369 lines
15 KiB
Python
369 lines
15 KiB
Python
|
'''Network architecture for DUF:
|
||
|
|
Deep Video Super-Resolution Network Using Dynamic Upsampling Filters
|
||
|
|
Without Explicit Motion Compensation (CVPR18)
|
||
|
|
https://github.com/yhjo09/VSR-DUF
|
||
|
|
|
||
|
|
For all the models below, [adapt_official] is only necessary when
|
||
|
|
loading the weights converted from the official TensorFlow weights.
|
||
|
|
Please set it to [False] if you are training the model from scratch.
|
||
|
|
'''
|
||
|
|
|
||
|
|
import numpy as np
|
||
|
|
import torch
|
||
|
|
import torch.nn as nn
|
||
|
|
import torch.nn.functional as F
|
||
|
|
|
||
|
|
|
||
|
|
def adapt_official(Rx, scale=4):
|
||
|
|
'''Adapt the weights translated from the official tensorflow weights
|
||
|
|
Not necessary if you are training from scratch'''
|
||
|
|
x = Rx.clone()
|
||
|
|
x1 = x[:, ::3, :, :]
|
||
|
|
x2 = x[:, 1::3, :, :]
|
||
|
|
x3 = x[:, 2::3, :, :]
|
||
|
|
|
||
|
|
Rx[:, :scale**2, :, :] = x1
|
||
|
|
Rx[:, scale**2:2 * (scale**2), :, :] = x2
|
||
|
|
Rx[:, 2 * (scale**2):, :, :] = x3
|
||
|
|
|
||
|
|
return Rx
|
||
|
|
|
||
|
|
|
||
|
|
class DenseBlock(nn.Module):
|
||
|
|
'''Dense block
|
||
|
|
for the second denseblock, t_reduced = True'''
|
||
|
|
|
||
|
|
def __init__(self, nf=64, ng=32, t_reduce=False):
|
||
|
|
super(DenseBlock, self).__init__()
|
||
|
|
self.t_reduce = t_reduce
|
||
|
|
if self.t_reduce:
|
||
|
|
pad = (0, 1, 1)
|
||
|
|
else:
|
||
|
|
pad = (1, 1, 1)
|
||
|
|
self.bn3d_1 = nn.BatchNorm3d(nf, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_1 = nn.Conv3d(nf, nf, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0), bias=True)
|
||
|
|
self.bn3d_2 = nn.BatchNorm3d(nf, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_2 = nn.Conv3d(nf, ng, (3, 3, 3), stride=(1, 1, 1), padding=pad, bias=True)
|
||
|
|
self.bn3d_3 = nn.BatchNorm3d(nf + ng, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_3 = nn.Conv3d(nf + ng, nf + ng, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.bn3d_4 = nn.BatchNorm3d(nf + ng, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_4 = nn.Conv3d(nf + ng, ng, (3, 3, 3), stride=(1, 1, 1), padding=pad, bias=True)
|
||
|
|
self.bn3d_5 = nn.BatchNorm3d(nf + 2 * ng, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_5 = nn.Conv3d(nf + 2 * ng, nf + 2 * ng, (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
self.bn3d_6 = nn.BatchNorm3d(nf + 2 * ng, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_6 = nn.Conv3d(nf + 2 * ng, ng, (3, 3, 3), stride=(1, 1, 1), padding=pad,
|
||
|
|
bias=True)
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''x: [B, C, T, H, W]
|
||
|
|
C: nf -> nf + 3 * ng
|
||
|
|
T: 1) 7 -> 7 (t_reduce=False);
|
||
|
|
2) 7 -> 7 - 2 * 3 = 1 (t_reduce=True)'''
|
||
|
|
x1 = self.conv3d_1(F.relu(self.bn3d_1(x), inplace=True))
|
||
|
|
x1 = self.conv3d_2(F.relu(self.bn3d_2(x1), inplace=True))
|
||
|
|
if self.t_reduce:
|
||
|
|
x1 = torch.cat((x[:, :, 1:-1, :, :], x1), 1)
|
||
|
|
else:
|
||
|
|
x1 = torch.cat((x, x1), 1)
|
||
|
|
|
||
|
|
