def downsample(imgs, s, size=None):
if size is None:
size = utils.get_data_size(imgs)
if size == 3:
return np_downsample_3d(imgs, s)
elif size == 2:
return np_downsample_2d(imgs, s)
elif size == 1:
return np_downsample_1d(imgs, s)
else:
raise ValueError("Size should be 1,2,3 or None")
def up_sampler(x, s=2, size=None, smoothout=False):
"""Up sampling operation"""
if size is None:
size = utils.get_data_size(x)
if not (size in [1, 2, 3]):
raise ValueError("Wrong size parameter")
bs = tf.shape(x)[0]
dims = x.shape.as_list()[1:]
if size == 3:
filt = tf.constant(1, dtype=tf.float32, shape=[s, s, s, 1, 1])
output_shape = [bs, dims[0] * s, dims[1] * s, dims[2] * s, dims[3]]
return tf.nn.conv3d_transpose(x,
filt,
output_shape=output_shape,
strides=[1, s, s, s, 1],
padding='SAME')
elif size == 2:
filt = tf.constant(1, dtype=tf.float32, shape=[s, s, 1, 1])
output_shape = [bs, dims[0] * s, dims[1] * s, 1]
if dims[-1] == 1:
x = tf.nn.conv2d_transpose(x,
filt,
output_shape=output_shape,
strides=[1, s, s, 1],
padding='SAME')
if smoothout:
paddings = tf.constant([[0, 0], [(s - 1) // 2, s // 2],
[(s - 1) // 2, s // 2], [0, 0]])
x = tf.pad(x, paddings, "SYMMETRIC")
x = tf.nn.conv2d(x,
filt / (s * s),
strides=[1, 1, 1, 1],
padding='VALID')
return x
else:
res = []
for sl in tf.split(x, dims[-1], axis=3):
tx = tf.nn.conv2d_transpose(sl,
filt,
output_shape=output_shape,
strides=[1, s, s, 1],
padding='SAME')
if smoothout:
paddings = tf.constant([[0, 0], [(s - 1) // 2, s // 2],
[(s - 1) // 2, s // 2], [0, 0]])
tx = tf.pad(tx, paddings, "SYMMETRIC")
tx = tf.nn.conv2d(tx,
filt / (s * s),
strides=[1, 1, 1, 1],
padding='VALID')
res.append(tx)
return tf.concat(res, axis=3)
else:
filt = tf.constant(1, dtype=tf.float32, shape=[s, 1, 1])
output_shape = [bs, dims[0] * s, dims[1]]
return tf.nn.conv1d_transpose(x,
filt,
output_shape=output_shape,
strides=s,
padding='SAME')
def down_sampler(x=None, s=2, size=None):
'''
Op to downsample 2D or 3D images by factor 's'.
This method works for both inputs: tensor or placeholder
'''
if size is None:
size = utils.get_data_size(x)
if not (size in [1, 2, 3]):
raise ValueError("Wrong size parameter")
# The input to the downsampling operation is a placeholder.
if x is None:
if size == 3:
addname = '3d_'
elif size == 2:
addname = '2d_'
else:
addname = '1d_'
placeholder_name = 'down_sampler_in_' + addname + str(s)
down_sampler_x = tf.placeholder(dtype=tf.float32,
name=placeholder_name)
op_name = 'down_sampler_out_' + addname + str(s)
# The input to the downsampling operation is the input tensor x.
else:
down_sampler_x = x
op_name = None
if size == 3:
filt = tf.constant(1 / (s * s * s),
dtype=tf.float32,
shape=[s, s, s, 1, 1])
return tf.nn.conv3d(down_sampler_x,
filt,
strides=[1, s, s, s, 1],
padding='SAME',
name=op_name)
elif size == 2:
filt = tf.constant(1 / (s * s), dtype=tf.float32, shape=[s, s, 1, 1])
ldim = down_sampler_x.shape[-1]
if ldim == 1 or not (isinstance(ldim, int)):
return tf.nn.conv2d(down_sampler_x,
filt,
strides=[1, s, s, 1],
padding='SAME',
name=op_name)
else:
res = []
for sl in tf.split(down_sampler_x,
down_sampler_x.shape[-1],
axis=3):
if op_name is not None:
op_name += 'I'
res.append(
tf.nn.conv2d(sl,
filt,
strides=[1, s, s, 1],
padding='SAME',
name=op_name))
return tf.concat(res, axis=3)
else:
filt = tf.constant(1 / s, dtype=tf.float32, shape=[s, 1, 1])
return tf.nn.conv1d(down_sampler_x,
filt,
stride=s,
padding='SAME',
name=op_name)