__author__ = 'Felix Kleinert'
__date__ = '2020-03-02'
import tensorflow as tf
import numpy as np
import keras.backend as K
from keras.layers.convolutional import _ZeroPadding
from keras.legacy import interfaces
from keras.utils import conv_utils
from keras.utils.generic_utils import transpose_shape
from keras.backend.common import normalize_data_format
class PadUtils:
"""
Helper class for advanced paddings
"""
@staticmethod
def get_padding_for_same(kernel_size, strides=1):
"""
This methods calculates the padding size to keep input and output dimensions equal for a given kernel size
(STRIDES HAVE TO BE EQUAL TO ONE!)
:param kernel_size:
:return:
"""
if strides != 1:
raise NotImplementedError("Strides other than 1 not implemented!")
if not all(isinstance(k, int) for k in kernel_size):
raise ValueError(f"The `kernel_size` argument must have a tuple of integers. Got: {kernel_size} "
f"of type {[type(k) for k in kernel_size]}")
ks = np.array(kernel_size, dtype=np.int64)
if any(k <= 0 for k in ks):
raise ValueError(f"All values of kernel_size must be > 0. Got: {kernel_size} ")
if all(k % 2 == 1 for k in ks): # (d & 0x1 for d in ks):
pad = ((ks - 1) / 2).astype(np.int64)
# convert numpy int to base int
pad = [np.asscalar(v) for v in pad]
return tuple(pad)
# return tuple(PadUtils.check_padding_format(pad))
else:
raise NotImplementedError(f"even kernel size not implemented. Got {kernel_size}")
@staticmethod
def spatial_2d_padding(padding=((1, 1), (1, 1)), data_format=None):
"""Pads the 2nd and 3rd dimensions of a 4D tensor.
# Arguments
x: Tensor or variable.
padding: Tuple of 2 tuples, padding pattern.
data_format: string, `"channels_last"` or `"channels_first"`.
# Returns
A padded 4D tensor.
# Raises
ValueError: if `data_format` is neither `"channels_last"` or `"channels_first"`.
"""
assert len(padding) == 2
assert len(padding[0]) == 2
assert len(padding[1]) == 2
data_format = normalize_data_format(data_format)
pattern = [[0, 0],
list(padding[0]),
list(padding[1]),
[0, 0]]
pattern = transpose_shape(pattern, data_format, spatial_axes=(1, 2))
return pattern
@staticmethod
def check_padding_format(padding):
if isinstance(padding, int):
normalized_padding = ((padding, padding), (padding, padding))
elif hasattr(padding, '__len__'):
if len(padding) != 2:
raise ValueError('`padding` should have two elements. '
'Found: ' + str(padding))
for idx_pad, sub_pad in enumerate(padding):
if isinstance(sub_pad, str):
raise ValueError(f'`padding[{idx_pad}]` is str but must be int')
if hasattr(sub_pad, '__len__'):
if len(sub_pad) != 2:
raise ValueError(f'`padding[{idx_pad}]` should have one or two elements. '
f'Found: {padding[idx_pad]}')
if not all(isinstance(sub_k, int) for sub_k in padding[idx_pad]):
raise ValueError(f'`padding[{idx_pad}]` should have one or two elements of type int. '
f"Found:{padding[idx_pad]} of type {[type(sub_k) for sub_k in padding[idx_pad]]}")
height_padding = conv_utils.normalize_tuple(padding[0], 2,
'1st entry of padding')
if not all(k >= 0 for k in height_padding):
raise ValueError(f"The `1st entry of padding` argument must be >= 0. Received: {padding[0]} of type {type(padding[0])}")
width_padding = conv_utils.normalize_tuple(padding[1], 2,
'2nd entry of padding')
if not all(k >= 0 for k in width_padding):
raise ValueError(f"The `2nd entry of padding` argument must be >= 0. Received: {padding[1]} of type {type(padding[1])}")
normalized_padding = (height_padding, width_padding)
else:
raise ValueError('`padding` should be either an int, '
'a tuple of 2 ints '
'(symmetric_height_pad, symmetric_width_pad), '
'or a tuple of 2 tuples of 2 ints '
'((top_pad, bottom_pad), (left_pad, right_pad)). '
f'Found: {padding} of type {type(padding)}')
return normalized_padding
class ReflectionPadding2D(_ZeroPadding):
"""
Reflection padding layer for 2D input. This custum padding layer is built on keras' zero padding layers. Doc is copy
pasted from the original functions/methods:
This layer can add rows and columns of reflected values
at the top, bottom, left and right side of an image like tensor.
Example:
6, 5, 4, 5, 6, 5, 4
_________
1, 2, 3 RefPad(padding=[[1, 1,], [2, 2]]) 3, 2,| 1, 2, 3,| 2, 1
4, 5, 6 =============================>>>> 6, 5,| 4, 5, 6,| 5, 4
_________
3, 2, 1, 2, 3, 2, 1
'# Arguments
padding: int, or tuple of 2 ints, or tuple of 2 tuples of 2 ints.
- If int: the same symmetric padding
is applied to height and width.