x2 = self.conv3d_3(F.relu(self.bn3d_3(x1), inplace=True))
|
||
|
|
x2 = self.conv3d_4(F.relu(self.bn3d_4(x2), inplace=True))
|
||
|
|
if self.t_reduce:
|
||
|
|
x2 = torch.cat((x1[:, :, 1:-1, :, :], x2), 1)
|
||
|
|
else:
|
||
|
|
x2 = torch.cat((x1, x2), 1)
|
||
|
|
|
||
|
|
x3 = self.conv3d_5(F.relu(self.bn3d_5(x2), inplace=True))
|
||
|
|
x3 = self.conv3d_6(F.relu(self.bn3d_6(x3), inplace=True))
|
||
|
|
if self.t_reduce:
|
||
|
|
x3 = torch.cat((x2[:, :, 1:-1, :, :], x3), 1)
|
||
|
|
else:
|
||
|
|
x3 = torch.cat((x2, x3), 1)
|
||
|
|
return x3
|
||
|
|
|
||
|
|
|
||
|
|
class DynamicUpsamplingFilter_3C(nn.Module):
|
||
|
|
'''dynamic upsampling filter with 3 channels applying the same filters
|
||
|
|
filter_size: filter size of the generated filters, shape (C, kH, kW)'''
|
||
|
|
|
||
|
|
def __init__(self, filter_size=(1, 5, 5)):
|
||
|
|
super(DynamicUpsamplingFilter_3C, self).__init__()
|
||
|
|
# generate a local expansion filter, used similar to im2col
|
||
|
|
nF = np.prod(filter_size)
|
||
|
|
expand_filter_np = np.reshape(np.eye(nF, nF),
|
||
|
|
(nF, filter_size[0], filter_size[1], filter_size[2]))
|
||
|
|
expand_filter = torch.from_numpy(expand_filter_np).float()
|
||
|
|
self.expand_filter = torch.cat((expand_filter, expand_filter, expand_filter),
|
||
|
|
0) # [75, 1, 5, 5]
|
||
|
|
|
||
|
|
def forward(self, x, filters):
|
||
|
|
'''x: input image, [B, 3, H, W]
|
||
|
|
filters: generate dynamic filters, [B, F, R, H, W], e.g., [B, 25, 16, H, W]
|
||
|
|
F: prod of filter kernel size, e.g., 5*5 = 25
|
||
|
|
R: used for upsampling, similar to pixel shuffle, e.g., 4*4 = 16 for x4
|
||
|
|
Return: filtered image, [B, 3*R, H, W]
|
||
|
|
'''
|
||
|
|
B, nF, R, H, W = filters.size()
|
||
|
|
# using group convolution
|
||
|
|
input_expand = F.conv2d(x, self.expand_filter.type_as(x), padding=2,
|
||
|
|
groups=3) # [B, 75, H, W] similar to im2col
|
||
|
|
input_expand = input_expand.view(B, 3, nF, H, W).permute(0, 3, 4, 1, 2) # [B, H, W, 3, 25]
|
||
|
|
filters = filters.permute(0, 3, 4, 1, 2) # [B, H, W, 25, 16]
|
||
|
|
out = torch.matmul(input_expand, filters) # [B, H, W, 3, 16]
|
||
|
|
return out.permute(0, 3, 4, 1, 2).view(B, 3 * R, H, W) # [B, 3*16, H, W]
|
||
|
|
|
||
|
|
|
||
|
|
class DUF_16L(nn.Module):
|
||
|
|
'''Official DUF structure with 16 layers'''
|
||
|
|
|
||
|
|
def __init__(self, scale=4, adapt_official=False):
|
||
|
|
super(DUF_16L, self).__init__()
|
||
|
|
self.conv3d_1 = nn.Conv3d(3, 64, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=True)
|
||
|
|
self.dense_block_1 = DenseBlock(64, 64 // 2, t_reduce=False) # 64 + 32 * 3 = 160, T = 7
|
||
|
|
self.dense_block_2 = DenseBlock(160, 64 // 2, t_reduce=True) # 160 + 32 * 3 = 256, T = 1
|
||
|
|
self.bn3d_2 = nn.BatchNorm3d(256, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_2 = nn.Conv3d(256, 256, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1),
|
||
|
|
bias=True)
|
||
|
|
|
||
|
|
