def aligned_dataset(opt):
opt = opt
root = opt.dataroot
### input A (label maps)
dir_A = '_A' if opt.label_nc == 0 else '_label'
dir_A = os.path.join(opt.dataroot, opt.phase + dir_A)
A_paths = sorted(make_dataset(dir_A))
### input B (real images)
if opt.isTrain or opt.use_encoded_image:
dir_B = '_B' if opt.label_nc == 0 else '_img'
dir_B = os.path.join(opt.dataroot, opt.phase + dir_B)
B_paths = sorted(make_dataset(dir_B))
assert len(A_paths) == len(B_paths)
A_dataset = tl.disk_image_batch_dataset(
A_paths,
opt.batchSize,
drop_remainder=not opt.no_drop_remainder,
map_fn=preprocess_image_A,
shuffle=not opt.no_shuffle,
repeat=opt.repeat_num)
B_dataset = tl.disk_image_batch_dataset(
B_paths,
opt.batchSize,
drop_remainder=not opt.no_drop_remainder,
map_fn=preprocess_image_B,
shuffle=not opt.no_shuffle,
repeat=opt.repeat_num)
A_B_dataset = tf.data.Dataset.zip((A_dataset, B_dataset))
len_dataset = len(A_paths) // opt.batchSize
return A_B_dataset, len_dataset
def make_dataset(img_paths, batch_size, load_size, crop_size, training, drop_remainder=True, shuffle=True, repeat=1):
if training:
@tf.function
def _map_fn(img): # preprocessing
img = tf.image.random_flip_left_right(img)
img = tf.image.random_flip_up_down(img)
# img = tf.image.resize(img, [load_size, load_size])
img = tf.image.random_crop(img, [crop_size, crop_size, tf.shape(img)[-1]])
img = tf.clip_by_value(tf.cast(img, tf.float32), 0, 255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return img
else:
@tf.function
def _map_fn(img): # preprocessing
# img = tf.image.resize(img, [crop_size, crop_size]) # or
img = tf.image.resize(img, [load_size, load_size]);
img = tl.center_crop(img, load_size)
img = tf.clip_by_value(tf.cast(img, tf.float32), 0, 255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return img
return tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
def make_custom_datset(img_paths, batch_size, resize=64, drop_remainder=True, shuffle=True, repeat=1):
@tf.function
def _map_fn(img):
# ======================================
# = custom =
# ======================================
img = ... # custom preprocessings, should output img in [0.0, 255.0]
# ======================================
# = custom =
# ======================================
img = tf.image.resize(img, [resize, resize])
img = tf.clip_by_value(img, 0, 255)
img = img / 127.5 - 1
return img
dataset = tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
img_shape = (resize, resize, 3)
len_dataset = len(img_paths) // batch_size
return dataset, img_shape, len_dataset
def make_celeba_dataset(img_paths,
batch_size,
resize=64,
drop_remainder=True,
shuffle=True,
repeat=1):
@tf.function
def _map_fn(img):
crop_size = 108
img = tf.image.crop_to_bounding_box(img, (218 - crop_size) // 2,
(178 - crop_size) // 2, crop_size,
crop_size)
img = tf.image.resize(img, [resize, resize])
img = tf.clip_by_value(img, 0, 255)
img = img / 127.5 - 1
return img
dataset = tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
img_shape = (resize, resize, 3)
len_dataset = len(img_paths) // batch_size
return dataset, img_shape, len_dataset
def make_dataset(img_paths,
batch_size,
load_size,
crop_size,
training,
drop_remainder=True,
shuffle=True,
repeat=1):
_map_fn = img_preprocessing_fn(load_size, crop_size, training)
return tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
def make_PETCT_dataset(img_paths, batch_size, resize=128, drop_remainder=True, shuffle=True, repeat=1):
@tf.function
def _map_fn(img):
img = tf.image.resize(img, [resize, resize])
img = tf.clip_by_value(img, 0, 255)
img = img / 127.5 - 1
return img
dataset = tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
