def main(args):
# 输入
img = io.imread(args.image_path)
img = np.float32(cv2.resize(img, (224, 224))) / 255
inputs = prepare_input(img)
# 输出图像
image_dict = {}
# 网络
net = get_net(args.network, args.weight_path)
# Grad-CAM
layer_name = get_last_conv_name(
net) if args.layer_name is None else args.layer_name
grad_cam = GradCAM(net, layer_name)
mask = grad_cam(inputs, args.class_id) # cam mask
image_dict['cam'], image_dict['heatmap'] = gen_cam(img, mask)
grad_cam.remove_handlers()
# Grad-CAM++
grad_cam_plus_plus = GradCamPlusPlus(net, layer_name)
mask_plus_plus = grad_cam_plus_plus(inputs, args.class_id) # cam mask
image_dict['cam++'], image_dict['heatmap++'] = gen_cam(img, mask_plus_plus)
grad_cam_plus_plus.remove_handlers()
# GuidedBackPropagation
gbp = GuidedBackPropagation(net)
inputs.grad.zero_() # 梯度置零
grad = gbp(inputs)
gb = gen_gb(grad)
image_dict['gb'] = gb
# 生成Guided Grad-CAM
cam_gb = gb * mask[..., np.newaxis]
image_dict['cam_gb'] = norm_image(cam_gb)
save_image(image_dict, os.path.basename(args.image_path), args.network,
args.output_dir)
def grad_cam(fid, cam_model, net, im, im_size, class_idx=None, cuda=False):
im_shape = im.shape
#logger.info(f'grad_cam im_shape:{im_shape}')
#img = np.float32(cv2.resize(im, (im_size, im_size))) / 255
img = np.float32(im) / 255
inputs = prepare_input(img)
if cuda:
inputs = inputs.cuda()
#logger.info(f'grad_cam inputs:{type(inputs)}')
# 输出图像
image_dict = {}
layer_name = get_last_conv_name(net)
#logger.info(f'grad_cam layer_name:{layer_name}')
# Grad-CAM
if cam_model == 'GradCAM':
grad_cam = GradCAM(net, layer_name)
mask = grad_cam(inputs, class_idx) # cam mask
# mask = cv2.resize(mask, (im.shape[1], im.shape[0]), interpolation=cv2.INTER_CUBIC)
image_dict['GradCAM'], image_dict['heatmap'] = gen_cam(img, mask)
grad_cam.remove_handlers()
# Grad-CAM++
if cam_model == 'GradCAMpp':
grad_cam_plus_plus = GradCamPlusPlus(net, layer_name)
mask_plus_plus = grad_cam_plus_plus(inputs, class_idx) # cam mask
image_dict['GradCAMpp'], image_dict['heatmap++'] = gen_cam(
img, mask_plus_plus)
# cv2.imwrite(f'./output/GradCAMpp:{fid}.jpg', image_dict['GradCAMpp'])
# cv2.imwrite(f'./output/heatmappp:{fid}.jpg', image_dict['heatmap++'])
grad_cam_plus_plus.remove_handlers()
return image_dict[cam_model]
def heat_map(image,
tensors,
model,
class_id=None,
layer_name=None,
save_path='./heatmap/'):
"""
生成热值图
:param image: 图片原始numpy对象
:param tensors: 图像tensor对象（之所以单独读取，是因为数据增强方式会变化）
:param model: 模型对象
:param class_id: 识别类别索引，为None时去概率最大的类别
:param layer_name: 可视化的卷积层名称，请参考get_conv_name读取，默认最后一层
:param save_path: 结果保存路径，默认'./heatmap/'
:return:
"""
image_dict = {}
if layer_name is None:
layer_name = get_last_conv_name(model)
grad_cam = GradCAM(model, layer_name)
mask = grad_cam(tensors, class_id) # cam mask
image_dict['cam'], image_dict['heatmap'] = gen_cam(image, mask)
grad_cam.remove_handlers()
# Grad-CAM++
grad_cam_plus_plus = GradCamPlusPlus(model, layer_name)
mask_plus_plus = grad_cam_plus_plus(tensors, class_id) # cam mask
image_dict['cam++'], image_dict['heatmap++'] = gen_cam(
image, mask_plus_plus)
grad_cam_plus_plus.remove_handlers()
# GuidedBackPropagation
gbp = GuidedBackPropagation(model)
# image_wakan.zero_grad() # 梯度置零
grad = gbp(tensors)
gb = gen_gb(grad)
image_dict['gb'] = norm_image(gb)
# 生成Guided Grad-CAM
cam_gb = gb * mask[..., np.newaxis]
image_dict['cam_gb'] = norm_image(cam_gb)
if not os.path.exists(save_path):
os.mkdir(save_path)
save_image(image_dict, 'heatmap', save_path)
def get_attribution_maps_methods(net, layer_name, method_name):
'''
Get_attribution_maps_methods
method_name: include GradCAM, GradCAM++
'''
if method_name in ["GradCAM"]:
attr_maps = GradCAM(net, layer_name)
elif method_name in ["GradCAM++"]:
attr_maps = GradCamPlusPlus(net, layer_name)
return attr_maps
def CAM(net, img_list, date, input_type):
# img_list = glob.glob("./../eye_data/dataset/test/*")
correct = 0
for img_path in img_list:
file_name = img_path.split("/")[-1]
class_id = 1 if file_name[0] == 'n' else 2
class_id -= 1
img = Image.open(img_path)
transform_test = transforms.Compose([
transforms.Resize(size=(224, 224), interpolation=2),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
# img = transform_test(img)
# img = img[np.newaxis, ...] # 增加batch维
# img = Variable(img).cuda()
# output = net(img) # size(1, 2)
inputs = transform_test(img)
inputs = inputs[np.newaxis, ...] # 增加batch维
inputs = torch.tensor(inputs, requires_grad=True)
inputs = Variable(inputs, requires_grad=True).cuda()
inputs.retain_grad()
output = net(inputs)
img = io.imread(img_path)
img = np.float32(cv2.resize(img, (224, 224))) / 255
# print(output.data, class_id-1)
_, predicted = output.max(1)
# print(predicted[0].item())
if (predicted[0].item() == class_id - 1):
correct += 1
image_dict = {}
# Grad-CAM
layer_name = get_last_conv_name(net)
grad_cam = GradCAM(net, layer_name)
mask = grad_cam(inputs, class_id) # cam mask
image_dict['cam'], image_dict['heatmap'] = gen_cam(img, mask)
grad_cam.remove_handlers()
# Grad-CAM++
grad_cam_plus_plus = GradCamPlusPlus(net, layer_name)
mask_plus_plus = grad_cam_plus_plus(inputs, class_id) # cam mask
image_dict['cam++'], image_dict['heatmap++'] = gen_cam(
img, mask_plus_plus)
grad_cam_plus_plus.remove_handlers()
# GuidedBackPropagation
gbp = GuidedBackPropagation(net)
inputs.grad.zero_() # 梯度置零
grad = gbp(inputs)
gb = gen_gb(grad)
image_dict['gb'] = norm_image(gb)
# 生成Guided Grad-CAM
cam_gb = gb * mask[..., np.newaxis]
image_dict['cam_gb'] = norm_image(cam_gb)
save_image(image_dict, file_name, "EfficientNet-B0",
f"./grad_cam_result/{input_type}_{date}")
print("Test the accuracy on Grad-CAM:", correct / len(img_list))