def _california(reduce=False):
""" Load California dataset from .mat """
mat = loadmat("data/California/UiT_HCD_California_2017.mat")
t1 = np.array(mat["t1_L8_clipped"], dtype=np.float32)
t2 = np.array(mat["logt2_clipped"], dtype=np.float32)
change_mask = tf.convert_to_tensor(mat["ROI"], dtype=tf.bool)
assert t1.shape[:2] == t2.shape[:2] == change_mask.shape[:2]
if change_mask.ndim == 2:
change_mask = change_mask[..., np.newaxis]
if reduce:
print("Reducing")
reduction_ratios = (4, 4)
new_dims = list(
map(lambda a, b: a // b, change_mask.shape, reduction_ratios))
t1 = tf.cast(tf.image.resize(t1, new_dims, antialias=True),
dtype=tf.float32)
t2 = tf.cast(tf.image.resize(t2, new_dims, antialias=True),
dtype=tf.float32)
change_mask = tf.cast(
tf.image.resize(tf.cast(change_mask, tf.uint8),
new_dims,
antialias=True),
tf.bool,
)
t1, t2, change_mask = (
remove_borders(t1, 2),
remove_borders(t2, 2),
remove_borders(change_mask, 2),
)
return t1, t2, change_mask
def _france(reduce=True):
""" Load France dataset from .mat """
mat = loadmat("data/France/France.mat")
t1 = np.array(mat["t1"], dtype=np.single)
t2 = np.array(mat["t2"], dtype=np.single)
t1, t2 = _clip(t1), _clip(t2)
change_mask = tf.convert_to_tensor(mat["ROI"], dtype=tf.bool)
assert t1.shape[:2] == t2.shape[:2] == change_mask.shape[:2]
if change_mask.ndim == 2:
change_mask = change_mask[..., np.newaxis]
if reduce:
print("Reducing")
reduction_ratios = (5, 5)
new_dims = list(
map(lambda a, b: a // b, change_mask.shape, reduction_ratios))
t1 = tf.cast(tf.image.resize(t1, new_dims, antialias=True),
dtype=tf.float32)
t2 = tf.cast(tf.image.resize(t2, new_dims, antialias=True),
dtype=tf.float32)
change_mask = tf.cast(
tf.image.resize(tf.cast(change_mask, tf.uint8),
new_dims,
antialias=True),
tf.bool,
)
t1, t2, change_mask = (
remove_borders(t1, 2),
remove_borders(t2, 2),
remove_borders(change_mask, 2),
)
return t1, t2, change_mask
def fetch_CGAN(name, **kwargs):
"""
Input:
name - dataset name, should be in DATASETS
kwargs - config {key: value} pairs.
Key should be in DATASET_DEFAULT_CONFIG
Output:
training_data - tf.data.Dataset with (x, y, prior)
shapes like (inf, patch_size, patch_size, ?)
evaluation_data - tf.data.Dataset with (x, y, change_map)
shapes (1, h, w, ?)
channels - tuple (c_x, c_y), number of channels for domains x and y
"""
ps = kwargs.get("patch_size")
y_im, x_im, target_cm = DATASETS[name](prepare_data[name])
if not tf.config.list_physical_devices("GPU"):
dataset = [
tf.image.central_crop(tensor, 0.1)
for tensor in [x_im, y_im, target_cm]
]
else:
dataset = [x_im, y_im, target_cm]
chs = [tensor.shape[-1] for tensor in dataset]
dataset = [remove_borders(tensor, ps) for tensor in dataset]
dataset = [tf.expand_dims(tensor, 0) for tensor in dataset]
evaluation_data = tf.data.Dataset.from_tensor_slices(tuple(dataset))
dataset = [image_in_patches(tensor, ps) for tensor in dataset]
tot_patches = dataset[0].shape[0]
return dataset[0], dataset[1], evaluation_data, (chs[0],
chs[1]), tot_patches
def _italy(reduce=False):
""" Load Italy dataset from .mat """
mat = loadmat("data/Italy/Italy.mat")
t1 = np.array(mat["t1"], dtype=np.single)
t2 = np.array(mat["t2"], dtype=np.single)
change_mask = np.array(mat["ROI"], dtype=np.bool)
if t1.shape[-1] == 3:
t1 = t1[..., 0]
t1, t2, change_mask = (
remove_borders(t1, 2),
remove_borders(t2, 2),
remove_borders(change_mask, 2),
)
t1, t2 = _clip(t1[..., np.newaxis]), _clip(t2)
change_mask = tf.convert_to_tensor(change_mask, dtype=tf.bool)
assert t1.shape[:2] == t2.shape[:2] == change_mask.shape[:2]
if change_mask.ndim == 2:
change_mask = change_mask[..., np.newaxis]
change_mask = change_mask[..., :1]
return t1, t2, change_mask
def _texas(clip=True):
""" Load Texas dataset from .mat """
mat = loadmat("data/Texas/Cross-sensor-Bastrop-data.mat")
t1 = np.array(mat["t1_L5"], dtype=np.single)
t2 = np.array(mat["t2_ALI"], dtype=np.single)
if clip:
print("clipping")
t1, t2 = _clip(t1), _clip(t2)
change_mask = tf.convert_to_tensor(mat["ROI_1"], dtype=tf.bool)
assert t1.shape[:2] == t2.shape[:2] == change_mask.shape[:2]
if change_mask.ndim == 2:
change_mask = change_mask[..., np.newaxis]
t1, t2, change_mask = (
remove_borders(t1, 2),
remove_borders(t2, 2),
remove_borders(change_mask, 2),
)
return t1, t2, change_mask