Modulus Configuration
Modulus employs an extension of the Hydra configuration framework to offer a highly customizable but user friendly method for configuring the majority of Modulus’ features. This is achieved by using easy to understand YAML files which contain essential hyperparameters for any physics-informed deep learning model. While you can still achieve the same level of customization in Modulus as any deep learning library, our built in configuration framework allows many of the internal features to be much more accessible. This section provides an overview of the built in configurable API Modulus provides.
Generally speaking, Modulus follows the same work flow as Hydra with just some minor differences.
For each Modulus program you should create a YAML configuration file that is then loaded into
a Python ModulusConfig
object which is used by Modulus. Consider the following example:
Listing 15 conf/config.yaml
defaults:
- modulus_default
- arch:
- fully_connected
- optimizer: adam
- scheduler: exponential_lr
- loss: sum
Listing 16 main.py
import modulus
from modulus.hydra import to_yaml
from modulus.hydra.config import ModulusConfig
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
print(to_yaml(cfg))
if __name__ == "__main__":
run()
Here, a minimal configuration (config) YAML file is shown.
A defaults list in config.yaml
is used to load predefined configurations that are supported by Modulus.
This config file is then loaded into python using the @modulus.main()
decorator, in which you specify
the location and name of your custom config.
The config object, cfg
, is then ingested into Modulus and used to setup all sorts of internals all of which
can be individually customized as discussed in the following sections.
For this example, Modulus has been configured to load a fully connected neural network, ADAM optimizer, exponential decay LR scheduler and a summation loss aggregation. Each of the included examples present in this user guide has its own config file which can be referenced.
Configs in Modulus are required to follow a common structure to ensure that all necessary parameters are provided independent
of the user explicitly providing them.
This is done by specifying the modulus_default
schema at the top of the defaults list in every configuration file which will
create the following config structure:
config
|
| <General parameters>
|
|- arch
| <Architecture parameters>
|- training
| <Training parameters>
|- loss
| <Loss parameters>
|- optimizer
| <Optimizer parameters>
|- scheduler
| <Scheduler parameters>
|- profiler
| <Profiler parameters>
This config object has multiple configuration groups that each contain separate parameters pertaining to various
features needed inside of Modulus.
As seen in the example above, these groups can be quickly populated in the defaults list (e.g. optimizer: adam
will
populate the optimizer
configuration group with parameters needed for ADAM).
The next section takes a look at each of these groups in greater detail.
- modulus_default
should always be placed at the top of your defaults list in Modulus config files. Without this, essential parameters
will not be initialized and Modulus will not run!
The --help
flag can be used with your Modulus program to bring
up some useful information on different config groups or get documentation links.
Global Parameters
Some essential parameters that you will find in a Modulus configuration include:
jit
: Turn on TorchScriptsave_filetypes
: Types of file outputs from constraints, validators and inferencersdebug
: Turn on debug logginginitialization_network_dir
: Custom location to load pretrained models from
Architecture
The architecture config group holds a list of model configurations that can be used to create different built in neural networks present within Modulus. While not required by the Modulus solver, this parameter group allows you to tune model architectures through the YAML config file or even the command line.
To initialize an architecture using the config, Modulus provides an instantiate_arch()
method that allows different architectures
to be initialized easily.
The following two examples initialize the same neural network.
Listing 17 Config model intialization
# config/config.yaml
defaults:
- modulus_default
- arch:
- fully_connected
# Python code
import modulus
from modulus.hydra import instantiate_arch
from modulus.hydra.config import ModulusConfig
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v"), Key("p")],
cfg=cfg.arch.fully_connected,
)
if __name__ == "__main__":
run()
Listing 18 Explicit model intialization
# Python code
import modulus
from modulus.hydra.config import ModulusConfig
from modulus.models.fully_connected import FullyConnectedArch
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model = FullyConnectedArch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v"), Key("p")],
layer_size: int = 512,
nr_layers: int = 6,
...
)
if __name__ == "__main__":
run()
Both of these approaches yield the same model. The instantiate_arch approach allows the model architecture to be controlled through the YAML file and CLI without loss of control. This can streamline the tuning of architecture hyperparameters.
