def lrp_visualize(R_lrp, gamma=args.gamma):
heatmaps_lrp = []
for h, heat in enumerate(R_lrp):
heat = heat.permute(1, 2, 0).detach().cpu().numpy()
maps = render.heatmap(heat, reduce_axis=-1, gamma_=gamma)
heatmaps_lrp.append(maps)
return heatmaps_lrp
def visualize3(R_lrp, R_lrp34, R_grad, image_tensor, epoch, j, prediction):
try:
os.makedirs(args.img_dir + args.img_name + str('/')) ##
except OSError:
pass
heatmaps_lrp = []
heatmaps_lrp34 = []
heatmaps_grad = []
# 0. input images
input_shape = image_tensor.shape
images = image_tensor.permute(0, 2, 3, 1).cpu().detach().numpy()
images = images - images.min(axis=(1, 2, 3), keepdims=True)
images = images / images.max(axis=(1, 2, 3), keepdims=True)
# images = images - images.min()
# images = images / images.max()
# 1. LRP
for h, heat in enumerate(R_lrp):
heat = heat.permute(1, 2, 0).detach().cpu().numpy()
maps = render.heatmap(heat, reduce_axis=-1)
heatmaps_lrp.append(maps)
# 2. LRP34
R_lrp34 = R_lrp34.squeeze(1).permute(1, 2, 0)
R_lrp34 = R_lrp34.detach().cpu().detach().numpy()
R_lrp34 = cv2.resize(R_lrp34, (224, 224))
R_lrp34.reshape(input_shape[2], input_shape[3], input_shape[0])
for i in range(input_shape[0]):
heatmap = np.float32(
cv2.applyColorMap(np.uint8(
(1 - R_lrp34[:, :, i]) * 255), cv2.COLORMAP_JET)) / 255
cam = heatmap + np.float32(images[i])
cam = cam / np.max(cam)
heatmaps_lrp34.append(cam)
# 3. Grad_CAM
R_grad = R_grad.squeeze(1).permute(1, 2, 0)
R_grad = R_grad.cpu().detach().numpy()
R_grad = cv2.resize(R_grad, (224, 224))
R_grad.reshape(input_shape[2], input_shape[3], input_shape[0])
for i in range(input_shape[0]):
heatmap = np.float32(
cv2.applyColorMap(np.uint8(
(1 - R_grad[:, :, i]) * 255), cv2.COLORMAP_JET)) / 255
cam = heatmap + np.float32(images[i])
cam = cam / np.max(cam)
heatmaps_grad.append(cam)
R_lrp = np.array(heatmaps_lrp, dtype=np.float32)
R_lrp34 = np.array(heatmaps_lrp34, dtype=np.float32)
R_grad = np.array(heatmaps_grad, dtype=np.float32)
image_path = args.img_dir
img_name = args.img_name
prediction_dic = {}
l = min(args.batch_size_test, args.num_visualize_plot)
for i in range(l):
file_name = 'lrp_epoch' + str(epoch) + '_no-' + str(i + l * j) + '.png'
prediction_dic['lrp_epoch' + str(epoch) + '_no-' +
str(i + l * j)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, R_lrp[i], plugin='pil')
for i in range(l):
file_name = 'lrp34_epoch' + str(epoch) + '_no-' + str(i +
l * j) + '.png'
prediction_dic['lrp34_epoch' + str(epoch) + '_no-' +
str(i + l * j)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, R_lrp34[i], plugin='pil')
for i in range(l):
file_name = 'grad_epoch' + str(epoch) + '_no-' + str(i +
l * j) + '.png'
prediction_dic['grad_epoch' + str(epoch) + '_no-' +
str(i + l * j)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, R_grad[i], plugin='pil')
for i in range(l):
file_name = 'ori_epoch' + str(epoch) + '_no-' + str(i + l * j) + '.png'
prediction_dic['ori_epoch' + str(epoch) + '_no-' +
str(i + l * j)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, images[i], plugin='pil')
prediction_dic = sorted(prediction_dic.items(), key=operator.itemgetter(0))
np.save(
args.img_dir + args.img_name + str('/') + args.img_name +
'_prediction.npy', prediction_dic)
def visualize2(relevances, epoch, prediction, images_tensor=None, shape=None):
try:
os.makedirs(args.img_dir + args.img_name + str('/')) ##
except OSError:
pass
print('relevance shape', relevances.shape)
n, dim, h, w = relevances.shape
heatmaps = []
oris = []
# import pdb;pdb.set_trace()
if args.model in ['CNN8', 'VGG19']:
images_tensor = images_tensor.reshape(
[args.batch_size_test, 3, 224, 224])
else:
images_tensor = images_tensor.reshape(
[args.batch_size_test, 1, 28, 28])
if images_tensor is not None:
assert relevances.shape == images_tensor.shape, 'Relevances shape != Images shape'
# heatmaps = render.heatmap(relevances.cpu().numpy(),reduce_axis=1)
# oris = render.image(images_tensor.cpu().numpy())
# ori = oris - oris.min(axis=0)
# ori = ori / ori.max(axis=0)
for h, heat in enumerate(relevances):
if images_tensor is not None:
print(2, heat.shape)
input_image = images_tensor[h].permute(1, 2,
0).detach().cpu().numpy()
heat = heat.permute(1, 2, 0).detach().cpu().numpy()
#maps = render.hm_to_rgb(heat, input_image, scaling=3, sigma=2, shape=shape)
maps = render.heatmap(heat, reduce_axis=-1)
#ori = render.image(input_image)
ori = input_image - input_image.min()
ori = ori / ori.max()
else:
heat = heat.cpu().numpy()
#maps = render.hm_to_rgb(heat, scaling=3, sigma=2, shape=shape)
maps = render.heatmap(heat, reduce_axis=0)
#ori = render.image(input_image)
ori = input_image - input_image.min()
ori = ori / ori.max()
heatmaps.append(maps)
oris.append(ori)
R = np.array(heatmaps, dtype=np.float32)
#R = np.transpose(R,(0,3,2,1))
ori = np.array(oris, dtype=np.float32)
#ori= np.transpose(ori,(0,3,2,1))
# save R to img file
# print('---- R shape---',R.shape)
# print('--- img saved in utils.py in def visualize. change file name every time----')
# imsave('../results/test/180225_lrp_module.png', R[1])
# np.save('../results/test/180225_R.npy',R)
image_path = args.img_dir
img_name = args.img_name
prediction_dic = {}
for i in range(min(args.batch_size_test, args.num_visualize_plot)):
#path = os.path.join(image_path, 'results{}.png'.format(i))
#file_name = img_name + str(i) + '.png'
file_name = 'lrp_epoch' + str(epoch) + '_no-' + str(i) + '.png'
prediction_dic['lrp_epoch' + str(epoch) + '_no-' +
str(i)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, R[i], plugin='pil')
for i in range(min(args.batch_size_test, args.num_visualize_plot)):
#path = os.path.join(image_path, 'results{}.png'.format(i))
#file_name = img_name + str(i) + '.png'
file_name = 'ori_epoch' + str(epoch) + '_no-' + str(i) + '.png'
prediction_dic['ori_epoch' + str(epoch) + '_no-' +
str(i)] = prediction[i]
path = os.path.join(image_path + img_name + str('/'), file_name) ##
imsave(path, ori[i], plugin='pil')
prediction_dic = sorted(prediction_dic.items(), key=operator.itemgetter(0))
np.save(
args.img_dir + args.img_name + str('/') + args.img_name +
'_prediction.npy', prediction_dic)