import torch
import torch.nn as nn
from mmcv.cnn import ConvModule, constant_init, xavier_init
from torch.nn import functional as F
[docs]class GCAModule(nn.Module):
"""Guided Contextual Attention Module.
From https://arxiv.org/pdf/2001.04069.pdf.
Based on https://github.com/nbei/Deep-Flow-Guided-Video-Inpainting.
This module use image feature map to augment the alpha feature map with
guided contextual attention score.
Image feature and alpha feature are unfolded to small patches and later
used as conv kernel. Thus, we refer the unfolding size as kernel size.
Image feature patches have a default kernel size 3 while the kernel size of
alpha feature patches could be specified by `rate` (see `rate` below). The
image feature patches are used to convolve with the image feature itself
to calculate the contextual attention. Then the attention feature map is
convolved by alpha feature patches to obtain the attentioned alpha feature.
At last, the attentioned alpah feature is added to the input alpha feature.
Args:
in_channels (int): Input channels of the guided contextual attention
module.
out_channels (int): Output channels of the guided contextual attention
module.
kernel_size (int): Kernel size of image feature patches. Default 3.
stride (int): Stride when unfolding the image feature. Default 1.
rate (int): The downsample rate of image feature map. The corresponding
kernel size and stride of alpha feature patches will be `rate x 2`
and `rate`. It could be regarded as the granularity of the gca
module. Default: 2.
pad_args (dict): Parameters of padding when convolve image feature with
image feature patches or alpha feature patches. Allowed keys are
`mode` and `value`. See torch.nn.functional.pad() for more
information. Default: dict(mode='reflect').
interpolation (str): Interpolation method in upsampling and
downsampling.
penalty (float): Punishment hyperparameter to avoid a large correlation
between each unknown patch and itself.
eps (float): A small number to avoid dividing by 0 when calculating
the normed image feature patch. Default: 1e-4.
"""
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
rate=2,
pad_args=dict(mode='reflect'),
interpolation='nearest',
penalty=-1e4,
eps=1e-4):
super(GCAModule, self).__init__()
self.kernel_size = kernel_size
self.stride = stride
self.rate = rate
self.pad_args = pad_args
self.interpolation = interpolation
self.penalty = penalty
self.eps = eps
# reduced the channels of input image feature.
self.guidance_conv = nn.Conv2d(in_channels, in_channels // 2, 1)
# convolution after the attentioned alpha feature
self.out_conv = ConvModule(
out_channels,
out_channels,
1,
norm_cfg=dict(type='BN'),
act_cfg=None)
self.init_weights()
def init_weights(self):
xavier_init(self.guidance_conv, distribution='uniform')
xavier_init(self.out_conv.conv, distribution='uniform')
constant_init(self.out_conv.norm, 1e-3)
[docs] def forward(self, img_feat, alpha_feat, unknown=None, softmax_scale=1.):
"""Forward function of GCAModule.
Args:
img_feat (Tensor): Image feature map of shape
(N, ori_c, ori_h, ori_w).
alpha_feat (Tensor): Alpha feature map of shape
(N, alpha_c, ori_h, ori_w).
unknown (Tensor, optional): Unknown area map generated by trimap.
If specified, this tensor should have shape
(N, 1, ori_h, ori_w).
softmax_scale (float, optional): The softmax scale of the attention
if unknown area is not provided in forward. Default: 1.
Returns:
Tensor: The augmented alpha feature.
"""
if alpha_feat.shape[2:4] != img_feat.shape[2:4]:
raise ValueError(
'image feature size does not align with alpha feature size: '
f'image feature size {img_feat.shape[2:4]}, '
f'alpha feature size {alpha_feat.shape[2:4]}')
if unknown is not None and unknown.shape[2:4] != img_feat.shape[2:4]:
raise ValueError(
'image feature size does not align with unknown mask size: '
f'image feature size {img_feat.shape[2:4]}, '
f'unknown mask size {unknown.shape[2:4]}')
# preprocess image feature
img_feat = self.guidance_conv(img_feat)
img_feat = F.interpolate(
img_feat, scale_factor=1 / self.rate, mode=self.interpolation)
# preprocess unknown mask
unknown, softmax_scale = self.process_unknown_mask(
unknown, img_feat, softmax_scale)
img_ps, alpha_ps, unknown_ps = self.extract_feature_maps_patches(
img_feat, alpha_feat, unknown)
# create self correlation mask with shape:
# (N, img_h*img_w, img_h, img_w)
self_mask = self.get_self_correlation_mask(img_feat)
# split tensors by batch dimension; tuple is returned
img_groups = torch.split(img_feat, 1, dim=0)
img_ps_groups = torch.split(img_ps, 1, dim=0)
alpha_ps_groups = torch.split(alpha_ps, 1, dim=0)
unknown_ps_groups = torch.split(unknown_ps, 1, dim=0)
scale_groups = torch.split(softmax_scale, 1, dim=0)
groups = (img_groups, img_ps_groups, alpha_ps_groups,
unknown_ps_groups, scale_groups)
out = []
# i is the virtual index of the sample in the current batch
for img_i, img_ps_i, alpha_ps_i, unknown_ps_i, scale_i in zip(*groups):
similarity_map = self.compute_similarity_map(img_i, img_ps_i)
gca_score = self.compute_guided_attention_score(
similarity_map, unknown_ps_i, scale_i, self_mask)
out_i = self.propagate_alpha_feature(gca_score, alpha_ps_i)
out.append(out_i)
out = torch.cat(out, dim=0)
out.reshape_as(alpha_feat)
out = self.out_conv(out) + alpha_feat
return out
[docs] def compute_similarity_map(self, img_feat, img_ps):
"""Compute similarity between image feature patches.
