def GetGraySmoothGrad(net, img:torch.Tensor, label:torch.Tensor, DEVICE):
grad_map = GetSmoothGrad(net, img, label, DEVICE, stdev_spread=0.10)
# 3x224x224
grad = GetSmoothGrad(net, img, label, DEVICE, stdev_spread = 0.0, num=1)
gray_grad = clip_gradmap_to_gray(grad_map)
grad = clip_gradmap_to_gray(grad)
#print(gray_grad.shape)
gray_grad = gray_grad[:, :, np.newaxis]
grad = grad[:, :, np.newaxis]
gray_grad = np.repeat(gray_grad, 3, axis=2) # (224, 224, 3)
grad = np.repeat(grad, 3, axis=2)
#print(gray_grad.shape)
return gray_grad, grad
def l1_for_without_smooth(net, dl, DEVICE):
net.eval()
net.to(DEVICE)
#criterion = nn.CrossEntropyLoss().to(DEVICE)
l1s = []
for i, (batch_img, batch_label) in enumerate(dl):
#if i> 5:
# break
for j in range(int(batch_img.size(0))):
img = batch_img[j]
label = batch_label[j]
img = img.to(DEVICE)
label = label.to(DEVICE)
#print(img.size())
grad_map = GetSmoothGrad(net,
img,
label,
DEVICE,
stdev_spread=0.05,
num=32)
#print(grad_maps.size(), batch_img.size())
l1s.append(torch.norm(grad_map, 1).item())
l1s = np.array(l1s)
print("Min: {:.4f} -- Max: {:.2f} -- Mean:{:.2f}".format(
l1s.min(), l1s.max(), l1s.mean()))
def GetSmoothRes(net, Data, DEVICE, save_path='./SmoothRes/Fashion_MNIST'):
for i, (img, label) in enumerate(zip(Data.X, Data.Y)):
#print(i)
#print(img.shape, label.shape)
img = img.astype(np.float32)
#label = label.astype(np.float32)
img = img[np.newaxis, :]
img = torch.tensor(img)
#print(img.type())
label = torch.tensor(label).type(torch.LongTensor)
grad_map = GetSmoothGrad(net, img, label, DEVICE=DEVICE)
grad_map = grad_map.cpu().detach().numpy()
grad_map = clip_gradmap(grad_map)
#print(grad_map.shape, grad_map.mean())
save_p = os.path.join(save_path, '{}.png'.format(i))
#print(grad_map.shape)
imwrite(save_p, grad_map)
print('{} imgs saved in {}'.format(i + 1, save_path))
def get_result(net, dl, DEVICE, net_name = ''):
save_bench = '../data/benchmark/'
save_path = os.path.join('../SmoothRes/', net_name)
labels = []
net.eval()
mean = torch.tensor(np.array([0.485, 0.456, 0.406]).astype(np.float32)[np.newaxis, :, np.newaxis, np.newaxis])
std = torch.tensor(np.array([0.229, 0.224, 0.225]).astype(np.float32)[np.newaxis, :, np.newaxis, np.newaxis])
mean = mean.to(DEVICE)
std = std.to(DEVICE)
for i, (batch_img, batch_label) in enumerate(dl):
if i> 5:
break
for j in range(int(batch_img.size(0))):
img = batch_img[j]
label = batch_label[j]
img = img.to(DEVICE)
label = label.to(DEVICE)
#print(img.size())
grad_map = GetSmoothGrad(net, img, label, DEVICE, stdev_spread = 0.05)
#print(grad_map.shape)
clip_and_save_single_img(grad_map, i * batch_img.size(0) + j, save_dir=save_path)
#print(grad.shape)
#simg = (img + mean) * std
simg = img * std + mean
#print('rb', simg.max(), simg.min())
simg = torch.clamp(simg, 0, 1)
#print('r', simg.max(), simg.min())
simg = simg.detach().cpu().numpy() * 255.0
#print(simg.shape)
#print(simg.shape)
simg = simg[0]
simg = np.transpose(simg, (1, 2, 0)).astype(np.uint8)
#print('r', simg.max(), simg.min())
#imwrite(os.path.join(save_bench, '{}.png'.format(i * batch_img.size(0) + j)), simg)
io.imsave(os.path.join(save_bench, '{}.png'.format(i * batch_img.size(0) + j)), simg)