self.conv3d_r1 = nn.Conv3d(256, 256, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_r2 = nn.Conv3d(256, 3 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.conv3d_f1 = nn.Conv3d(256, 512, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_f2 = nn.Conv3d(512, 1 * 5 * 5 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.dynamic_filter = DynamicUpsamplingFilter_3C((1, 5, 5))
|
||
|
|
|
||
|
|
self.scale = scale
|
||
|
|
self.adapt_official = adapt_official
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''
|
||
|
|
x: [B, T, C, H, W], T = 7. reshape to [B, C, T, H, W] for Conv3D
|
||
|
|
Generate filters and image residual:
|
||
|
|
Fx: [B, 25, 16, H, W] for DynamicUpsamplingFilter_3C
|
||
|
|
Rx: [B, 3*16, 1, H, W]
|
||
|
|
'''
|
||
|
|
B, T, C, H, W = x.size()
|
||
|
|
x = x.permute(0, 2, 1, 3, 4) # [B, C, T, H, W] for Conv3D
|
||
|
|
x_center = x[:, :, T // 2, :, :]
|
||
|
|
|
||
|
|
x = self.conv3d_1(x)
|
||
|
|
x = self.dense_block_1(x)
|
||
|
|
x = self.dense_block_2(x) # reduce T to 1
|
||
|
|
x = F.relu(self.conv3d_2(F.relu(self.bn3d_2(x), inplace=True)), inplace=True)
|
||
|
|
|
||
|
|
# image residual
|
||
|
|
Rx = self.conv3d_r2(F.relu(self.conv3d_r1(x), inplace=True)) # [B, 3*16, 1, H, W]
|
||
|
|
|
||
|
|
# filter
|
||
|
|
Fx = self.conv3d_f2(F.relu(self.conv3d_f1(x), inplace=True)) # [B, 25*16, 1, H, W]
|
||
|
|
Fx = F.softmax(Fx.view(B, 25, self.scale**2, H, W), dim=1)
|
||
|
|
|
||
|
|
# Adapt to official model weights
|
||
|
|
if self.adapt_official:
|
||
|
|
adapt_official(Rx, scale=self.scale)
|
||
|
|
|
||
|
|
# dynamic filter
|
||
|
|
out = self.dynamic_filter(x_center, Fx) # [B, 3*R, H, W]
|
||
|
|
out += Rx.squeeze_(2)
|
||
|
|
out = F.pixel_shuffle(out, self.scale) # [B, 3, H, W]
|
||
|
|
|
||
|
|
return out
|
||
|
|
|
||
|
|
|
||
|
|
class DenseBlock_28L(nn.Module):
|
||
|
|
'''The first part of the dense blocks used in DUF_28L
|
||
|
|
Temporal dimension remains the same here'''
|
||
|
|
|
||
|
|
def __init__(self, nf=64, ng=16):
|
||
|
|
super(DenseBlock_28L, self).__init__()
|
||
|
|
pad = (1, 1, 1)
|
||
|
|
|
||
|
|
dense_block_l = []
|
||
|
|
for i in range(0, 9):
|
||
|
|
dense_block_l.append(nn.BatchNorm3d(nf + i * ng, eps=1e-3, momentum=1e-3))
|
||
|
|
dense_block_l.append(nn.ReLU())
|
||
|
|
dense_block_l.append(
|
||
|
|
nn.Conv3d(nf + i * ng, nf + i * ng, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True))
|
||
|
|
|
||
|
|
dense_block_l.append(nn.BatchNorm3d(nf + i * ng, eps=1e-3, momentum=1e-3))
|
||
|
|
dense_block_l.append(nn.ReLU())
|
||
|
|
dense_block_l.append(
|
||
|
|
nn.Conv3d(nf + i * ng, ng, (3, 3, 3), stride=(1, 1, 1), padding=pad, bias=True))
|
||
|
|
|
||
|
|
self.dense_blocks = nn.ModuleList(dense_block_l)
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''x: [B, C, T, H, W]
|
||
|
|
C: 1) 64 -> 208;
|
||
|
|
T: 1) 7 -> 7; (t_reduce=True)'''
|
||
|
|
for i in range(0, len(self.dense_blocks), 6):
|
||
|
|
y = x
|
||
|
|
for j in range(6):
|
||
|
|
y = self.dense_blocks[i + j](y)
|
||
|
|
x = torch.cat((x, y), 1)
|
||
|
|
return x
|
||
|
|
|
||
|
|
|
||
|
|
class DUF_28L(nn.Module):
|
||
|
|
'''Official DUF structure with 28 layers'''
|
||
|
|
|
||
|
|