Args:
img_feat (Tensor): Image feature map of shape
(1, img_c, img_h, img_w).
img_ps (Tensor): Image feature patches tensor of shape
(1, img_h*img_w, img_c, img_ks, img_ks).
Returns:
Tensor: Similarity map between image feature patches with shape \
(1, img_h*img_w, img_h, img_w).
"""
img_ps = img_ps[0] # squeeze dim 0
# convolve the feature to get correlation (similarity) map
escape_NaN = torch.FloatTensor([self.eps]).to(img_feat)
img_ps_normed = img_ps / torch.max(self.l2_norm(img_ps), escape_NaN)
img_feat = self.pad(img_feat, self.kernel_size, self.stride)
similarity_map = F.conv2d(img_feat, img_ps_normed)
return similarity_map
[docs] def compute_guided_attention_score(self, similarity_map, unknown_ps, scale,
self_mask):
"""Compute guided attention score.
Args:
similarity_map (Tensor): Similarity map of image feature with shape
(1, img_h*img_w, img_h, img_w).
unknown_ps (Tensor): Unknown area patches tensor of shape
(1, img_h*img_w, 1, 1).
scale (Tensor): Softmax scale of known and unknown area:
[unknown_scale, known_scale].
self_mask (Tensor): Self correlation mask of shape
(1, img_h*img_w, img_h, img_w). At (1, i*i, i, i) mask value
equals -1e4 for i in [1, img_h*img_w] and other area is all
zero.
Returns:
Tensor: Similarity map between image feature patches with shape \
(1, img_h*img_w, img_h, img_w).
"""
# scale the correlation with predicted scale factor for known and
# unknown area
unknown_scale, known_scale = scale[0]
out = similarity_map * (
unknown_scale * unknown_ps.gt(0.).float() +
known_scale * unknown_ps.le(0.).float())
# mask itself, self-mask only applied to unknown area
out = out + self_mask * unknown_ps
gca_score = F.softmax(out, dim=1)
return gca_score
[docs] def propagate_alpha_feature(self, gca_score, alpha_ps):
"""Propagate alpha feature based on guided attention score.
Args:
gca_score (Tensor): Guided attention score map of shape
(1, img_h*img_w, img_h, img_w).
alpha_ps (Tensor): Alpha feature patches tensor of shape
(1, img_h*img_w, alpha_c, alpha_ks, alpha_ks).
Returns:
Tensor: Propagted alpha feature map of shape \
(1, alpha_c, alpha_h, alpha_w).
"""
alpha_ps = alpha_ps[0] # squeeze dim 0
if self.rate == 1:
gca_score = self.pad(gca_score, kernel_size=2, stride=1)
alpha_ps = alpha_ps.permute(1, 0, 2, 3)
out = F.conv2d(gca_score, alpha_ps) / 4.
else:
out = F.conv_transpose2d(
gca_score, alpha_ps, stride=self.rate, padding=1) / 4.
return out
[docs] def process_unknown_mask(self, unknown, img_feat, softmax_scale):
"""Process unknown mask.
Args:
unknown (Tensor, optional): Unknown area map generated by trimap of
shape (N, 1, ori_h, ori_w)
img_feat (Tensor): The interpolated image feature map of shape
(N, img_c, img_h, img_w).
softmax_scale (float, optional): The softmax scale of the attention
if unknown area is not provided in forward. Default: 1.
Returns:
tuple: 2-tuple of
``Tensor``: Interpolated unknown area map of shape \
(N, img_h*img_w, img_h, img_w).
``Tensor``: Softmax scale tensor of known and unknown area of \
shape (N, 2).
"""
n, _, h, w = img_feat.shape
if unknown is not None:
unknown = unknown.clone()
unknown = F.interpolate(
unknown, scale_factor=1 / self.rate, mode=self.interpolation)
unknown_mean = unknown.mean(dim=[2, 3])
known_mean = 1 - unknown_mean
unknown_scale = torch.clamp(
torch.sqrt(unknown_mean / known_mean), 0.1, 10).to(img_feat)
known_scale = torch.clamp(
torch.sqrt(known_mean / unknown_mean), 0.1, 10).to(img_feat)
softmax_scale = torch.cat([unknown_scale, known_scale], dim=1)
else:
unknown = torch.ones((n, 1, h, w)).to(img_feat)
softmax_scale = torch.FloatTensor(
[softmax_scale,
softmax_scale]).view(1, 2).repeat(n, 1).to(img_feat)
return unknown, softmax_scale
def pad(self, x, kernel_size, stride):
left = (kernel_size - stride + 1) // 2
right = (kernel_size - stride) // 2
pad = (left, right, left, right)
return F.pad(x, pad, **self.pad_args)
def get_self_correlation_mask(self, img_feat):
_, _, h, w = img_feat.shape
# As ONNX does not support dynamic num_classes, we have to convert it
# into an integer
self_mask = F.one_hot(
torch.arange(h * w).view(h, w), num_classes=int(h * w))
self_mask = self_mask.permute(2, 0, 1).view(1, h * w, h, w)
# use large negative value to mask out self-correlation before softmax
self_mask = self_mask * self.penalty
return self_mask.to(img_feat)
@staticmethod
def l2_norm(x):
x = x**2
x = x.sum(dim=[1, 2, 3], keepdim=True)
return torch.sqrt(x)