S"
+
+features:
+  - coe:air_temperature_ensavg
+  - coe:dew_point_temperature_ensavg
+  - coe:dew_point_depression_ensavg
+  - coe:surface_air_pressure_ensavg
+  - coe:relative_humidity_ensavg
+  - coe:water_vapor_mixing_ratio_ensavg
+  - coe:leadtime
+  - time:cos_hourofday
+  - time:sin_hourofday
+  - time:cos_dayofyear
+  - time:sin_dayofyear
+
+targets:
+  - obs:air_temperature
+  - obs:dew_point_temperature
+  - obs:surface_air_pressure
+  - obs:relative_humidity
+  - obs:water_vapor_mixing_ratio
+
+random_seeds: [100, 200, 300]
+
+devices: 1
+
+data_partitioning:
+  forecast_reference_time:
+    p: [0.6, 0.2, 0.2]
+    n_splits: 4
+
+# data
+data.reduction: # 0.1
+data.train_season: # ["SON","DJF","MAM"]
+data.test_season: # JJA
+
+# net
+net.l1: 256
+net.l2: 256
+net.embedding_size: 6
+net.constraint: False
+net.out_size: 5
+
+# loss
+loss.alpha: 0.0
+loss.mask: [True, True, True, True, True]
+loss.trainable: True 
+# loss.log_var_init: [0., 0., 0., 0., 0.]
+# loss.log_var_init: [1.5, 1.7, 0.5, 4.5, -0.5]
+
+# fit
+# fit.lr_patience: 4
+fit.lr: 0.0007
+fit.batch_size: 512
+fit.max_epochs: 25
+fit.patience: 5
+
+# approach-specific parameters
+unconstrained:
+  {}
+  # fit.lr: 0.005
+  # fit.batch_size: 1024
+architecture_constrained:
+  net.constraint: True
+  # fit.lr: 0.005
+  # fit.batch_size: 1024
+loss_constrained:
+  # fit.lr: 0.005
+  # fit.batch_size: 1024
+  loss.alpha: 0.995
+offline_constrained:
+  net.constraint: True
+  loss.mask: [True, True, True, False, False]
+  # fit.lr: 0.005
+  # fit.batch_size: 1024

config/experiments.yaml

@@ -0,0 +1,84 @@
+experiments:
+  # generic experiments (do not have their own notebook)
+  default:
+    approaches:
+      - unconstrained
+      - architecture_constrained
+      - loss_constrained
+      - offline_constrained
+    splits: [0, 1, 2, 3]
+    seeds: [100, 200, 300]
+
+  no_loss_weights:
+    approaches:
+      - unconstrained
+      - architecture_constrained
+      - loss_constrained
+      - offline_constrained
+    param_override:
+      loss.trainable: False
+    splits: [0, 1, 2, 3]
+    seeds: [100, 200, 300]
+
+
+  data_efficiency:
+    approaches:
+      - unconstrained
+      - architecture_constrained
+      - loss_constrained
+      - offline_constrained
+    param_runs:
+      - { data.reduction: 1.0,    net.l1:  256, net.l2:  256, net.embedding_size: 6, fit.lr:  0.007,  fit.batch_size:  512}
+      - { data.reduction: 0.2,    net.l1:  64, net.l2:  64, net.embedding_size: 3, fit.lr:  0.004,  fit.batch_size:  64}
+      - { data.reduction: 0.05,   net.l1:  64, net.l2:  32, net.embedding_size: 3, fit.lr:  0.0035, fit.batch_size:  64}
+      - { data.reduction: 0.01,  net.l1:  32, net.l2:  32, net.embedding_size: 3, fit.lr:  0.003,  fit.batch_size:  64}
+    param_override:
+      fit.max_epochs: 50
+      fit.patience: 5
+    splits: [0, 1, 2, 3]
+    seeds: [100, 200, 300]
+
+  time_generalization:
+    approaches:
+      - unconstrained
+      - architecture_constrained
+      - loss_constrained
+      - offline_constrained
+    param_override:
+      data.train_season: ["SON", "DJF", "MAM"]
+      data.test_season: ["JJA"]
+      net.l1: 64
+      net.l2: 64
+      net.embedding_size: 5
+      fit.lr: 0.002
+      fit.batch_size: 256
+      fit.max_epochs: 15
+      fit.patience: 5
+    splits: [0, 1, 2, 3]
+    seeds: [100, 200, 300]
+
+  # parameter-specific experiments (have their own notebooks)
+  loss_alpha:
+    approaches:
+      - loss_constrained
+    param_grid:
+      loss.alpha: [0., 0.2, 0.5, 0.8, 0.9, 0.95, 0.99, 0.995, 0.999, 0.9999, 0.99999]
+    splits: [0]
+    seeds: [10]
+
+  data_reduction_consistency:
+    approaches:
+      - unconstrained
+      - architecture_constrained
+      - loss_constrained
+    param_override:
+      data.reduction: 0.001
+      net.l1:  32
+      net.l2:  32
+      net.embedding_size: 3
+      fit.lr:  0.001
+      fit.batch_size:  16
+      fit.max_epochs: 30
+      fit.patience: 5
+    splits: [0, 1, 2, 3]
+    seeds: [100, 200, 300]

config/extras.yaml

@@ -0,0 +1,34 @@
+plotting:
+  rcparams:
+    savefig.facecolor: "#ffffff"
+    savefig.dpi: 300
+    font.size: 13
+
+  var_long_names:
+    air_temperature: "Air temperature"
+    dew_point_temperature: "Dew point temperature"
+    surface_air_pressure: "Surface pressure"
+    relative_humidity: "Relative Humidity"
+    water_vapor_mixing_ratio: "Mixing ratio"
+
+  var_short_names:
+    air_temperature: "$T$"
+    dew_point_temperature: "$T_d$"
+    surface_air_pressure: "$P$"
+    relative_humidity: "$RH$"
+    water_vapor_mixing_ratio: "$r$"
+
+  var_units:
+    air_temperature: "°C"
+    dew_point_temperature: "°C"
+    surface_air_pressure: "hPa"
+    relative_humidity: "%"
+    water_vapor_mixing_ratio: "g kg$^{-1}$"
+
+  approach_names:
+    unconstrained: "Unconstrained"
+    architecture_constrained: "Architecture constrained"
+    loss_constrained: "Loss constrained"
+    offline_constrained: "Offline constrained"
+
+  approach_colors: ["#57B1FF", "#D81B60", "#FFC107", "#07846F"]