Awesome

CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR2022 Oral)

Paul Gavrikov, Janis Keuper

Paper | ArXiv | HQ Poster | Oral Presentation

Abstract: Currently, many theoretical as well as practically relevant questions towards the transferability and robustness of Convolutional Neural Networks (CNNs) remain unsolved. While ongoing research efforts are engaging these problems from various angles, in most computer vision related cases these approaches can be generalized to investigations of the effects of distribution shifts in image data. In this context, we propose to study the shifts in the learned weights of trained CNN models. Here we focus on the properties of the distributions of dominantly used 3x3 convolution filter kernels. We collected and publicly provide a data set with over 1.4 billion filters from hundreds of trained CNNs, using a wide range of data sets, architectures, and vision tasks. In a first use case of the proposed data set, we can show highly relevant properties of many publicly available pre-trained models for practical applications: I) We analyze distribution shifts (or the lack thereof) between trained filters along different axes of meta-parameters, like visual category of the data set, task, architecture, or layer depth. Based on these results, we conclude that model pre-training can succeed on arbitrary data sets if they meet size and variance conditions. II) We show that many pre-trained models contain degenerated filters which make them less robust and less suitable for fine-tuning on target applications.

⚠ News

11/2022: We published a follow-up paper on our previous observations of medical imaging models at MedNeurIPS'22: https://arxiv.org/abs/2210.13799

07/2022: We have received interest of many individuals in using the metrics we propose in our paper to better understand CNN training. Therefore, we have released a torch library here: https://github.com/paulgavrikov/torchconvquality

Versions

If you are looking for our specialized dataset on Robustness head to https://github.com/paulgavrikov/cvpr22w_RobustnessThroughTheLens.

Environment

We have executed this with Python 3.8.8 on Linux 3.10.0-1160.24.1.el7.x86_64. The scripts should however work with most python3 versions and OS.

To install all necessary modules please run:

pip install -r requirements.txt

or install these modules manually with your desired package manager:

numpy==1.21.2
scipy
scikit-learn==0.24.1
matplotlib==3.4.1
pandas==1.1.4
fast-histogram==0.10
KDEpy==1.1.0
tqdm==4.53.0
colorcet==2.0.6
h5py==3.1.0
tables==3.6.1

Prepare

Download dataset.h5. This file contains the filters and meta information as individual datasets. If the filename ends with .xz you first need to decompress it: xz -dv data.csv.xz. Note that this will increase the size by 225%.

The filters are linked as a Nx9 numpy.float32 array under the /filter dataset. Every row is one filter and the row number is also the filter ID (i.e. the first row is filter ID 0). To reshape a filter f back to its original shape use f.reshape(3, 3).

The meta information is stored as a pandas.DataFrame under /meta. Following is an out of order list of column keys with a short description. Other column keys can and should be ignored. The table has a Multiindex on [model_id, conv_depth, conv_depth].

Run

Adjust dataset_path in https://github.com/paulgavrikov/CNN-Filter-DB/blob/main/main.ipynb and run the cells.

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{Gavrikov_2022_CVPR,
    author    = {Gavrikov, Paul and Keuper, Janis},
    title     = {CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {19066-19076}
}

and the dataset:

@dataset{cnnfilterdb2022,
  author       = {Paul Gavrikov and
                  Janis Keuper},
  title        = {CNN-Filter-DB},
  month        = jun,
  year         = 2022,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.6371680},
  url          = {https://doi.org/10.5281/zenodo.6371680}
}

Legal

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Funded by the Ministry for Science, Research and Arts, Baden-Wuerttemberg, Germany Grant 32-7545.20/45/1 (Q-AMeLiA).