Conditional Invertible Neural Network (cINN) Project

This project implements a Conditional Invertible Neural Network (cINN) designed for modeling complex relationships between input features and target outputs. The architecture includes a condition network, coupling subnet constructors, and a main cINN model wrapper.

Project Structure

conditional-inn-project
├── src
│   ├── cINN_cnn.py          # Implementation of the cINN architecture
│   ├── data
│   │   └── dataprep.py      # Data preparation functions
│   ├── models
│   │   └── conditional_inn.py # Definition of the ConditionalINN class
│   ├── train.py             # Training script for the cINN model
│   └── types
│       └── index.py         # Custom types and interfaces
├── requirements.txt          # Project dependencies
└── README.md                 # Project documentation

Setup Instructions

1. Clone the repository:

   git clone <repository-url>
   cd conditional-inn-project
   

2. Install the required dependencies:

   pip install -r requirements.txt
   

Usage

To train the cINN model, run the following command:

python src/train.py

Make sure to prepare your dataset and adjust the data loading parameters in src/train.py as needed.

Realistic Conditioning Experiment

The current src/testcinn.py experiment conditions on Ux, Uy, Depth, and Pressure change, then predicts X, Dip, Poisson ratio, Shear Modulus, Biot's coefficient, Offset, Thickness, and Width.

Run from the repository root:

python src/testcinn.py

Outputs are saved into timestamped folders under src/cinn_results/.

For a quick server smoke test before a full run:

CINN_EPOCHS=1 CINN_MAX_ROWS=256 CINN_COVERAGE_K=2 CINN_POSTERIOR_SAMPLES_EVAL=8 CINN_POSTERIOR_SAMPLES_PLOT=32 python src/testcinn.py

Multi-Point Synthetic Dataset

Each scenario retains the approved parameter distributions and stores Ux and Uy in metres at 2,001 common positions from -10 km to +10 km. Depth and Pressure change are stored as known conditioning values; the other seven physical parameters are stored as inversion targets. The analytical solution is included directly in this repository, so no external solver script is called.

Small validation run:

python src/synthetic_data_multipoint.py --n-scenarios 4 --chunk-size 2

When --output-dir is omitted, the run is written to a timestamped directory under src/multipoint_data/. For a named larger server run:

python src/synthetic_data_multipoint.py \
  --n-scenarios 10000 \
  --chunk-size 100 \
  --output-dir /path/to/multipoint_10k

Run it in tmux and keep a terminal log:

tmux new -s multipoint
python src/synthetic_data_multipoint.py \
  --n-scenarios 10000 \
  --chunk-size 100 \
  --output-dir /path/to/multipoint_10k 2>&1 | tee multipoint_10k.log

Detach with Ctrl-b, then d. Reconnect later with:

tmux attach -t multipoint

Multi-Point cINN Training

Train one cINN from every NPZ chunk listed by a completed dataset manifest:

python src/testcinn_multipoint.py \
  --data-dir src/multipoint_data/20260622_153845

The script loads all scenarios into CPU memory, splits by scenario, encodes the full Ux/Uy profiles with a 1D CNN, and conditions the cINN on the profile features plus Depth and Pressure change. Results are written to timestamped directories under src/cinn_results_multipoint/.

Shear modulus uses the synthetic prior support of 0.1-20 GPa and is displayed in GPa in posterior figures. Checkpoints created before this bounded transform must not be reused; retrain from the existing NPZ dataset. Dataset regeneration is not required.

Each run also reports condition-shuffle and posterior-contraction diagnostics. A positive shuffled-minus-correct NLL indicates that the cINN uses its conditioning input. A posterior/prior 68% width ratio below one indicates information gain; values near one indicate little posterior contraction. These quantities are diagnostics, not hard identifiability thresholds.

Dataset distribution figures compare the scenario-level training and validation parameter distributions, summarize Ux/Uy profile envelopes and peak amplitudes, and show the training-split Pearson correlation matrix. Figures are saved as 300 dpi PNG and vector PDF files in the timestamped result directory.

Run a small CPU smoke experiment before starting a full server job:

python src/testcinn_multipoint.py \
  --data-dir src/multipoint_data/20260622_153845 \
  --epochs 1 \
  --max-scenarios 32 \
  --posterior-samples-plot 32 \
  --posterior-samples-eval 8 \
  --coverage-scenarios 2 \
  --num-blocks 2 \
  --device cpu

License

This project is licensed under the MIT License. See the LICENSE file for more details.