def _process_data(self, kspace_data, gt_data):
mask = self.mask_func(kspace_data.shape).expand_as(kspace_data)
kspace_data = unprep_fft_channel(kspace_data.unsqueeze(0))
mask = unprep_fft_channel(mask.unsqueeze(0))
kspace_masked = kspace_data * mask
gt_data = to_complex(gt_data)
if self.keep_mask_as_func:
mask = self.mask_func
return kspace_masked, mask, gt_data
def text_perturbation(shape):
# create empty image
img = Image.new("P", shape)
# adapt font choice depending on font availability on your system
system_fonts = font_manager.findSystemFonts(fontpaths=None, fontext="ttf")
filtered_fonts = [
font for font in system_fonts if "DejaVuSansMono.ttf" in font
]
# draw text on image
draw = ImageDraw.Draw(img)
font = ImageFont.truetype(filtered_fonts[0], size=16)
draw.text(
(0.27 * shape[0], 0.350 * shape[1]),
"I",
255,
font=font,
stroke_width=0,
)
font = ImageFont.truetype(filtered_fonts[0], size=13)
draw.text(
(0.30 * shape[0], 0.360 * shape[1]),
"E",
255,
font=font,
stroke_width=0,
)
font = ImageFont.truetype(filtered_fonts[0], size=10)
draw.text(
(0.33 * shape[0], 0.3725 * shape[1]),
"E",
255,
font=font,
stroke_width=0,
)
font = ImageFont.truetype(filtered_fonts[0], size=7)
draw.text(
(0.355 * shape[0], 0.385 * shape[1]),
"E",
255,
font=font,
stroke_width=0,
)
t_img = (torch.tensor(
np.array(img.getdata()).astype(np.float32).reshape(shape)) / 255)
return to_complex(t_img.unsqueeze(0)).unsqueeze(0).to(device)
Flatten(0, -3),
Normalize(reduction="mean", use_target=True),
], ),
}
test_data = AlmostFixedMaskDataset
test_data = test_data("val", **test_data_params)
vols = range(30)
slices_in_vols = [test_data.get_slices_in_volume(vol_idx) for vol_idx in vols]
slices_selected = [
range((lo + hi) // 2, (lo + hi) // 2 + 1) for lo, hi in slices_in_vols
]
samples = np.concatenate(slices_selected)
X_0 = torch.stack([test_data[s][2] for s in samples])
X_0 = to_complex(X_0.to(device))
Y_0 = torch.stack([test_data[s][0] for s in samples])
Y_0 = im2vec(to_complex(Y_0.to(device)))[..., im2vec(mask[0, 0, :, :].bool())]
# ----- noise setup --------
noise_min = 1e-3
noise_max = 0.03
noise_steps = 50
noise_rel = torch.tensor(
np.logspace(np.log10(noise_min), np.log10(noise_max),
num=noise_steps)).float()
noise_rel = (torch.cat([
torch.zeros(1).float(),
noise_rel,
0.060 * torch.ones(1).float(),
Flatten(0, -3),
Normalize(reduction="mean", use_target=True),
], ),
}
test_data = AlmostFixedMaskDataset
test_data = test_data("val", **test_data_params)
vols = range(30)
slices_in_vols = [test_data.get_slices_in_volume(vol_idx) for vol_idx in vols]
slices_selected = [
range((lo + hi) // 2, (lo + hi) // 2 + 1) for lo, hi in slices_in_vols
]
samples = np.concatenate(slices_selected)
X_0 = torch.stack([test_data[s][2] for s in samples])
X_0 = to_complex(X_0.to(device))
Y_0 = torch.stack([test_data[s][0] for s in samples])
Y_0 = im2vec(to_complex(Y_0.to(device)))[...,
im2vec(cfg_rob.mask[0,
0, :, :].bool())]
# no meas samples
it_init = 6
keep_init = 3
noise_type = noise_gaussian
# select range relative noise
noise_rel = torch.tensor([0.00, 0.0025, 0.005, 0.01, 0.015, 0.02, 0.025])
True,
"transform":
torchvision.transforms.Compose([
CropOrPadAndResimulate((320, 320)),
Flatten(0, -3),
Normalize(reduction="mean", use_target=True),
], ),
}
test_data = AlmostFixedMaskDataset
test_data = test_data("val", **test_data_params)
lo, hi = test_data.get_slices_in_volume(sample_vol)
print("volume slices from {} to {}, selected {}".format(
lo, hi, lo + sample_sl))
X_VOL = to_complex(
torch.stack([test_data[sl_idx][2] for sl_idx in range(lo, hi)],
dim=0)).to(device)
X_MAX = rotate_real(X_VOL)[:, 0:1, ...].max().cpu()
X_0 = to_complex(test_data[lo + sample_sl][2].to(device)).unsqueeze(0)
X_0 = X_0.repeat(it_init, *((X_0.ndim - 1) * (1, )))
Y_0 = cfg_rob.OpA(X_0)
# set range for plotting and similarity indices
v_min = 0.05
v_max = 4.50
print("Pixel values between {} and {}".format(v_min, v_max))
# create result table and load existing results from file
results = pd.DataFrame(columns=[
"name",
"X_adv_err",
# select measure for reconstruction error
err_measure = err_measure_l2
# select reconstruction methods
methods_include = ["L1", "UNet jit", "Tiramisu EE jit", "UNet It jit"]
methods = methods.loc[methods_include]
# select methods excluded from (re-)performing attacks
methods_no_calc = ["L1", "UNet jit", "Tiramisu EE jit", "UNet It jit"]
# ----- perform attack -----
# select sample
test_data = IPDataset("test", config.DATA_PATH)
X_0 = test_data[sample][0]
X_0 = to_complex(X_0.to(device)).unsqueeze(0)
X_0 = X_0.repeat(it, *((X_0.ndim - 1) * (1, )))
Y_0 = cfg_rob.OpA(X_0)
# create result table and load existing results from file
results = pd.DataFrame(columns=["name", "X_err", "X_psnr", "X_ssim", "X", "Y"])
results.name = methods.index
results = results.set_index("name")
# load existing results from file
if os.path.isfile(save_results):
results_save = pd.read_pickle(save_results)
for idx in results_save.index:
if idx in results.index:
results.loc[idx] = results_save.loc[idx]
else:
results_save = results
# select methods excluded from (re-)performing attacks
methods_no_calc = [
"L1",
"UNet jit",
"UNet EE jit",
"Tiramisu jit",
"Tiramisu EE jit",
"UNet It jit",
]
# ----- perform attack -----
# select samples
test_data = IPDataset("test", config.DATA_PATH)
X_0 = torch.stack([test_data[s][0] for s in samples])
X_0 = to_complex(X_0.to(device))
Y_0 = cfg_rob.OpA(X_0)
# create result table
results = pd.DataFrame(
columns=[
"name",
"X_adv_err",
"X_ref_err",
"X_adv_psnr",
"X_ref_psnr",
"X_adv_ssim",
"X_ref_ssim",
]
)
results.name = methods.index
def __call__(self, imgs):
return tuple([to_complex(img) for img in imgs])
def square_perturbation(shape):
t_img = torch.zeros(shape)
x, y = int(0.4675 * shape[0]), int(0.39 * shape[1])
t_img[(x - 1):(x + 2), (y - 1):(y + 2)] = 1
return to_complex(t_img.unsqueeze(0)).unsqueeze(0).to(device)
