____ ____ ___ _____ ___ _____ _____ ___ ____
| \| \ / \ | |/ _]/ ___// ___/ / \ | \
| o ) D ) || __/ [_( \_( \_ | || D )
| _/| /| O || |_| _]\__ |\__ || O || /
| | | \| || _] [_ / \ |/ \ || || \
| | | . \ || | | |\ |\ || || . \
|__| |__|\_|\___/ |__| |_____| \___| \___| \___/ |__|\_|
Professor (commonly abbreviated as Prof) is a person who professes to be an expert in some art or science.
Professor is a tool to help you study complicated physical phenomena by providing tools to 1) fit machine learning models to 2D image arrays from simulations and 2) interactively explore these machine learning models in real time. Professor is most useful when studying ensembles of simulations.
A typically workflow would look like:
- A user is interested in how parameters A, B, C, & D influence some complicated physics
- User setups up parameterized simulations to study ABCD and the results of these simulations to be image arrays (a 2d matrix of float32 values)
- User runs an ensemble of simulations studying ABCD creating a dataset of image arrays
- User runs
prof-trainerto fit a machine learning model to learn the mapping from [A,B,C,D] to the image arrays - User then uses
prov-visto interactively explore the machine learning model in real time, gaining their insight into how those parameters influence the physics - Go profess your idea about ABCD!
Clone and then cd into this repo.
The default installation only checks the minimal dependencies. This process was setup to allow users to install some parts (e.g. just distributed training, or just machine learning visualization).
To install just the machine learning training:
python -m pip install .[trainer]To install just the gui machine learning visualiztion:
python -m pip install .[gui]To install both the gui and machine learning training:
python -m pip install .[all]For developers, we recommend to install dev which installs everything and a few developer tools.
python -m pip install -e .[dev]The basic requirements are:
torch>=1.12.1
torchinfo>=1.8.0
omegaconf>=2.3.0
schema>=0.7.5
torchlayers
kornia
wheel
Some example datasets will be hosted TBD. This will allow users to reproduce our machine learning model results, and to experiment with other models that may improve accuracy.
rayleigh-taylor-single is a 328 GiB dataset of shape (2000, 51, 6, 768, 256) studying the Rayleigh-Taylor instability of two fluids mixing.
pchip-hvimpact is a 2.1 TiB dataset of shape (2000, 51, 6, 1024, 1024) studying the Richtmyer-Meshkov instability as a shockwave passes through copper.
The examples folder contains example batch scripts to fit the machine learning model on the Lassen HPC. These training runs utilized distributed data parallelism to train the ML model using many GPUs using prof-trainer.
The distributed training will create folders like the following:
runs
├── Jul10_14-08-27_rzadams1003GEN_NGPU_4_LR_0.0019200000000000003_Batch_48_Loss_l1_MaxFeat_1024_MinFeat_128_keys_density,velocity_x,velocity_y,pressure,energy,materials
│ ├── 0000.pt
│ ├── 0020.pt
│ ├── 0040.pt
│ ├── 0060.pt
│ ├── 0080.pt
│ ├── 0099.pt
│ ├── args.txt
│ ├── events.out.tfevents.1720645707.rzadams1003.1180058.0
│ └── model.info
├── Jun22_11-50-13_rzvernal14GEN_NGPU_40_LR_0.019200000000000002_Batch_48_Loss_l1_MaxFeat_4096_MinFeat_256_keys_density,velocity_x,velocity_y,pressure,energy,materials
│ ├── 0000.pt
│ ├── 0020.pt
│ ├── 0040.pt
│ ├── args.txt
│ ├── events.out.tfevents.1719082213.rzvernal14.901987.0
│ └── model.info
├── Jun22_11-50-14_rzvernal24GEN_NGPU_40_LR_0.019200000000000002_Batch_48_Loss_l1_MaxFeat_4096_MinFeat_512_keys_density,velocity_x,velocity_y,pressure,energy,materials
│ ├── 0000.pt
│ ├── 0020.pt
│ ├── 0040.pt
│ ├── args.txt
│ ├── events.out.tfevents.1719082214.rzvernal24.3706721.0
│ └── model.info
where the .pt files are model checkpoints, args.txt are the command line arguments that went into prof-trainer, model.info is the layer by layer information of the machine learning model, and events* are the tensorboard training history.
You can use tensorboard to visualize the results from many training runs with
tensorboard --logdir runs
The configs folder contains example yaml files to allow you to visualize the machine learning models. These can be run like the following:
prof-vis configs/configs/pchip.yamlYou must point the checkpoint key within the yaml file to the path of the machine learning checkpoint, as well as provide details for the expected inputs to the model.
Example model checkpoints will be hosted TBD.
Jekel, C.F., Sterbentz, D.M., Stitt, T.M., Mocz, P., Rieben, R.N., White, D.A. and Belof, J.L., 2024. Machine learning visualization tool for exploring parameterized hydrodynamics. Machine Learning: Science and Technology, 5(4), p.045048. https://iopscience.iop.org/article/10.1088/2632-2153/ad8daa/meta
@article{jekel2024machine,
title={Machine learning visualization tool for exploring parameterized hydrodynamics},
author={Jekel, CF and Sterbentz, DM and Stitt, TM and Mocz, P and Rieben, RN and White, DA and Belof, JL},
journal={Machine Learning: Science and Technology},
volume={5},
number={4},
pages={045048},
year={2024},
publisher={IOP Publishing}
}This work was supported by the LLNL-LDRD Program under Project No. 21-SI-006.
SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
LLNL-CODE-2015991