Configuration Guide

This guide explains how to configure and customize the benchmark. As mentioned in the quickstart, you need two configuration files:

--parallel-config — SLURM cluster settings (CPUs, GPUs, walltime, modules)
--config — Experiment definition (datasets, solvers, parameters, execution grid)

Available Configuration Files

All configuration files are located in the configs/ directory:

File	Purpose
`config_parallel.yml`	SLURM cluster configuration
`highres_imaging.yml`	High-resolution image reconstruction
`tomography_2d.yml`	2D tomography experiments
`tomography_3d.yml`	3D tomography experiments

Parallel Configuration (SLURM)

The --parallel-config file (e.g., configs/config_parallel.yml) defines cluster-level settings. Here is an example for a SLURM cluster (Jean-Zay):

backend: submitit
slurm_time: 30                    # Job walltime in minutes
slurm_additional_parameters:
  cpus-per-task: 10               # CPU cores per job
  qos: qos_gpu-dev                # Queue used for the jobs
  account: your_account_here      # Your cluster account
  constraint: v100-32g            # GPU type (e.g., V100 with 32GB)
slurm_setup:                      # Commands run before job starts
  - module purge
  - module load pytorch-gpu/py3/2.7.0
  - export NCCL_DEBUG=INFO

Key parameters to customize:

cpus-per-task — Number of CPU cores per job
slurm_time — Maximum runtime (minutes)
constraint — GPU type (e.g., v100-32g, a100)
slurm_setup — Environment modules and variables for your cluster

Note: The number of GPUs per job is specified in the main config file, not here.

Main Configuration (Experiments)

The --config file (e.g., configs/highres_imaging.yml) defines what experiments to run. Let’s walk through its structure:

1. Dataset Configuration

Specify the dataset and its parameters:

dataset:
  - highres_color_image:
      ...

Dataset implementations are in datasets/ (e.g., datasets/highres_color_image.py).

2. Solver Configuration

Define which reconstruction algorithms to run:

solver:
  - PnP:
      ...

Available solvers are in solvers/ (e.g., solvers/pnp.py).

3. Execution Grid

This section defines GPU configurations and parallelization strategies. Each row specifies one complete experiment configuration:

slurm_gres, slurm_ntasks_per_node, slurm_nodes, distribute_physics, distribute_denoiser, patch_size, overlap, max_batch_size: [
  ["gpu:1", 1, 1, false, false, 0,   0,  0],
  ["gpu:2", 2, 1, true,  true,  448, 32, 0],
  ["gpu:4", 4, 1, true,  true,  448, 32, 0],
]

Key fields:

slurm_gres — GPU resource request ("gpu:1" = 1 GPU, "gpu:2" = 2 GPUs)
slurm_ntasks_per_node — Number of parallel tasks per node
slurm_nodes — Number of compute nodes
distribute_physics — Parallelize physics operators across GPUs
distribute_denoiser — Enable spatial tiling: split large images into patches and process them in parallel on available GPUs

Example configurations:

Single GPU (no parallelization):

["gpu:1", 1, 1, false, false, 0, 0, 0]

Uses 1 GPU with no distributed processing.

Multi-GPU (2 GPUs in parallel):

["gpu:2", 2, 1, true, true, 448, 32, 0]

Uses 2 GPUs in parallel on one node, distributing both physics operators and denoiser. For more details on distributed computing, see the DeepInv distributed documentation.

Customizing Your Configuration

To create your own experiment:

Copy an existing config:

cp configs/highres_imaging.yml configs/my_experiment.yml

Edit dataset parameters (image size, noise levels, etc.)
Adjust solver hyperparameters (step size, iterations, denoiser settings)
Configure GPU resources in the execution grid

Run your benchmark:

benchopt run . \
    --parallel-config ./configs/config_parallel.yml \
    --config ./configs/my_experiment.yml