A Simple TF1.15 version of Full-Gradient Saliency Maps . The project investigates the concept of using input_gradients and bias_gradients to plot saliency maps of conv models. This technique produces fine saliency maps when compared to ClassActivationMaps according to author , I will update with comparison in coming days.
.png)
https://github.com/idiap/fullgrad-saliency
@inproceedings{srinivas2019fullgrad,
title={Full-Gradient Representation for Neural Network Visualization},
author={Srinivas, Suraj and Fleuret, François},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2019}
}
'''
Tested on vgg , resnet, densenet, Xception
Feel free to use in custom models and other
architectures and report issues
'''
from tensorflow.keras.applications.resnet50 import ResNet50,preprocess_input
from tensorflow.keras import backend as K
import numpy as np
import os
from fullgradsaliency_tf.fullgrad import FullGrad
K.clear_session()
base_model=ResNet50(weights='imagenet')
fullgrad=FullGrad(base_model)
input_=np.ones(shape=(1,224,224,3))
preprocessed_input=preprocess_input(input_)
'''
check if completeness test is satisfied. Refer example.ipynb
'''
fullgrad.checkCompleteness(input_)
'''
now get saliency map of highest class from fullgrad model
'''
saliency=fullgrad.saliency(preprocessed_input)
saliency=fullgrad.postprocess_saliency_map(saliency[0])
'''
more detailed usage is available in example.ipynb
'''
Infer on various models and report any bugs.
Compare the results of model by editing or
removing the part of image having high
confidence on saliency heatmap.