Benchmarks#

This section provides benchmark results for various datasets and physics models.

Note

Benchmarks are defined in the deepinv/benchmarks repository. To contribute a new benchmark or add your solver to an existing benchmark, please refer to this repository.

List of benchmarks#

Benchmark	Dataset	Physics	Noise Model
CBSD68 denoising	`CBSD68`	`Denoising`	`GaussianNoise`
DIV2K Gaussian Deblurring	`DIV2K`	`Blur`	`GaussianNoise`
DIV2K Inpainting (30% random)	`DIV2K`	`Inpainting`	`ZeroNoise`
DIV2K Super Resolution (2x)	`DIV2K`	`Downsampling`	`ZeroNoise`

Testing your method on benchmarks#

To evaluate your own reconstruction methods on these benchmarks, install deepinv_bench:

pip install git+https://github.com/deepinv/benchmarks.git#egg=deepinv_bench

If you have already installed benchmarks, you can update it with:

pip install --upgrade --force-reinstall --no-deps git+https://github.com/deepinv/benchmarks.git#egg=deepinv_bench

and then run on python:

from deepinv_bench import run_benchmark
import deepinv as dinv
my_solver = ... # replace with your reconstruction method
results = run_benchmark(my_solver, "benchmark_name")

where benchmark_name is the name of the benchmark and my_solver is your reconstruction method which receives (y, physics, **kwargs) where

y is a torch.Tensor containing the measurements,
physics is the forward operator

and outputs a torch.Tensor containing the reconstructed image.