This repository reproduces the results of the paper "Beyond calibration: estimating the grouping loss of modern neural networks" by Alexandre Perez-Lebel, Marine Le Morvan, and Gaël Varoquaux (ICLR 2023).

A separate package to easily estimate the grouping loss of a classifier is available at:

git clone https://github.com/aperezlebel/beyond_calibration

conda install --file requirements.txt -c conda-forge

src/test_figures.py generates the figures present in the paper. Each is written as a test function that can be run with pytest.

Example:

# Generate the first figure of the paper.
pytest src/test_figures.py::test_fig1 -s

Depending on the figures you want to reproduce, you may need to install data.

• Figures 1 to 5 and 9 to 12: no data required. You can already run the commands.
• Figures 6 to 8 and 13 to 27: ⚠️ data required. You should install the data before running the commands: see the section 'Full data build' below.

This procedure builds all the datasets, enabling the reproduction of all the figures (main text appendix). If you want to reproduce only a subset of the figures, jump to the 'Partial data build for specific figures only' section.

This downloads all the dataset archives (ImageNet-1K validation set, ImageNet-R, and ImageNet-C), extracts them, and builds the merged version of ImageNet-C.

pytest src/test_data.py::test_make_datasets -s

pytest src/test_data.py::test_download_datasets -s

pytest src/test_data.py::test_extract_datasets -s

pytest src/test_data.py::test_make_imagenet_c_merged_no_rep -s

pytest -n 15 src/test_data.py::test_download_vision_networks -s
pytest -n 2 src/test_data.py::test_download_nlp_network -s

Since we work in the last layer's feature space, we forward once and for all the datasets through each network, creating as many datasets of embeddings. The evaluation then only looks at those smaller datasets.

pytest -n 30 src/test_data.py::test_forward_vision_networks -s
pytest -n 2 src/test_data.py::test_forward_nlp_network -s

Depending on the figures you want to reproduce, build a subset of the data as follows:

Figure	Requires data	Resource intensive	Command
Figure 1			pytest src/test_figures.py::test_fig1 -s
Figure 2			pytest src/test_figures.py::test_fig2 -s
Figure 3			pytest src/test_figures.py::test_fig3 -s
Figure 4			pytest src/test_figures.py::test_fig4 -s --njobs 120
Figure 5			pytest src/test_figures.py::test_fig5 -s
Figure 6			pytest src/test_figures.py::test_fig6 -s -n 4 --njobs 15
Figure 7			pytest src/test_figures.py::test_fig7 -s --njobs 15
Figure 8			pytest src/test_figures.py::test_fig8 -s -n 4 --njobs 15

Click for appendix Figures 9 to 27.

Figure	Requires data	Resource intensive	Command
Figure 9			pytest src/test_figures.py::test_fig9 -s
Figure 10			pytest src/test_figures.py::test_fig10 -s
Figure 11			pytest src/test_figures.py::test_fig11 -s
Figure 12			pytest src/test_figures.py::test_fig12 -s
Figure 13			pytest src/test_figures.py::test_fig13 -s --njobs 15
Figure 14			pytest src/test_figures.py::test_fig14 -s --njobs 120
Figure 15			pytest src/test_figures.py::test_fig15 -s -n 15 --njobs 8
Figure 16			pytest src/test_figures.py::test_fig16 -s -n 15 --njobs 8
Figure 17			pytest src/test_figures.py::test_fig17 -s -n 15 --njobs 8
Figure 18			pytest src/test_figures.py::test_fig18 -s -n 15 --njobs 8
Figure 19			pytest src/test_figures.py::test_fig19 -s -n 15 --njobs 8
Figure 20			pytest src/test_figures.py::test_fig20 -s -n 15 --njobs 8
Figure 21			pytest src/test_figures.py::test_fig21 -s -n 15 --njobs 8
Figure 22			pytest src/test_figures.py::test_fig22 -s -n 15 --njobs 8
Figure 23			pytest src/test_figures.py::test_fig23 -s -n 15 --njobs 8
Figure 24			pytest src/test_figures.py::test_fig24 -s -n 15 --njobs 8
Figure 25			pytest src/test_figures.py::test_fig25 -s -n 15 --njobs 8
Figure 26			pytest src/test_figures.py::test_fig26 -s -n 15 --njobs 8
Figure 27			pytest src/test_figures.py::test_fig27 -s -n 15 --njobs 8

Comments:

• Figures marked as 'resource intensive' are recommended to be run on a computing cluster. The complete experiments were run on a 256-CPU node for several days. The expensive part is to forward the datasets through the networks to create datasets of embeddings of inputs in the last layer feature space. Then, the evaluation of the grouping loss with the partitioning is fast.
• Some tests are parallelized using the pytest-xdist plugin through the -n argument or internally using the --njobs argument. When specified, adjust the number of workers (-n or --njobs) depending on your node's CPU count.
• Add --disable-warnings to the pytest command to silent warnings.

• src/test_data.py: code building the datasets necessary to reproduce the experiments.

• src/test_figures.py: code generating the figures present in the paper.

• src/partitioning.py: main partitioning algorithm (implemented in the cluster_evaluate function). It partitions the feature space in each level set and returns the bins' region scores, counts, and average confidence scores.

• src/networks/*: code related to vision and NLP networks. All networks inherit the BaseNet class in src/networks/base.py, which implements functions that load the networks, forward samples, extract transformed samples in the high-level feature space, confidence scores, etc...

• _utils.py, _plot.py, _linalg.py are implementing helper functions.

• tests/*: unit tests to test the functions of the repository.

Should you have any questions, comments, or feedback, please open an issue or reach out!

• Email: alexandre [dot] perez [at] inria [dot] fr
• Twitter: @aperezlebel
• Website: https://perez-lebel.com