TejaGollapudi/PyTorch-CNN-Visualizations-Saliency-Maps

Pytorch implementation of convolutional neural network visualization techniques

Demo for visualizing CNNs using Guided_Grad_Gam and Grad_cam

Sivateja Gollapudi

Built on the work of utkuozbulak/pytorch-cnn-visualizations

vis_grad file contains model_compare function which is used to visualize guided_gradcam_back_prop and model_compare_cam perfroms grad_cam

from vis_grad import model_compare_cam , model_compare

import pretrained models using torch vision models (custom models can be used)

from torchvision import models

using 3 models , alex net , dense net 121 and resnet 152

md=models.alexnet(pretrained=True)
md2=models.densenet121(pretrained=True)
md3=models.resnet152(pretrained=True)
md4 = models.vgg16(pretrained=True)

input image size used by the network

size=[224,224]

create a list containing (model,'model name to print',[input image size,input image size]) for each model

list=[[md,'alexnet',size],[md2,'densenet121',size],[md3,'resnet152',size],[md4,'vgg',size]]

pass the list , class number , layer to visualize , input_image to visualize on

model_compare(list,56,6,'../input_images/snake.jpg')

Grad cam completed
Guided backpropagation completed
Guided grad cam completed
Grad cam completed
Guided backpropagation completed
Guided grad cam completed
Grad cam completed
Guided backpropagation completed
Guided grad cam completed
Grad cam completed
Guided backpropagation completed
Guided grad cam completed

Images are automatically saved in result folder

For visualizing grad_cam

model_compare_cam(list,56,10,'../input_images/snake.jpg')

Grad cam completed
Grad cam completed
Grad cam completed
Grad cam completed