dsanno/chainer-grad-cam

Experimental Chainer implementation of Grad-CAM.

Chainer Grad-CAM

Experimental Chainer implementation of Grad-CAM without modifying network structure.

Caution

This implementation depends on Chainer internal implementation, so it may not work on later version of Chainer.

Requirements

• Chainer 3.0
• OpenCV

Sample Usage

import package

from lib import gradcam

Make Chainer neural network model

from chainercv.links import VGG16

model = VGG16(pretrained_model='imagenet')

Load image and make input array

image = cv2.imread(image_path)
image_size = (224, 224)
image = cv2.resize(image, image_size)

x = np.float32(image)
x = x.transpose((2, 0, 1))[np.newaxis,::-1,:,:]
x -= model.mean

Make gradient for model output

y_grad = np.zeros((1, 1000), dtype=np.float32)
y_grad[0, args.label] = 1.0

Calculate Grad-CAM

gcam = gradcam.gradcam(model, x, [model.conv5_3.conv, F.relu], y_grad=y_grad)

Make Grad-CAM heatmap

gcam = gcam[0]
heatmap_image = gradcam.heatmap(gcam, image_size)
cv2.imwrite(heatmap_path, heatmap_image)

Make Grad-CAM heatmap over input image

overlay_image = gradcam.overlay(image, gcam)
cv2.imwrite(overlay_path, overlay_image)

Sample program usage

$ python src/vgg16.py image_path heatmap_path overlay_path label [-g gpu_device]

Parameters

• image_path: (Required) Input image file path
• heatmap_path: (Required) File path for heatmap image
• overlay_path: (Required) File path for heatmap over input image
• label: (Required) Class label index
• gpu_device: (Optional) GPU device ID, negative value indicates CPU (default: -1)

Example

$ python src/vgg16.py image/cat_dog.jpg cat_dog_heatmap.jpg cat_dog_overlay.jpg 242

Input image	Heatmap	Heatmap over input image