def __init__(self):
super(TDAN_VSR, self).__init__()
self.name = 'TDAN'
self.conv_first = nn.Conv2d(3, 64, 3, padding=1, bias=True)
self.residual_layer = self.make_layer(Res_Block, 5)
self.relu = nn.ReLU(inplace=True)
# deformable
self.cr = nn.Conv2d(128, 64, 3, padding=1, bias=True)
self.off2d_1 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.dconv_1 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d_2 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.deconv_2 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d_3 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.deconv_3 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.dconv = ConvOffset2d(64,
64, (3, 3),
padding=(1, 1),
num_deformable_groups=8)
self.recon_lr = nn.Conv2d(64, 3, 3, padding=1, bias=True)
fea_ex = [nn.Conv2d(5 * 3, 64, 3, padding=1, bias=True), nn.ReLU()]
self.fea_ex = nn.Sequential(*fea_ex)
self.recon_layer = self.make_layer(Res_Block, 10)
upscaling = [
Upsampler(default_conv, 4, 64, act=False),
nn.Conv2d(64, 3, 3, padding=1, bias=False)
]
self.up = nn.Sequential(*upscaling)
# xavier initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
def __init__(self):
super(Net_Deform_CUDA, self).__init__()
self.relu = nn.ReLU()
self.pool = nn.MaxPool2d(2)
self.conv_off1 = nn.Conv2d(1, 1 * 2 * 5 * 5, 3, 1, 1,
bias=False).cuda()
self.conv1 = ConvOffset2d(1,
16, (5, 5),
stride=1,
padding=2,
num_deformable_groups=1).cuda()
self.conv_off2 = nn.Conv2d(16, 4 * 2 * 5 * 5, 3, 1, 1,
bias=False).cuda()
self.conv2 = ConvOffset2d(16,
32, (5, 5),
stride=1,
padding=2,
num_deformable_groups=4).cuda()
self.out = nn.Linear(32 * 7 * 7, 10)
def __init__(self):
super(align_net_w_feat, self).__init__()
self.conv_first = nn.Conv2d(3, 64, 3, padding=1, bias=True)
self.residual_layer = self.make_layer(Res_Block, 5)
self.relu = nn.ReLU(inplace=True)
# deformable
self.cr = nn.Conv2d(128, 64, 3, padding=1, bias=True)
self.off2d_1 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.dconv_1 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d_2 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.deconv_2 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d_3 = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.deconv_3 = ConvOffset2d(64,
64,
3,
padding=1,
num_deformable_groups=8)
self.off2d = nn.Conv2d(64, 18 * 8, 3, padding=1, bias=True)
self.dconv = ConvOffset2d(64,
64, (3, 3),
padding=(1, 1),
num_deformable_groups=8)
self.recon_lr = nn.Conv2d(64, 3, 3, padding=1, bias=True)
# xavier initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
def __init__(self, in_channels, growth_rate):
super(DenseLayer, self).__init__()
# self.add_module('norm', nn.BatchNorm2d(num_features=in_channels))
self.norm = nn.BatchNorm2d(num_features=in_channels)
# self.add_module('prelu', nn.PReLU())
self.prelu = nn.PReLU()
# author's impl - lasange 'same' pads with half
# filter size (rounded down) on "both" sides
self.conv_offset = nn.Conv2d(in_channels=in_channels, out_channels=num_deformable_groups * 2 * kH * kW,
kernel_size=(kH, kW), stride=(1, 1), padding=(1, 1), bias=True)
self.deform_conv = ConvOffset2d(
in_channels, growth_rate, (kH, kW), stride=1, padding=1, num_deformable_groups=num_deformable_groups)
self.drop = nn.Dropout2d(0.2)
def __init__(self,
in_planes,
out_planes,
kernel_size,
stride=1,
padding=0,
dilation=1,
groups=1,
bias=False):
super(Deform_Conv, self).__init__()
self.out_channels = out_planes
self.conv = nn.Conv2d(in_planes,
groups * 2 * kernel_size * kernel_size,
kernel_size=(kernel_size, kernel_size),
stride=(stride, stride),
padding=(padding, padding),
bias=bias)
self.conv_offset2d = ConvOffset2d(in_planes,
out_planes,
(kernel_size, kernel_size),
stride=stride,
padding=padding,
num_deformable_groups=groups)
else:
from dcn import ConvOffset2d
num_deformable_groups = 2
N, inC, inH, inW = 1, 6, 512, 512
outC, outH, outW = 4, 512, 512
kH, kW = 3, 3
conv = nn.Conv2d(inC,
num_deformable_groups * 2 * kH * kW,
kernel_size=(kH, kW),
stride=(1, 1),
padding=(1, 1),
bias=False).cuda()
conv_offset2d = ConvOffset2d(
inC,
outC, (kH, kW),
stride=1,
padding=1,
num_deformable_groups=num_deformable_groups).cuda()
pdb.set_trace()
inputs = Variable(torch.randn(N, inC, inH, inW).cuda())
offset = conv(inputs)
output = conv_offset2d(inputs, offset)
output.backward(output.data)
print(output.size())
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.autograd import gradcheck
from torch.nn.modules.utils import _single, _pair
from modules import ConvOffset2d
num_deformable_group = 1
N, inC, inH, inW = 1, 3, 512, 512
outC, outH, outW = 4, 512, 512
kH, kW = 3, 3
conv = nn.Conv2d(inC,
num_deformable_group * 2 * kH * kW,
kernel_size=(kH, kW),
stride=(1, 1),
padding=(1, 1),
bias=False).cuda()
conv_offset2d = ConvOffset2d(inC, outC, (kH, kW), stride=1, padding=1).cuda()
inputs = Variable(torch.randn(N, inC, inH, inW).cuda())
offset = conv(inputs)
output = conv_offset2d(inputs, offset)
output.backward(output.data)
print(output.size())