VarDACAE
This module is used to create Convolutional AutoEncoders for Variational Data Assimilation. A user can define, create and train an AE for Data Assimilation with just a few lines of code. It is the accompanying code to the paper here, published in Computer Methods in Applied Mechanics and Engineering.
Introduction
Data Assimilation (DA) is an uncertainty quantification technique used to reduce the error in predictions by combining forecasting data with observation of the state. The most common techniques for DA are Variational approaches and Kalman Filters.
In this work, we propose a method of using Autoencoders to model the Background error covariance matrix, to greatly reduce the computational cost of solving 3D Variational DA while increasing the quality of the Data Assimilation.
Data
The data used in this paper is owned by the Data Science Institute, Imperial College, London. If you do not have access to this data, please see the section below on training a model with your own data.
Installation
-
Install
vtk
by navigating to this link and installing the version applicable to your system. -
Navigate to the base directory and run:
pip install -e .
-
Run
pytest
from the home directory to ensure correct installation.
Tests
From the project home directory run pytest
.
Getting Started
To train and evaluate a Tucodec model on Fluidity data:
from VarDACAE import TrainAE, BatchDA
from VarDACAE.settings.models.CLIC import CLIC
model_kwargs = {"model_name": "Tucodec", "block_type": "NeXt", "Cstd": 64}
settings = CLIC(**model_kwargs) # settings describing experimental setup
expdir = "experiments/expt1/" # dir to save results data and models
trainer = TrainAE(settings, expdir, batch_sz=16)
model = trainer.train(num_epochs=150) # this will take approximately 8 hrs on a K80
# evaluate DA on the test set:
results_df = BatchDA(settings, AEModel=model).run()
Settings Instance
The API is based around a monolithic settings
object that is used to define all configuration parameters, from the model definition to the seed. This single point of truth is used so that, an experiment can be repeated exactly by simply loading a pickled settings
object. All key classes like TrainAE
and BatchDA
require a settings
object at initialisation.
Train a model on your own data
To train a model on your own 3D data you must do the following:
- Override the default
get_X(...)
method in theGetData
loader class:
from VarDACAE import GetData
class NewLoaderClass(GetData):
def get_X(self, settings):
"Arguments:
settings: (A settings.Config class)
returns:
np.array of dimensions B x nx x ny x nz "
# ... calculate / load or download X
# For an example see VarDACAE.data.load.GetData.get_X"""
return X
- Create a new settings class that inherits from your desired model's settings class (e.g.
VarDACAE.settings.models.CLIC.CLIC
) and update the data dimensions:
from VarDACAE.settings.models.CLIC import CLIC
class NewConfig(CLIC):
def __init__(self, CLIC_kwargs, opt_kwargs):
super(CLIC, self).__init__(**CLIC_kwargs)
self.n3d = (100, 200, 300) # Define input domain size
# This is used by ConvScheduler
self.X_FP = "SET_IF_REQ_BY_get_X"
# ... use opt_kwargs as desired
CLIC_kwargs = {"model_name": "Tucodec", "block_type": "NeXt",
"Cstd": 64, "loader": NewLoaderClass}
# NOTE: do not initialize NewLoaderClass
settings = NewConfig(CLIC_kwargs, opt_kwargs)
This settings
object can now be used to train a model with the TrainAE
method as shown above.