run-crom 1.0.0

A package for training and testing CROM weights

CROM_offline_training

This repository contains the offline training pipeline for CROM

Prerequisites

We assume a fresh install of Ubuntu 20.04. For example,

docker run --gpus all --shm-size 128G -it --rm -v $HOME:/home/ubuntu ubuntu:20.04

Install python and pip:

apt-get update
apt install python3-pip

Dependencies

From the project directory install through pip:

pip install .

Alternatively, you may install via PyPI directly

pip install run_crom

Usage

Training

run_crom -mode train -d [data directory] -initial_lr [learning rate constant] -epo [epoch sequence] -lr [learning rate scaling sequence] -batch_size [batch size] -lbl [label length] -scale_mlp [network width scale] -ks [kernel size] -strides [stride size] [-siren_dec] [-dec_omega_0 [decoder siren omega]] [-siren_enc] [-enc_omega_0 [encoder siren omega]] 

For example

run_crom -mode train -d /home/ubuntu/sim_data/libTorchFem_data/extreme_pig/test_tension011_pig_long_l-0.01_p2d -lbl 6 -lr 1 0.1 0.05 0.02 0.01 -epo 3000 3000 3000 3000 3000 -siren_dec -batch_size 4 -scale_mlp 64 -dec_omega_0 30 --gpus 1

Testing

run_crom -mode test -m [path to .ckpt file to test]

You may also provide any built-in flags for PytorchLightning's Trainer

Data

Simulation data should be stored in a directory with the following structure. For example,

├───sim_data_parent_directory (contain multiple simulation sequences; each entry in this directory is a simulation sequence)
    ├───sim_seq_ + suffix
        ├───h5_f_0000000000.h5
        ├───h5_f_0000000001.h5
        ├───...

    ├───....

See SimulationState under simulation.py for the structure of the h5 file.