Configure and Run Solver

Overview

Use this skill when you have a directory of measurements (2D images) and want to reconstruct a 3D volume using the point-based solver entrypoint driver_point.py.

This skill covers:

• •creating a config YAML that points to your measurement data and output/cache paths
• •running the solver on remote GPU nodes
• •verifying success by locating outputs and optionally computing a reconstruction metric against a GT volume

Instructions

Search data/run*.sh to see whether there are driver_point.py calling, if so, use them as reference. We prefer parallel execute commands below for different measurement cases for a fast turnaround.

1. •

   Create a solver config YAML (for example, config/solve_my_experiment.yaml).

   Include (at minimum) these fields:

   • •data.raw_b_dir: path to the measurement images directory
   • •data.points_dir: path to cache/store processed point clouds
   • •data.crop_box_b: crop for the images in [x0, y0, x1, y1]
   • •data.crop_box_A: crop for the 3D volume in [x0, y0, z0, x1, y1, z1]
   • •solver.type: solver backend (for example newton, ista, etc.)

If the YAML already exist and from the history you know that there is change in raw_b_dir after last run, delete points_dir before running.

1. •

   Decide which machine/GPU(s) to run on, see skill use_gpu_node

2. •

   Run the solver via driver_point.py.

3. •

   Check outputs.

   A successful run should produce a reconstruction artifact under directory output_dir/latest_date (often a reconstruction.pt). If you want to monitor logs, monitor run.log file inside output_dir/latest_date.

4. •

   (Optional) Compute a metric vs GT.

   If you have a ground-truth .pt volume, compute a quantitative metric using scripts/compute_3d_metric.py.

What success looks like

You should consider this skill successful when:

• •driver_point.py finishes without error, and
• •the configured output folder under result/ contains a reconstructed volume (commonly reconstruction.pt), and
• •(optional) your metric script runs and reports reasonable values.

Examples

Example: run on a remote GPU node

SSH to a GPU host, activate the environment, choose GPUs via CUDA_VISIBLE_DEVICES, then run the solver:

PYTHON="/data/volume3/share_storage/ym.xiao/miniconda/envs/torch/bin/python"
ssh gpu04 "cd /home/ym.xiao/workspace/lightfield_linearsys && \
CUDA_VISIBLE_DEVICES=0,1 $PYTHON driver_point.py --config config/solve_my_experiment.yaml'"

Example: compute metrics against ground truth

Run the metric script on the produced reconstruction:

$PYTHON scripts/compute_3d_metric.py result/.../reconstruction.pt --gt-path data/.../gt.pt