Currently the architectures that are shipped internally in Modulus that have a configuration group include:
fully_connected
: Fully connected neural network modelfourier_net
: Fourier neural networkhighway_fourier
: Highway Fourier Network - Fourier neural network with adaptive gating unitsmodified_fourier
: Modified Fourier Network - Fourier neural network with two layers of Fourier featuresmultiplicative_fourier
: Fourier feature neural network with frequency connectionsmultiscale_fourier
: Multi-scale Fourier Feature Network - Multi-scale Fourier feature neural networksiren
: Sinusoidal Representation Networks (SiReNs) - Sinusoidal representation networkshash_net
: Neural network augmented by a multiresolution hash tablefno
: Fourier Neural Operator - 1D, 2D, or 3D Fourier neural operatorafno
: Adaptive Fourier Neural Operator - Fourier neural operator based transformer modelsuper_res
: Super Resolution Net - Convolutional super resolution modelpix2pix
: Pix2Pix Net - A pix2pix based convolutional encoder-decoder
Examples
Listing 19 Initialization of fully-connected model with 5 layers of size 128
# config.yaml
defaults:
- modulus_default
- arch:
- fully_connected
arch:
fully_connected:
layer_size: 512
nr_layers: 6
# Python code
import modulus
from modulus.hydra import instantiate_arch
from modulus.hydra.config import ModulusConfig
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
cfg=cfg.arch.fully_connected,
)
if __name__ == "__main__":
run()
Listing 20 Initialization of modified fourier model and siren model
# config.yaml
defaults:
- modulus_default
- arch:
- modified_fourier
- siren
# Python code
import modulus
from modulus.hydra import instantiate_arch
from modulus.hydra.config import ModulusConfig
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model_1 = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
frequencies=("axis,diagonal", [i / 2.0 for i in range(10)]),
cfg=cfg.arch.modified_fourier,
)
model_2 = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
cfg=cfg.arch.siren,
)
if __name__ == "__main__":
run()
Not all model parameters are controllable through the configs. Parameters that are not supported can be specified through
additional keyword arguments in the instantiate_arch
method. Alternatively, the model can be initialized in the standard
Pythonic approach.
Training
The training config group contains parameters essential to the training process of the model. This is set by default with modulus_default, but many of the parameters contained in this group are often essential to modify.
default_training
: Default training parameters (set automatically)
Parameters
Some essential parameters that you will find under the training
config group include:
max_steps
: Number of training iterations.grad_agg_freq
: Number of iterations to aggregate gradients over (default is 1). Effectively, settinggrad_agg_freq=2
will double the batch size per iteration, compared to a case with no gradient aggregation.rec_results_freq
: Frequency to record results. This value will be used as the default frequency for recording constraints, validators, inferencers and monitors. See Results Frequency for more details.save_network_freq
: Frequency to save a network checkpoint.amp
: Use automatic mixed precision. This will set the precision for GPU operations to improve performance (default is'float16'
set usingamp_dtype
).ntk.use_ntk
: Use neural tangent kernel in training (default set to False)
Loss
The loss config group is used to select different loss aggregations that are supported by Modulus. A loss aggregation is the method used to combine the losses from different constraints. Different methods can yield improved performance for some problems.
sum
: Simple summation aggregation (default)grad_norm
: Gradient normalization for adaptive loss balancinghomoscedastic
: Homoscedastic Task Uncertainty for Loss Weightinglr_annealing
: Learning Rate Annealingsoft_adapt
: Adaptive loss weightingrelobralo
: Relative loss balancing with random lookback
Optimizer
The loss optimizer group contains the supported optimizers that can be used in Modulus which includes ones that are built into PyTorch as well as from Torch Optimizer package. Some of the most commonly used optimizers include:
adam
: ADAM optimizersgd
: Standard stochastic gradient descentrmsprop
: The RMSProp algorithmadahessian
: Second order stochastic optimization algorithmbfgs
: L-BFGS iterative optimization method
as well as these more unique optimizers:
a2grad_exp
, a2grad_inc
, a2grad_uni
, accsgd
, adabelief
, adabound
,
adadelta
, adafactor
, adagrad
, adamax
, adamod
, adamp
, adamw
, aggmo
,
apollo
, asgd
, diffgrad
, lamb
, madgrad
, nadam
, novograd
, pid
, qhadam
, qhm
, radam
,
ranger
, ranger_qh
, ranger_va
, rmsprop
, rprop
, sgdp
, sgdw
, shampoo
, sparse_adam
, swats
, yogi
.
Scheduler
The scheduler optimizer group contains the supported learning rate schedulers that can be used in Modulus. By default none is specified for which a constant learning rate will be used.
exponential_lr
: PyTorch exponential learning rate decayinitial_learning_rate * gamma ^ (step)
tf_exponential_lr
: Tensorflow parameterization of exponential learning rate decayinitial_learning_rate * decay_rate ^ (step / decay_steps)
As previously mentioned, a particular benefit using Hydra configs to control Modulus is that any of these parameters can be controlled through CLI. This can be particularly useful during hyperparameter tuning or queuing up multiple runs using Hydra multirun. Here are a couple of examples which may be particularly useful when developing physics-informed models.
