forked from mrq/DL-Art-School
ae3301c0ea
Renames AttentiveRRDB to SwitchedRRDB. Moves SwitchedConv to an external repo (neonbjb/switchedconv). Switchs RDB blocks instead of conv blocks. Works good!
131 lines
4.3 KiB
Python
131 lines
4.3 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.init as init
|
|
import torch.nn.functional as F
|
|
import torch.nn.utils.spectral_norm as SpectralNorm
|
|
from math import sqrt
|
|
|
|
def pixel_norm(x, epsilon=1e-8):
|
|
return x * torch.rsqrt(torch.mean(torch.pow(x, 2), dim=1, keepdims=True) + epsilon)
|
|
|
|
def initialize_weights(net_l, scale=1):
|
|
if not isinstance(net_l, list):
|
|
net_l = [net_l]
|
|
for net in net_l:
|
|
for m in net.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
init.kaiming_normal_(m.weight, a=0, mode='fan_in')
|
|
m.weight.data *= scale # for residual block
|
|
if m.bias is not None:
|
|
m.bias.data.zero_()
|
|
elif isinstance(m, nn.Linear):
|
|
init.kaiming_normal_(m.weight, a=0, mode='fan_in')
|
|
m.weight.data *= scale
|
|
if m.bias is not None:
|
|
m.bias.data.zero_()
|
|
elif isinstance(m, nn.BatchNorm2d):
|
|
init.constant_(m.weight, 1)
|
|
init.constant_(m.bias.data, 0.0)
|
|
|
|
|
|
def make_layer(block, n_layers, return_layers=False):
|
|
layers = []
|
|
for _ in range(n_layers):
|
|
layers.append(block())
|
|
if return_layers:
|
|
return nn.Sequential(*layers), layers
|
|
else:
|
|
return nn.Sequential(*layers)
|
|
|
|
|
|
class ResidualBlock(nn.Module):
|
|
'''Residual block with BN
|
|
---Conv-BN-ReLU-Conv-+-
|
|
|________________|
|
|
'''
|
|
|
|
def __init__(self, nf=64):
|
|
super(ResidualBlock, self).__init__()
|
|
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
|
|
self.conv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
|
self.BN1 = nn.BatchNorm2d(nf)
|
|
self.conv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
|
self.BN2 = nn.BatchNorm2d(nf)
|
|
|
|
# initialization
|
|
initialize_weights([self.conv1, self.conv2], 0.1)
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
out = self.lrelu(self.BN1(self.conv1(x)))
|
|
out = self.BN2(self.conv2(out))
|
|
return identity + out
|
|
|
|
class ResidualBlockSpectralNorm(nn.Module):
|
|
'''Residual block with Spectral Normalization.
|
|
---SpecConv-ReLU-SpecConv-+-
|
|
|________________|
|
|
'''
|
|
|
|
def __init__(self, nf, total_residual_blocks):
|
|
super(ResidualBlockSpectralNorm, self).__init__()
|
|
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
|
|
self.conv1 = SpectralNorm(nn.Conv2d(nf, nf, 3, 1, 1, bias=True))
|
|
self.conv2 = SpectralNorm(nn.Conv2d(nf, nf, 3, 1, 1, bias=True))
|
|
|
|
initialize_weights([self.conv1, self.conv2], 1)
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
out = self.lrelu(self.conv1(x))
|
|
out = self.conv2(out)
|
|
return identity + out
|
|
|
|
class ResidualBlock_noBN(nn.Module):
|
|
'''Residual block w/o BN
|
|
---Conv-ReLU-Conv-+-
|
|
|________________|
|
|
'''
|
|
|
|
def __init__(self, nf=64):
|
|
super(ResidualBlock_noBN, self).__init__()
|
|
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
|
|
self.conv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
|
self.conv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
|
|
|
# initialization
|
|
initialize_weights([self.conv1, self.conv2], 0.1)
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
out = self.lrelu(self.conv1(x))
|
|
out = self.conv2(out)
|
|
return identity + out
|
|
|
|
|
|
def flow_warp(x, flow, interp_mode='bilinear', padding_mode='zeros'):
|
|
"""Warp an image or feature map with optical flow
|
|
Args:
|
|
x (Tensor): size (N, C, H, W)
|
|
flow (Tensor): size (N, H, W, 2), normal value
|
|
interp_mode (str): 'nearest' or 'bilinear'
|
|
padding_mode (str): 'zeros' or 'border' or 'reflection'
|
|
|
|
Returns:
|
|
Tensor: warped image or feature map
|
|
"""
|
|
assert x.size()[-2:] == flow.size()[1:3]
|
|
B, C, H, W = x.size()
|
|
# mesh grid
|
|
grid_y, grid_x = torch.meshgrid(torch.arange(0, H), torch.arange(0, W))
|
|
grid = torch.stack((grid_x, grid_y), 2).float() # W(x), H(y), 2
|
|
grid.requires_grad = False
|
|
grid = grid.type_as(x)
|
|
vgrid = grid + flow
|
|
# scale grid to [-1,1]
|
|
vgrid_x = 2.0 * vgrid[:, :, :, 0] / max(W - 1, 1) - 1.0
|
|
vgrid_y = 2.0 * vgrid[:, :, :, 1] / max(H - 1, 1) - 1.0
|
|
vgrid_scaled = torch.stack((vgrid_x, vgrid_y), dim=3)
|
|
output = F.grid_sample(x, vgrid_scaled, mode=interp_mode, padding_mode=padding_mode)
|
|
return output
|