FilLTP89/SeismoALICE

Adversarially Learned Inference (ALI) + Conditional Entropy (CE) for seismic signal prediction

SeismoALICE

Adversarially Learned Inference (ALI) + Conditional Entropy (CE) for prediction of seismic signals based on Physics-Based numerical simulations.

Gatti, F. and Clouteau, D. (2020). "Towards blending Physics-Based numerical simulations and seismic databases using Generative Adversarial Network". Submitted at Computer Methods in Applied Mechanics and Engineering, Special Issue: "AI in Computational Mechanics and Engineering Sciences". This work was inspired by the following original works:

Adversarially Learned Inference with Conditional Entropy (ALICE)

Li, C. et al. (2017). "ALICE: Towards Understanding Adversarial Learning for Joint Distribution Matching". Duke University. NIPS, 2017. https://arxiv.org/abs/1709.01215

DCGAN.torch: Train your own image generator

Radford, A. et al. (2015). "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks". ICLR 2016. https://arxiv.org/abs/1511.06434

ANN2BB: NN-based broadband ground motion generator from 3D earthquake simulations

Paolucci, R., Gatti, F. et al. (2018). "Broadband Ground Motions from 3D Physics-Based Numerical Simulations Using Artificial Neural Networks". BSSA, 2018 htt ps://doi.org/10.1785/0120170293

Prerequisites

• Computer with Linux or OSX
• Torch

For training, an NVIDIA GPU is strongly recommended for speed. CPU is supported but training is very slow. GPU-based version was tested on cuda 9.0, cuda 9.2.

Installing dependencies

Before getting started, some prerequisites must be installed via pip:

pip install -r requirements.txt

Before running the scripts, it is necessary to add ./src to the PYTHONPATH:

export PYTHONPATH="./src"

Getting Started

Three Deep-Convolutional Adversarial AutoEncoders (DCAAE) can be trained and tested, according to the reconstruction frequency band:

1. broadband (bb) seismic signals (0-30 Hz)
2. filtered (fl) seismic signals (0- Hz, where fc can be set via --cutoff option
3. hybrid (hb) seismic signals (0- + -30 Hz)

The original contribution of this paper is hybrid DCAAE, which takes a low-frequency signal (0- Hz) and reconstructs the broad-band one.

Each DCAAE can undergo three different actions (to be listed in the actions.txt file [True/False])

1. tract: train
2. trplt: plot/generate
3. trcmp: compare with ANN2BB

Each action implies the choice of a corresponding strategy (to be specified in the strategy.txt file) for keywords encoder,decoder (...and others, see below) corresponding to the desired network (from scratch template or pre-trained model). For each keyword, two alternatives are possible:

1. None: a generic CNN will be created, with random weights
2. The path to pre-trained models (under .pth format) to be used in the analysis

Extra keywords can be added as column's headers in the strategy.txt file: they are needed for comparison purposes and/or to test the discriminator performances.

Training and testing are performed by alternative running ./src/aae_drive_bbfl.py (for broadband and filtered) and ./src/aae_drive_hb.py (for hybrid). The latter requires pre-trained broadband and filtered DCAAE: sequential training is possible (via actions.txt file) or pre-trained models can be adopted instead.

Signal Databases

To train/test the different DCAAE, an extraction of 100 signals from the STEAD database is provided in the database folder. Seismic signals are 40.96 s-long, sampled at 100 Hz (nt=4096 time steps).

• ths_trn_nt4096_ls128_nzf8_nzd32.pth: training set (80%)
• ths_tst_nt4096_ls128_nzf8_nzd32.pth: testing set (10%)
• ths_vld_nt4096_ls128_nzf8_nzd32.pth: validation set (10%)

ls : latent space vector size nzd: latent space channels (broadband) nzf: latent space channels (filtered)

The tag nt4096_ls128_nzf8_nzd32.pth is passed as input to the drive files --dataset. --dataroot flag indicates where the files with this tag are stored.

Train DCAAE

In the following, basics command line examples are provided to train each DCAAE (broadband,filtered and hybrid) over 5000 epochs (--niter) and with --cuda, over 1 GPU (--ngpu).

• broadband:

   python3 ./src/aae_drive_bbfl.py --dataroot='./database/stead' --dataset='nt4096_ls128_nzf8_nzd32.pth' --cutoff=1. --imageSize=4096 --latentSize=128  --niter=5000 --cuda --ngpu=1 --nzd=32 --rlr=0.0001 --glr=0.0001 --outf='./imgs' --workers=8 --nsy=100 --batchSize=10 --actions='./actions_bb.txt' --strategy='./strategy_bb.txt'

• filtered:

python3 ./src/aae_drive_bbfl.py --dataroot='./' --dataset='nt4096_ls128_nzf8_nzd32.pth'  --cutoff=1. --imageSize=4096 --latentSize=128  --niter=5000 --cuda --ngpu=1 --nzf=8 --rlr=0.0001 --glr=0.0001 --outf='./imgs' --workers=8 --nsy=100 --batchSize=100 --actions='./actions_fl.txt' --strategy='./strategy_fl.txt' 

• hybrid

python3 ./src/aae_drive_hb.py --dataroot='./database/stead' --dataset='nt4096_ls128_nzf8_nzd32.pth' --cutoff=1. --imageSize=4096 --latentSize=128  --niter=3000 --cuda --ngpu=1 --nzd=32 --rlr=0.0001 --glr=0.0001 --outf='./imgs' --workers=8 --nsy=100 --batchSize=10 --actions='./actions_hb.txt' --strategy='./strategy_hb.txt'