Skip to content

Latest commit

 

History

History
124 lines (82 loc) · 7.83 KB

README.md

File metadata and controls

124 lines (82 loc) · 7.83 KB
Raw

DExT: Detector Explanation Toolkit

The objective of the toolkit is to explain the detections of an object detector using the complete bounding box and class predictions.

As object detectors are prone to non-local effects and utilize context, providing the rationale behind each components of the detection, the complete detection, and all detections in an image is essential.

A poster with preliminary results can be found here.

The toolkit is a part of the thesis work supervised by Professor. Dr. Paul G Plöger, Prof. Dr. Matias Valdenegro-Toro, Professor. Dr. Nico Hochgeschwender and Octavio Arriaga. The thesis proposal can be found here.

Installation

pip install -e .

Note: The code is tested using tensorflow==2.5.1 and tensorflow_gpu==2.3.0.

Generate explanations for a single image

dext_explainer -c config/explainer.gin -i <interpretation_method> -m <model_name> --explain_mode single_image --input_image_path <image_path>

Generate explanations on COCO dataset images

dext_explainer -c config/explainer.gin -i <interpretation_method> -m <model_name> --num_images <number_of_images>

Evaluate explanations

dext_evaluator -c config/evaluator.gin -i GuidedBackpropagation -m SSD512

Explain images on a custom model

dext_explainer -c config/explainer.gin -m <model_name> --explain_mode single_image --input_image_path <image_path> --image_size <custom_model_image_size> --dataset_name <dataset_name>

Refactor the preprocessing, postprocessing, model building, class names of the data used to train the model, and conversion to backpropagate bounding box in image coordinates. Include all the necessary function in the factory module. An example custom model, 'MarineDebris', is included in the repository. MarineDebris model does not exist and the code lines are just an example. The preprocessing, postprocessing, and model building steps are included in /model/marine_debris_ssd/utils.py. Class names are added to class_names.py. All factories are updated with this information.

Important explainer arguments

Argument Description
config, c configuration file with dataset paths
interpretation_method, i name of the explanation method
model_name, m name of the model to explain
explain_mode single image explanation or explain random images from dataset
dataset_name name of the dataset

Object detectors available

  1. EFFICIENTDETD{x}, x = 0, 1, 2, 3, 4, 5, 6, 7, 7x
  2. SSD512
  3. FasterRCNN

All above detectors are available with COCO weights. The weights are ported from [1], [2], and [3], respectively.

Interpretation methods available

  1. IntergratedGradients (IG)
  2. GuidedBackpropagation (GBP)
  3. SmoothGrad_Integrated Gradients (SIG)
  4. SmoothGrad_GuidedBackpropagation (SGBP)
  5. GradCAM, Gradient-weighted Class Activation Mapping (GradCAM)
  6. LIME, Local Interpretable Model-agnostic Explanations (LIME)
  7. DeepExplainer, SHapley Additive exPlanations (SHAP) + DeepLIFT
  8. GradientExplainer, SHapley Additive exPlanations (SHAP) + Integrated Gradients (IG) + SmoothGrad (SG)

Interactive Application

A streamlit application to interactively visualize the explanations and deletion evaluation metrics is available here.

The application is currently hosting all the models discussed in the thesis. However, due to limited resources, if the application is down, please E-mail me to reboot the server.

This interactive application can be run locally (recommended due to limited resources) by installing DExT repository and running the below command.

streamlit run app.py

References

[1] Google AutoML, "EfficientDet", GitHub, Available at: https://github.com/google/automl/tree/master/efficientdet, Accessed on: 01. 06. 2021.

[2] Octavio Arriaga, Matias Valdenegro-Toro, Mohandass Muthuraja, Sushma Devaramani, and Frank Kirchner. "Perception for Autonomous Systems (PAZ)." arXiv preprint arXiv:2010.14541. 2020.

[3] Waleed Abdulla, "Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow", GitHub, Available at: https://github.com/matterport/Mask_RCNN, Accessed on: 01.06. 2021.

[4] Maximilian Alber, Sebastian Lapuschkin, Philipp Seegerer, Miriam Hägele, Kristof T. Schütt, Grégoire Montavon, Wojciech Samek, Klaus-Robert Müller, Sven Dähne, and Pieter-Jan Kindermans. "iNNvestigate neural networks!." Journal of Machine Learning Research 20, No. 93, pp. 1-8. 2019.

[5] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ""Why should I trust you?" Explaining the predictions of any classifier." In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pp. 1135-1144. 2016.

[6] Scott M. Lundberg and Su-In Lee. "A unified approach to interpreting model predictions." In Proceedings of the 31st international conference on neural information processing systems, pp. 4768-4777. 2017.

[7] François Chollet, "Visualizing what convnets learn", Keras documentation, Available at: https://keras.io/examples/vision/visualizing_what_convnets_learn/, Accessed on: 03. 05. 2021.

[8] Jacob Gildenblat, "Visualizing CNN filters with keras", GitHub, Available at: https://jacobgil.github.io/deeplearning/filter-visualizations, Accessed on: 03. 05. 2021.

[9] Vincent Kao, "Visualize output of CNN filter by gradient ascent", Kaggle, Available at: https://www.kaggle.com/vincentman0403/visualize-output-of-cnn-filter-by-gradient-ascent, Accessed on: 03. 05. 2021.

[10] Sicara, "tf-explain", GitHub, Available at: https://github.com/sicara/tf-explain, Accessed on: 03. 05. 2021.

[11] Cristian Vasta, Deep Learning Models with Tensorflow 2.0, Available at: https://morioh.com/p/64064daff26c, Accessed on: 03. 05. 2021.

[12] Hoa Nguyen, "CNN Visualization Keras TF2", GitHub, Available at: https://github.com/nguyenhoa93/cnn-visualization-keras-tf2, Accessed on: 03. 05. 2021.

[13] People + AI Research, Google Research, "Saliency Methods", GitHub, Available at: https://github.com/PAIR-code/saliency, Accessed on: 01. 06. 2021.

[14] Marco Ancona, "DeepExplain: attribution methods for Deep Learning", GitHub, Available at: https://github.com/marcoancona/DeepExplain, Accessed on: 01. 06. 2021.

[15] Maximilian Alber, "iNNvestigate neural networks!", GitHub, Available at: https://github.com/albermax/innvestigate, Accessed on: 03. 05. 2021.

[16] Marco Tulio Correia, "lime", GitHub, Available at: https://github.com/marcotcr/lime, Accessed on: 01. 06. 2021.

[17] Scott M. Lundberg, "shap", GitHub, Available at: https://github.com/slundberg/shap, Accessed on: 01. 06. 2021.

[18] Mingxing Tan, Ruoming Pang, and Quoc V. Le. "Efficientdet: Scalable and efficient object detection." In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10781-10790. 2020.

[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. "Ssd: Single shot multibox detector." In European conference on computer vision, pp. 21-37. Springer, Cham, 2016.

[20] Heungsub Lee, "Elo, the classic rating system", GitHub, Available at: https://github.com/sublee/elo/, Accessed on: 13. 01. 2022.