Listing 21 Changing optimizer and learning rate
$ python main.py optimizer=sgd optimizer.lr=0.01
Listing 22 Hyperparameter search over architecture parameters using multirun
$ python main.py -m arch.fully_connected.layer_size=128,256 arch.fully_connected.nr_layers=2,4,6
Listing 23 Training for a different number of iterations
$ python main.py training.max_steps=1000
Every parameter present in the config can be adjusted through CLI. For additional information please see the Hydra documentation.
Results Frequency
Modulus offers several different methods for recording the results of your training including recording validation, inference, batch,
and monitor results.
Each of these can be individually controlled in the training config group, however, typically it’s preferred for each to have the same frequency.
In these instances, the rec_results_freq
parameter can be used to control all of these parameters uniformly.
The following two config files are equivalent.
# config/config.yaml
defaults:
- modulus_default
training:
rec_results_freq: 1000
rec_constraint_freq: 2000
# config/config.yaml
defaults:
- modulus_default
training:
rec_validation_freq: 1000
rec_inference_freq: 1000
rec_monitor_freq: 1000
rec_constraint_freq: 2000
Changing Activation Functions
Activations functions are one of the most important hyperparameters to test for any deep learning model.
While all of Modulus’ networks have default activations functions that have been seen to provide the best performance,
specific activations may perform better than others on a case to case basis.
Changing a activation function is straight forward using the instantiate_arch
method:
Listing 24 Initializing a fully-connect model with Tanh activation functions
# Python code
import modulus
from modulus.hydra import instantiate_arch
from modulus.hydra.config import ModulusConfig
from modulus.models.layers.layers import Activation
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model_1 = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
cfg=cfg.arch.fully_connected,
activation_fn=Activation.TANH,
)
if __name__ == "__main__":
run()
Activation functions are not currently supported in the config files. They must be set in the Python script.
Many of Modulus’ models also include support for Adaptive Activation Functions which can be turned on in the config file or explicitly in the code:
# config/config.yaml
defaults:
- modulus_default
- arch:
- fully_connected
arch:
fully_connected:
adaptive_activations: true
Multiple Architectures
For some problems, its better to have multiple models to learn the solution of different state variables.
This may require the use of models that are the same architecture with different hyperparameters.
We can have multiple neural network models with the same architecture using config group overrides in Hydra.
Here the arch_schema
config group is used to access an architecture’s structured config.
Listing 25 Extending configs with customized architectures
# config/config.yaml
defaults:
- modulus_default
- /arch_schema/fully_connected@arch.model1
- /arch_schema/fully_connected@arch.model2
arch:
model1:
layer_size: 128
model2:
layer_size: 256
Listing 26 Initialization of two custom architectures
# Python code
import modulus
from modulus.hydra import instantiate_arch
from modulus.hydra.config import ModulusConfig
@modulus.main(config_path="conf", config_name="config")
def run(cfg: ModulusConfig) -> None:
model_1 = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
cfg=cfg.arch.model1,
)
model_2 = instantiate_arch(
input_keys=[Key("x"), Key("y")],
output_keys=[Key("u"), Key("v")],
cfg=cfg.arch.model2,
)
if __name__ == "__main__":
run()
Run Modes
Modulus has two different run modes available for training and evaluation:
train
: Default run mode. Trains the neural network.eval
: Evaluates provided inferencers, monitors and validators using the last saved training checkpoint. Useful for post-processing after the training is complete.
Listing 27 Changing run mode to evaluate
# config/config.yaml
defaults:
- modulus_default
run_mode: 'eval'
Criterion Based Stopping
Modulus supports early training termination, based on a user specified criterion, before the maximum number of iterations is reached.
metric
: Metric to be monitored during the training. This can be the total loss, individual loss terms, validation metrics, or metrics in the monitor domain. For example, in the annular ring example, you can choose loss, loss_continuity, momentum_imbalance, or l2_relative_error_u as the metric. Note the use of l2_relative_error_ for metrics from the validation domain, this is consistent with the tag used for tensorboard plots.min_delta
: Minimum required change in the metric to qualify as a training improvement.patience
: Number of training steps to wait for a training improvement to happen.mode
: Choose ‘min’ if the metric is to be minimized, or ‘max’ if the metric is to be maximized.freq
: Frequency of evaluating the stop criterion. Note that if using a metric from the validation or monitor domain, freq should be a multiplier of the rec_validation_freq or rec_monitor_freq.strict
: If True, raises an error in case the metric is not valid.
Listing 28 Defining a stopping criterion for training
# config/config.yaml
defaults:
- modulus_default
stop_criterion:
- metric: 'l2_relative_error_u'
- min_delta: 0.1
- patience: 5000
- mode: 'min'
- freq: 2000
- strict: true
When using a metric from the validation domain, criterion based stopping can also serve as an early stopping regularization method for data-driven models.