def __init__(self, scale=4, adapt_official=False):
|
||
|
|
super(DUF_28L, self).__init__()
|
||
|
|
self.conv3d_1 = nn.Conv3d(3, 64, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=True)
|
||
|
|
self.dense_block_1 = DenseBlock_28L(64, 16) # 64 + 16 * 9 = 208, T = 7
|
||
|
|
self.dense_block_2 = DenseBlock(208, 16, t_reduce=True) # 208 + 16 * 3 = 256, T = 1
|
||
|
|
self.bn3d_2 = nn.BatchNorm3d(256, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_2 = nn.Conv3d(256, 256, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1),
|
||
|
|
bias=True)
|
||
|
|
|
||
|
|
self.conv3d_r1 = nn.Conv3d(256, 256, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_r2 = nn.Conv3d(256, 3 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.conv3d_f1 = nn.Conv3d(256, 512, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_f2 = nn.Conv3d(512, 1 * 5 * 5 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.dynamic_filter = DynamicUpsamplingFilter_3C((1, 5, 5))
|
||
|
|
|
||
|
|
self.scale = scale
|
||
|
|
self.adapt_official = adapt_official
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''
|
||
|
|
x: [B, T, C, H, W], T = 7. reshape to [B, C, T, H, W] for Conv3D
|
||
|
|
Generate filters and image residual:
|
||
|
|
Fx: [B, 25, 16, H, W] for DynamicUpsamplingFilter_3C
|
||
|
|
Rx: [B, 3*16, 1, H, W]
|
||
|
|
'''
|
||
|
|
B, T, C, H, W = x.size()
|
||
|
|
x = x.permute(0, 2, 1, 3, 4) # [B,C,T,H,W] for Conv3D
|
||
|
|
x_center = x[:, :, T // 2, :, :]
|
||
|
|
x = self.conv3d_1(x)
|
||
|
|
x = self.dense_block_1(x)
|
||
|
|
x = self.dense_block_2(x) # reduce T to 1
|
||
|
|
x = F.relu(self.conv3d_2(F.relu(self.bn3d_2(x), inplace=True)), inplace=True)
|
||
|
|
|
||
|
|
# image residual
|
||
|
|
Rx = self.conv3d_r2(F.relu(self.conv3d_r1(x), inplace=True)) # [B, 3*16, 1, H, W]
|
||
|
|
|
||
|
|
# filter
|
||
|
|
Fx = self.conv3d_f2(F.relu(self.conv3d_f1(x), inplace=True)) # [B, 25*16, 1, H, W]
|
||
|
|
Fx = F.softmax(Fx.view(B, 25, self.scale**2, H, W), dim=1)
|
||
|
|
|
||
|
|
# Adapt to official model weights
|
||
|
|
if self.adapt_official:
|
||
|
|
adapt_official(Rx, scale=self.scale)
|
||
|
|
|
||
|
|
# dynamic filter
|
||
|
|
out = self.dynamic_filter(x_center, Fx) # [B, 3*R, H, W]
|
||
|
|
out += Rx.squeeze_(2)
|
||
|
|
out = F.pixel_shuffle(out, self.scale) # [B, 3, H, W]
|
||
|
|
return out
|
||
|
|
|
||
|
|
|
||
|
|
class DenseBlock_52L(nn.Module):
|
||
|
|
'''The first part of the dense blocks used in DUF_52L
|
||
|
|
Temporal dimension remains the same here'''
|
||
|
|
|
||
|
|
def __init__(self, nf=64, ng=16):
|
||
|
|
super(DenseBlock_52L, self).__init__()
|
||
|
|
pad = (1, 1, 1)
|
||
|
|
|
||
|
|
dense_block_l = []
|
||
|
|
for i in range(0, 21):
|
||
|
|
dense_block_l.append(nn.BatchNorm3d(nf + i * ng, eps=1e-3, momentum=1e-3))
|
||
|
|
dense_block_l.append(nn.ReLU())
|
||
|
|
dense_block_l.append(
|
||
|
|
nn.Conv3d(nf + i * ng, nf + i * ng, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True))
|
||
|
|
|
||
|
|
dense_block_l.append(nn.BatchNorm3d(nf + i * ng, eps=1e-3, momentum=1e-3))
|
||
|
|
dense_block_l.append(nn.ReLU())