img_shape = (resize, resize, 3)
len_dataset = len(img_paths) // batch_size
return dataset, img_shape, len_dataset
def make_dataset(img_paths,
batch_size,
load_size,
crop_size,
training,
drop_remainder=True,
grayscale=False,
shuffle=False,
repeat=1):
# if training:
# @tf.function
# def _map_fn(img): # preprocessing
# #toss = np.random.uniform(0,1)
# if grayscale:
# img = tf.image.rgb_to_grayscale(img)
# img = tf.image.grayscale_to_rgb(img)
# img = tf.image.random_flip_left_right(img)
# img = tf.image.resize_with_pad(img, load_size, load_size, antialias = True)
# img = tf.image.random_crop(img, [crop_size, crop_size, tf.shape(img)[-1]])
# img = tf.clip_by_value(img, 0, 255) / 255.0 # or img = tl.minmax_norm(img)
# img = img * 2 - 1
# return img
# else:
# @tf.function
# def _map_fn(img): # preprocessing
# img = tf.image.resize_with_pad(img,crop_size, crop_size, antialias = True) # or img = tf.image.resize(img, [load_size, load_size]); img = tl.center_crop(img, crop_size)
# if grayscale:
# img = tf.image.rgb_to_grayscale(img)
# img = tf.image.grayscale_to_rgb(img)
# img = tf.clip_by_value(img, 0, 255) / 255.0 # or img = tl.minmax_norm(img)
# img = img * 2 - 1
# return img
if training:
@tf.function
def _map_fn(img): # preprocessing
#toss = np.random.uniform(0,1)
if grayscale:
img = tf.image.rgb_to_grayscale(img)
img = tf.image.grayscale_to_rgb(img)
img = tf.image.random_flip_left_right(img)
maxside = tf.math.maximum(tf.shape(img)[0], tf.shape(img)[1])
minside = tf.math.minimum(tf.shape(img)[0], tf.shape(img)[1])
new_img = img
if tf.math.divide(maxside, minside) > 1.2:
repeating = tf.math.floor(tf.math.divide(maxside, minside))
new_img = img
if tf.math.equal(tf.shape(img)[1], minside):
for i in range(int(repeating)):
new_img = tf.concat((new_img, img), axis=1)
if tf.math.equal(tf.shape(img)[0], minside):
for i in range(int(repeating)):
new_img = tf.concat((new_img, img), axis=0)
new_img = tf.image.rot90(new_img)
else:
new_img = img
img = tf.image.resize(new_img, [crop_size, crop_size])
#im.imwrite(img.numpy(),'test.jpg')
#img = tf.image.central_crop(img, [crop_size, crop_size, tf.shape(img)[-1]])
img = tf.clip_by_value(img, 0,
255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return img
else:
@tf.function
def _map_fn(img): # preprocessing
maxside = tf.math.maximum(tf.shape(img)[0], tf.shape(img)[1])
minside = tf.math.minimum(tf.shape(img)[0], tf.shape(img)[1])
new_img = img
if tf.math.divide(maxside, minside) > 1.3:
repeating = tf.math.floor(tf.math.divide(maxside, minside))
new_img = img
if tf.math.equal(tf.shape(img)[1], minside):
for i in range(int(repeating)):
new_img = tf.concat((new_img, img), axis=1)
if tf.math.equal(tf.shape(img)[0], minside):
for i in range(int(repeating)):
new_img = tf.concat((new_img, img), axis=0)
new_img = tf.image.rot90(new_img)
else:
new_img = img
img = tf.image.resize(new_img, [crop_size, crop_size])
#padx = load_size - tf.shape(img)[0]
#pady = load_size -tf.shape(img)[1]
#paddings = [[padx/2,padx/2],[pady/2,pady/2],[0, 0]]
#img = tf.pad(img,paddings,'SYMMETRIC')
#img = tf.image.resize_with_pad(img,crop_size, crop_size, antialias = True) # or img = tf.image.resize(img, [load_size, load_size]); img = tl.center_crop(img, crop_size)
if grayscale:
img = tf.image.rgb_to_grayscale(img)
img = tf.image.grayscale_to_rgb(img)
#img = tf.image.random_crop(img, [crop_size, crop_size, tf.shape(img)[-1]])
img = tf.clip_by_value(img, 0,
255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return img
return tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
def make_dataset2(img_paths,
labels,
batch_size,