- If tuple of 2 ints:
interpreted as two different
symmetric padding values for height and width:
`(symmetric_height_pad, symmetric_width_pad)`.
- If tuple of 2 tuples of 2 ints:
interpreted as
`((top_pad, bottom_pad), (left_pad, right_pad))`
data_format: A string,
one of `"channels_last"` or `"channels_first"`.
The ordering of the dimensions in the inputs.
`"channels_last"` corresponds to inputs with shape
`(batch, height, width, channels)` while `"channels_first"`
corresponds to inputs with shape
`(batch, channels, height, width)`.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
# Input shape
4D tensor with shape:
- If `data_format` is `"channels_last"`:
`(batch, rows, cols, channels)`
- If `data_format` is `"channels_first"`:
`(batch, channels, rows, cols)`
# Output shape
4D tensor with shape:
- If `data_format` is `"channels_last"`:
`(batch, padded_rows, padded_cols, channels)`
- If `data_format` is `"channels_first"`:
`(batch, channels, padded_rows, padded_cols)`
'
"""
@interfaces.legacy_zeropadding2d_support
def __init__(self,
padding=(1, 1),
data_format=None,
**kwargs):
normalized_padding = PadUtils.check_padding_format(padding=padding)
super(ReflectionPadding2D, self).__init__(normalized_padding,
data_format,
**kwargs)
def call(self, inputs, mask=None):
pattern = PadUtils.spatial_2d_padding(padding=self.padding, data_format=self.data_format)
return tf.pad(inputs, pattern, 'REFLECT')
class SymmetricPadding2D(_ZeroPadding):
"""
Symmetric padding layer for 2D input. This custom padding layer is built on keras' zero padding layers. Doc is copy
pasted from the original functions/methods:
This layer can add rows and columns of symmetric values
at the top, bottom, left and right side of an image like tensor.
Example:
2, 1, 1, 2, 3, 3, 2
_________
1, 2, 3 SymPad(padding=[[1, 1,], [2, 2]]) 2, 1,| 1, 2, 3,| 3, 2
4, 5, 6 =============================>>>> 5, 4,| 4, 5, 6,| 6, 5
_________
5, 4, 4, 5, 6, 6, 5
'# Arguments
padding: int, or tuple of 2 ints, or tuple of 2 tuples of 2 ints.
- If int: the same symmetric padding
is applied to height and width.
- If tuple of 2 ints:
interpreted as two different
symmetric padding values for height and width:
`(symmetric_height_pad, symmetric_width_pad)`.
- If tuple of 2 tuples of 2 ints:
interpreted as
`((top_pad, bottom_pad), (left_pad, right_pad))`
data_format: A string,
one of `"channels_last"` or `"channels_first"`.
The ordering of the dimensions in the inputs.
`"channels_last"` corresponds to inputs with shape
`(batch, height, width, channels)` while `"channels_first"`
corresponds to inputs with shape
`(batch, channels, height, width)`.
It defaults to the `image_data_format` value found in your
Keras config file at `~/.keras/keras.json`.
If you never set it, then it will be "channels_last".
# Input shape
4D tensor with shape:
- If `data_format` is `"channels_last"`:
`(batch, rows, cols, channels)`
- If `data_format` is `"channels_first"`:
`(batch, channels, rows, cols)`
# Output shape
4D tensor with shape:
- If `data_format` is `"channels_last"`:
`(batch, padded_rows, padded_cols, channels)`
- If `data_format` is `"channels_first"`:
`(batch, channels, padded_rows, padded_cols)`
'
"""
@interfaces.legacy_zeropadding2d_support
def __init__(self,
padding=(1, 1),
data_format=None,
**kwargs):
normalized_padding = PadUtils.check_padding_format(padding=padding)
super(SymmetricPadding2D, self).__init__(normalized_padding,
data_format,
**kwargs)
def call(self, inputs, mask=None):
pattern = PadUtils.spatial_2d_padding(padding=self.padding, data_format=self.data_format)
return tf.pad(inputs, pattern, 'SYMMETRIC')
if __name__ == '__main__':
from keras.models import Model
from keras.layers import Conv2D, Flatten, Dense, Input
kernel_1 = (3, 3)
kernel_2 = (5, 5)
x = np.array(range(2000)).reshape(-1, 10, 10, 1)
y = x.mean(axis=(1, 2))
x_input = Input(shape=x.shape[1:])
pad1 = PadUtils.get_padding_for_same(kernel_size=kernel_1)
x_out = ReflectionPadding2D(padding=pad1, name="RefPAD")(x_input)
x_out = Conv2D(5, kernel_size=kernel_1, activation='relu')(x_out)
pad2 = PadUtils.get_padding_for_same(kernel_size=kernel_2)
x_out = SymmetricPadding2D(padding=pad2, name="SymPAD")(x_out)
x_out = Conv2D(2, kernel_size=kernel_2, activation='relu')(x_out)
x_out = Flatten()(x_out)
x_out = Dense(1, activation='linear')(x_out)
model = Model(inputs=x_input, outputs=x_out)
model.compile('adam', loss='mse')
model.summary()
model.fit(x, y, epochs=10)