|
||
|
|
dense_block_l.append(
|
||
|
|
nn.Conv3d(nf + i * ng, ng, (3, 3, 3), stride=(1, 1, 1), padding=pad, bias=True))
|
||
|
|
|
||
|
|
self.dense_blocks = nn.ModuleList(dense_block_l)
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''x: [B, C, T, H, W]
|
||
|
|
C: 1) 64 -> 400;
|
||
|
|
T: 1) 7 -> 7; (t_reduce=True)'''
|
||
|
|
for i in range(0, len(self.dense_blocks), 6):
|
||
|
|
y = x
|
||
|
|
for j in range(6):
|
||
|
|
y = self.dense_blocks[i + j](y)
|
||
|
|
x = torch.cat((x, y), 1)
|
||
|
|
return x
|
||
|
|
|
||
|
|
|
||
|
|
class DUF_52L(nn.Module):
|
||
|
|
'''Official DUF structure with 52 layers'''
|
||
|
|
|
||
|
|
def __init__(self, scale=4, adapt_official=False):
|
||
|
|
super(DUF_52L, self).__init__()
|
||
|
|
self.conv3d_1 = nn.Conv3d(3, 64, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1), bias=True)
|
||
|
|
self.dense_block_1 = DenseBlock_52L(64, 16) # 64 + 21 * 9 = 400, T = 7
|
||
|
|
self.dense_block_2 = DenseBlock(400, 16, t_reduce=True) # 400 + 16 * 3 = 448, T = 1
|
||
|
|
|
||
|
|
self.bn3d_2 = nn.BatchNorm3d(448, eps=1e-3, momentum=1e-3)
|
||
|
|
self.conv3d_2 = nn.Conv3d(448, 256, (1, 3, 3), stride=(1, 1, 1), padding=(0, 1, 1),
|
||
|
|
bias=True)
|
||
|
|
|
||
|
|
self.conv3d_r1 = nn.Conv3d(256, 256, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_r2 = nn.Conv3d(256, 3 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.conv3d_f1 = nn.Conv3d(256, 512, (1, 1, 1), stride=(1, 1, 1), padding=(0, 0, 0),
|
||
|
|
bias=True)
|
||
|
|
self.conv3d_f2 = nn.Conv3d(512, 1 * 5 * 5 * (scale**2), (1, 1, 1), stride=(1, 1, 1),
|
||
|
|
padding=(0, 0, 0), bias=True)
|
||
|
|
|
||
|
|
self.dynamic_filter = DynamicUpsamplingFilter_3C((1, 5, 5))
|
||
|
|
|
||
|
|
self.scale = scale
|
||
|
|
self.adapt_official = adapt_official
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''
|
||
|
|
x: [B, T, C, H, W], T = 7. reshape to [B, C, T, H, W] for Conv3D
|
||
|
|
Generate filters and image residual:
|
||
|
|
Fx: [B, 25, 16, H, W] for DynamicUpsamplingFilter_3C
|
||
|
|
Rx: [B, 3*16, 1, H, W]
|
||
|
|
'''
|
||
|
|
B, T, C, H, W = x.size()
|
||
|
|
x = x.permute(0, 2, 1, 3, 4) # [B,C,T,H,W] for Conv3D
|
||
|
|
x_center = x[:, :, T // 2, :, :]
|
||
|
|
x = self.conv3d_1(x)
|
||
|
|
x = self.dense_block_1(x)
|
||
|
|
x = self.dense_block_2(x)
|
||
|
|
x = F.relu(self.conv3d_2(F.relu(self.bn3d_2(x), inplace=True)), inplace=True)
|
||
|
|
|
||
|
|
# image residual
|
||
|
|
Rx = self.conv3d_r2(F.relu(self.conv3d_r1(x), inplace=True)) # [B, 3*16, 1, H, W]
|
||
|
|
|
||
|
|
# filter
|
||
|
|
Fx = self.conv3d_f2(F.relu(self.conv3d_f1(x), inplace=True)) # [B, 25*16, 1, H, W]
|
||
|
|
Fx = F.softmax(Fx.view(B, 25, self.scale**2, H, W), dim=1)
|
||
|
|
|
||
|
|
# Adapt to official model weights
|
||
|
|
if self.adapt_official:
|
||
|
|
adapt_official(Rx, scale=self.scale)
|
||
|
|
|
||
|
|
# dynamic filter
|
||
|
|
out = self.dynamic_filter(x_center, Fx) # [B, 3*R, H, W]
|
||
|
|
out += Rx.squeeze_(2)
|
||
|
|
out = F.pixel_shuffle(out, self.scale) # [B, 3, H, W]
|
||
|
|
return out
|