load_size,
crop_size,
training,
drop_remainder=True,
grayscale=False,
shuffle=False,
repeat=1):
if training:
@tf.function
def _map_fn(img, label): # preprocessing
#toss = np.random.uniform(0,1)
if grayscale:
img = tf.image.rgb_to_grayscale(img)
img = tf.image.grayscale_to_rgb(img)
img = tf.image.random_flip_left_right(img)
maxside = tf.math.maximum(tf.shape(img)[0], tf.shape(img)[1])
while tf.math.square(tf.shape(img)[0] - tf.shape(img)[1]) > 100:
padx = tf.math.minimum(
maxside - tf.shape(img)[0],
tf.math.minimum(tf.shape(img)[0],
tf.shape(img)[1]))
pady = tf.math.minimum(
maxside - tf.shape(img)[1],
tf.math.minimum(tf.shape(img)[0],
tf.shape(img)[1]))
paddings = [[padx / 2, padx / 2], [pady / 2, pady / 2], [0, 0]]
img = tf.pad(
img, paddings, 'SYMMETRIC'
) #tf.image.resize_with_pad(img, load_size, load_size, antialias = True)
img = tf.image.resize(img, [load_size * +10, load_size + 10],
preserve_aspect_ratio=True)
img = tf.image.random_crop(
img,
[crop_size, crop_size, tf.shape(img)[-1]])
img = tf.clip_by_value(img, 0,
255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return [img, label]
else:
@tf.function
def _map_fn(img, label): # preprocessing
img = _aspect_preserving_resize(
img,
load_size + 4) # tf.image.resize(img, [load_size,load_size])
#padx = load_size - tf.shape(img)[0]
#pady = load_size -tf.shape(img)[1]
#paddings = [[padx/2,padx/2],[pady/2,pady/2],[0, 0]]
#img = tf.pad(img,paddings,'SYMMETRIC')
#img = tf.image.resize_with_pad(img,crop_size, crop_size, antialias = True) # or img = tf.image.resize(img, [load_size, load_size]); img = tl.center_crop(img, crop_size)
if grayscale:
img = tf.image.rgb_to_grayscale(img)
img = tf.image.grayscale_to_rgb(img)
img = tf.image.random_crop(
img,
[crop_size, crop_size, tf.shape(img)[-1]])
img = tf.clip_by_value(img, 0,
255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return [img, label]
return tl.disk_image_batch_dataset(img_paths,
batch_size,
labels=labels,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)
def make_dataset(img_paths,
batch_size,
load_size,
crop_size,
training,
drop_remainder=True,
shuffle=True,
repeat=1,
inference=False,
augmentation_preset='Normal'):
#if training:
if not inference:
@tf.function
def _map_fn(img): # preprocessing
img = tf.image.random_flip_left_right(img)
if augmentation_preset == 'Normal':
shape_y = load_size
shape_x = load_size
elif augmentation_preset == 'Ratio':
scale = tf.random.uniform([], minval=1.0001, maxval=1.0625)
shape = tf.shape(img)
min_shape = tf.math.minimum(shape[0], shape[1])
ratio = tf.cast(load_size / min_shape, tf.float32)
shape_y = tf.cast(shape[0], tf.float32)
shape_x = tf.cast(shape[1], tf.float32)
shape_y = tf.cast(shape_y * scale * ratio, tf.int32)
shape_x = tf.cast(shape_x * scale * ratio, tf.int32)
img = tf.image.resize(img, [shape_y, shape_x])
img = (img - tf.reduce_min(img)) / (
tf.reduce_max(img) - tf.reduce_min(img)
) # or img = tf.clip_by_value(img, 0, 255) / 255.0 # or img = tl.minmax_norm(img)
img = tf.image.random_crop(
img,
[crop_size, crop_size, tf.shape(img)[-1]])
img = img * 2 - 1
return img
elif inference:
@tf.function
def _map_fn(img): # preprocessing
img = tf.clip_by_value(img, 0,
255) / 255.0 # or img = tl.minmax_norm(img)
img = img * 2 - 1
return img
else:
@tf.function
def _map_fn(img): # preprocessing
img = tf.image.resize(
img, [crop_size, crop_size]
) # or img = tf.image.resize(img, [load_size, load_size]); img = tl.center_crop(img, crop_size)
img = img * 2 - 1
return img
return tl.disk_image_batch_dataset(img_paths,
batch_size,
drop_remainder=drop_remainder,
map_fn=_map_fn,
shuffle=shuffle,
repeat=repeat)