File size: 2,745 Bytes
712b45c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 |
# for contextual attention
import torch
def extract_image_patches(images, ksizes, strides, rates, padding="same"):
"""
Extract patches from images and put them in the C output dimension.
:param padding:
:param images: [batch, channels, in_rows, in_cols]. A 4-D Tensor with shape
:param ksizes: [ksize_rows, ksize_cols]. The size of the sliding window for
each dimension of images
:param strides: [stride_rows, stride_cols]
:param rates: [dilation_rows, dilation_cols]
:return: A Tensor
"""
assert len(images.size()) == 4
assert padding in ["same", "valid"]
batch_size, channel, height, width = images.size()
if padding == "same":
images = same_padding(images, ksizes, strides, rates)
elif padding == "valid":
pass
else:
raise NotImplementedError(
'Unsupported padding type: {}.\
Only "same" or "valid" are supported.'.format(
padding
)
)
unfold = torch.nn.Unfold(
kernel_size=ksizes, dilation=rates, padding=0, stride=strides
)
patches = unfold(images)
return patches # [N, C*k*k, L], L is the total number of such blocks
def same_padding(images, ksizes, strides, rates):
assert len(images.size()) == 4
batch_size, channel, rows, cols = images.size()
out_rows = (rows + strides[0] - 1) // strides[0]
out_cols = (cols + strides[1] - 1) // strides[1]
effective_k_row = (ksizes[0] - 1) * rates[0] + 1
effective_k_col = (ksizes[1] - 1) * rates[1] + 1
padding_rows = max(0, (out_rows - 1) * strides[0] + effective_k_row - rows)
padding_cols = max(0, (out_cols - 1) * strides[1] + effective_k_col - cols)
# Pad the input
padding_top = int(padding_rows / 2.0)
padding_left = int(padding_cols / 2.0)
padding_bottom = padding_rows - padding_top
padding_right = padding_cols - padding_left
paddings = (padding_left, padding_right, padding_top, padding_bottom)
images = torch.nn.ZeroPad2d(paddings)(images)
return images
def reduce_mean(x, axis=None, keepdim=False):
if not axis:
axis = range(len(x.shape))
for i in sorted(axis, reverse=True):
x = torch.mean(x, dim=i, keepdim=keepdim)
return x
def reduce_std(x, axis=None, keepdim=False):
if not axis:
axis = range(len(x.shape))
for i in sorted(axis, reverse=True):
x = torch.std(x, dim=i, keepdim=keepdim)
return x
def reduce_sum(x, axis=None, keepdim=False):
if not axis:
axis = range(len(x.shape))
for i in sorted(axis, reverse=True):
x = torch.sum(x, dim=i, keepdim=keepdim)
return x
|