print(i * batch_img.size(0) + j)
#grad = imread(os.path.join(save_path, '{}-smooth.png'.format(i * batch_img.size(0) + j)))
grad = io.imread(os.path.join(save_path, '{}-smooth.png'.format(i * batch_img.size(0) + j)),
as_gray = False)
# if gray
# grad = grad[:, :, np.newaxis]
# grad = np.repeat(grad, 3, axis = 2)
gray_grad = np.mean(grad, axis = -1, keepdims = True)
gray_grad = gray_grad.astype(np.uint8)
gray_grad = np.repeat(gray_grad, 3, axis = 2)
pair_img = np.concatenate((gray_grad, grad, simg), axis=1)
#imwrite(os.path.join(save_path, '{}-pair.png'.format(i * batch_img.size(0) + j)), pair_img)
io.imsave(os.path.join(save_path, '{}-pair.png'.format(i * batch_img.size(0) + j)), pair_img)
labels.append(batch_label.numpy())
labels = np.array(labels)
np.savetxt(os.path.join(save_bench, 'label.txt'), labels.reshape(-1))
def get_result(net, dl, DEVICE, net_name='', dl_name='raw'):
PgdAttack = IPGD(eps=8 / 255.0,
sigma=2 / 255.0,
nb_iter=40,
norm=np.inf,
DEVICE=DEVICE)
save_path = os.path.join('../Maps/', dl_name, net_name)
if not os.path.exists(save_path):
os.mkdir(save_path)
labels = []
net.eval()
mean = torch.tensor(
np.array([0.485, 0.456,
0.406]).astype(np.float32)[np.newaxis, :, np.newaxis,
np.newaxis])
std = torch.tensor(
np.array([0.229, 0.224,
0.225]).astype(np.float32)[np.newaxis, :, np.newaxis,
np.newaxis])
mean = mean.to(DEVICE)
std = std.to(DEVICE)
for i, (batch_img, batch_label) in enumerate(dl):
batch_img = batch_img.to(DEVICE)
batch_label = batch_label.to(DEVICE)
batch_img = PgdAttack.attack(net, batch_img, batch_label).detach()
if i > 5:
break
for j in range(int(batch_img.size(0))):
img = batch_img[j]
label = batch_label[j]
#img = img.to(DEVICE)
#label = label.to(DEVICE)
#print(img.size())
grad_map = GetSmoothGrad(net,
img,
label,
DEVICE,
stdev_spread=0.10)
#print(grad_map.shape)
clip_and_save_single_img(grad_map,
i * batch_img.size(0) + j,
save_dir=save_path)
#print(grad.shape)
#simg = (img + mean) * std
simg = img * std + mean
#print('rb', simg.max(), simg.min())
simg = torch.clamp(simg, 0, 1)
#print('r', simg.max(), simg.min())
simg = simg.detach().cpu().numpy() * 255.0
#print(simg.shape)
#print(simg.shape)
simg = simg[0]
simg = np.transpose(simg, (1, 2, 0)).astype(np.uint8)
#print('r', simg.max(), simg.min())
#imwrite(os.path.join(save_bench, '{}.png'.format(i * batch_img.size(0) + j)), simg)
#io.imsave(os.path.join(save_bench, '{}.png'.format(i * batch_img.size(0) + j)), simg)
#print(i * batch_img.size(0) + j)
#grad = imread(os.path.join(save_path, '{}-smooth.png'.format(i * batch_img.size(0) + j)))
grad = io.imread(os.path.join(
save_path, '{}-smooth.png'.format(i * batch_img.size(0) + j)),
as_gray=False)
# if gray
# grad = grad[:, :, np.newaxis]
# grad = np.repeat(grad, 3, axis = 2)
gray_grad = np.mean(grad, axis=-1, keepdims=True)
gray_grad = gray_grad.astype(np.uint8)
gray_grad = np.repeat(gray_grad, 3, axis=2)
pair_img = np.concatenate((gray_grad, grad, simg), axis=1)
#imwrite(os.path.join(save_path, '{}-pair.png'.format(i * batch_img.size(0) + j)), pair_img)
io.imsave(
os.path.join(save_path,
'{}-pair.png'.format(i * batch_img.size(0) + j)),
pair_img)