Constrained Attention Filter (CAF)

(ECCV 2020) Tensorflow implementation of A Generic Visualization Approach for Convolutional Neural Networks

Paper | 1 Min Video | 10 Mins Video

Qualitative Evaluation -- L2-CAF Slow Motion Convergence

	One Object	Two Objects
Last Conv
Intermediate Conv

TL;DR

L2-CAF has three core components:

1- TF filter This is the function that inserts L2-CAF inside a network (E.g, inside a DenseNet). L2-CAF is by default disabled; it is passive during classification. To active/de-activate L2-CAF (turn on and off the filter), I use the bool atten_var_gate. False deactivate L2-CAF, while True activates the filter.

2- Optimization loop In this loop, we computes the class-oblivious and class-specific loss and leverage gradient descent to minimize it. When the loss stabilize (loss - prev_loss< 10e-5), break out of the loop.

3- Finalize filter before saving After convergence, the output filter is normalized (L2-Norm|Softmaxed|Gauss-ed) before generating the heatmap.

Requirements

• Python 3+ [Tested on 3.7]
• Tensorflow 1.X [Tested on 1.14]

ImageNet Pretrained Models

I used the following

• DenseNet
• InceptionV1
• ResNet

Usage example

Update base_config._load_user_setup with your configurations. Mainly, set the location of pre-trained model (e.g, densenet). The released code optimizes the constrained attention filter on samples images from the "input_imgs" directory. However, if you plan to run the code on a whole dataset (e.g, ImageNet), you shoud set the local_datasets_dir in _load_user_setup

The unit L2-Norm constrained attention filter has two operating modes.

• visualize_attention.py is the script for the vanilla "slow" (4 seconds) mode. I recommend running this first before experimenting with the fast L2-CAF version. The code of this mode is much easier to understand. The script's main function sets all the hyper-parameters needed. I will ellaborate more on each hyper-parameter soon.

• visualize_attention_fast.py is the script for the fast (0.3 seconds) mode. The script only supports denseNet. I will add support to Inception and ResNet soon. This script only works for visualizing attention is the last conv layer. I only use it for quantitative evaluation experiments, for instance, when I evaluate L2-CAF using ImageNet validation split.

TODO LIST

• Add Fast L2-CAF on DenseNet
• Add InceptionNet and ResNet support
• Document to use the code
• Document the intermediate layer visualization
• Document extra technical tricks not mentioned in the paper

Contributing

It would be great if someone re-implement this in pytorch. Let me know and I will add a link to your Pytorch implementation here

MISC Notes

• We did not write localization evaluation code. We used the matlab code released by CAM in Tables 1 and 3. We used the python code released by ADL in Table 2. Feel free to evaluate L2-CAF localization with other evaluation codes.

• The softmax and Gaussian filters are released upon a reviewer request. The current Gaussian filter implementation is hard-coded to support only 7x7 attention filter. It is straight forward to extend it for any odd filter-size (e.g., 13x13). However, for even filter-size I think more changes are required. The last conv layer in standard architectures is 7x7. So the current configuration should cover most case-scenarios.

• I used modules of this code (especially the nets package) in multiple projects, so there is a lot of code that is not related to L2-CAF. I will iteratively clean the code. The TL;DR section, at the top of the readme file, highlights the core functions related to L2-CAF.

Release History

• 1.0.0
  • First commit Vanilla L2-CAF on DenseNet, InceptionV1, and ResNet50V2 on 12, 15,18 July 2020
  • Add Fast L2-CAF on DenseNet 21 July 2020
  • Add Fast L2-CAF on Inception 22 July 2020
  • Add Fast L2-CAF on ResNet 23 July 2020

Citation

@inproceedings{taha2020generic,
title={A Generic Visualization Approach for Convolutional Neural Networks},
author={Taha, Ahmed and Yang, Xitong and Shrivastava, Abhinav and Davis, Larry},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
year={2020}
}