src/utils/grad_cam.py

import cv2
import torch
import numpy as np

class InfoHolder():

    def __init__(self, heatmap_layer):
        self.gradient = None
        self.activation = None
        self.heatmap_layer = heatmap_layer

    def get_gradient(self, grad):
        self.gradient = grad

    def hook(self, model, input, output):
        output.register_hook(self.get_gradient)
        self.activation = output.detach()

def generate_heatmap(weighted_activation):
    raw_heatmap = torch.mean(weighted_activation, 0)
    heatmap = np.maximum(raw_heatmap.detach().cpu(), 0)
    heatmap /= torch.max(heatmap) + 1e-10
    return heatmap.numpy()

def superimpose(input_img, heatmap):
    img = cv2.cvtColor(input_img,cv2.COLOR_BGR2RGB)
    heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0]))
    heatmap = np.uint8(255 * heatmap)
    heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)
    superimposed_img = np.uint8(heatmap * 0.6 + img * 0.4)
    pil_img = cv2.cvtColor(superimposed_img,cv2.COLOR_BGR2RGB)
    return pil_img

def superimpose_to_mask(input_img, heatmap):
    img = cv2.cvtColor(input_img,cv2.COLOR_BGR2RGB)
    heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0]))
    heatmap = np.uint8(255 * heatmap)
    heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_BONE)
    superimposed_img = np.uint8(heatmap * 1 + img * 0.0)
    pil_img = cv2.cvtColor(heatmap,cv2.COLOR_BGR2GRAY)
    (thresh, blackAndWhiteImage) = cv2.threshold(pil_img, 127, 255, cv2.THRESH_BINARY)
    return blackAndWhiteImage


def to_RGB(tensor):
    tensor = (tensor - tensor.min())
    tensor = tensor/(tensor.max() + 1e-10)
    image_binary = np.transpose(tensor.cpu().numpy(), (1, 2, 0))
    image = np.uint8(255 * image_binary)
    return image

def grad_cam(model, input_tensor, heatmap_layer, truelabel=None):
    info = InfoHolder(heatmap_layer)
    heatmap_layer.register_forward_hook(info.hook)
    
    output = model(input_tensor.unsqueeze(0))[0]
    truelabel = truelabel if truelabel else torch.argmax(output)

    output[truelabel].backward()

    weights = torch.mean(info.gradient, [0, 2, 3])
    activation = info.activation.squeeze(0)

    weighted_activation = torch.zeros(activation.shape)
    for idx, (weight, activation) in enumerate(zip(weights, activation)):
        weighted_activation[idx] = weight * activation

    heatmap = generate_heatmap(weighted_activation)
    input_image = to_RGB(input_tensor)

    return superimpose(input_image, heatmap)
    
    
def save_grad_grad_cam(model, input_tensor, heatmap_layer, truelabel=None):
    info = InfoHolder(heatmap_layer)
    heatmap_layer.register_forward_hook(info.hook)
    
    output = model(input_tensor.unsqueeze(0))[0]
    truelabel = truelabel if truelabel else torch.argmax(output)

    output[truelabel].backward()

    weights = torch.mean(info.gradient, [0, 2, 3])
    activation = info.activation.squeeze(0)

    weighted_activation = torch.zeros(activation.shape)
    for idx, (weight, activation) in enumerate(zip(weights, activation)):
        weighted_activation[idx] = weight * activation

    heatmap = generate_heatmap(weighted_activation)
    input_image = to_RGB(input_tensor)
    grad_image = superimpose_to_mask(input_image, heatmap)
    return input_image, grad_image