|
|
|
|
|
|
|
|
|
import numpy as np |
|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
from kornia.geometry.transform import rotate |
|
import torch.fft as fft |
|
from icecream import ic |
|
import PIL |
|
|
|
def save_image_grid(feats, fname, gridsize): |
|
gw, gh = gridsize |
|
idx = gw * gh |
|
|
|
max_num = torch.max(feats[:idx]).item() |
|
min_num = torch.min(feats[:idx]).item() |
|
feats = feats[:idx].cpu() * 255 / (max_num - min_num) |
|
feats = np.asarray(feats, dtype=np.float32) |
|
feats = np.rint(feats).clip(0, 255).astype(np.uint8) |
|
|
|
C, H, W = feats.shape |
|
|
|
feats = feats.reshape(gh, gw, 1, H, W) |
|
feats = feats.transpose(0, 3, 1, 4, 2) |
|
feats = feats.reshape(gh * H, gw * W, 1) |
|
feats = np.stack([feats]*3, axis=2).squeeze() * 10 |
|
feats = np.rint(feats).clip(0, 255).astype(np.uint8) |
|
|
|
from icecream import ic |
|
ic(feats.shape) |
|
|
|
feats = PIL.Image.fromarray(feats) |
|
feats.save(fname + '.png') |
|
|
|
def _conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1): |
|
return F.conv2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) |
|
|
|
class LearnableSpatialTransformWrapper(nn.Module): |
|
def __init__(self, impl, pad_coef=0.5, angle_init_range=80, train_angle=True): |
|
super().__init__() |
|
self.impl = impl |
|
self.angle = torch.rand(1) * angle_init_range |
|
if train_angle: |
|
self.angle = nn.Parameter(self.angle, requires_grad=True) |
|
self.pad_coef = pad_coef |
|
|
|
def forward(self, x): |
|
if torch.is_tensor(x): |
|
return self.inverse_transform(self.impl(self.transform(x)), x) |
|
elif isinstance(x, tuple): |
|
x_trans = tuple(self.transform(elem) for elem in x) |
|
y_trans = self.impl(x_trans) |
|
return tuple(self.inverse_transform(elem, orig_x) for elem, orig_x in zip(y_trans, x)) |
|
else: |
|
raise ValueError(f'Unexpected input type {type(x)}') |
|
|
|
def transform(self, x): |
|
height, width = x.shape[2:] |
|
pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef) |
|
x_padded = F.pad(x, [pad_w, pad_w, pad_h, pad_h], mode='reflect') |
|
x_padded_rotated = rotate(x_padded, angle=self.angle.to(x_padded)) |
|
return x_padded_rotated |
|
|
|
def inverse_transform(self, y_padded_rotated, orig_x): |
|
height, width = orig_x.shape[2:] |
|
pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef) |
|
|
|
y_padded = rotate(y_padded_rotated, angle=-self.angle.to(y_padded_rotated)) |
|
y_height, y_width = y_padded.shape[2:] |
|
y = y_padded[:, :, pad_h : y_height - pad_h, pad_w : y_width - pad_w] |
|
return y |
|
|
|
|
|
class SELayer(nn.Module): |
|
def __init__(self, channel, reduction=16): |
|
super(SELayer, self).__init__() |
|
self.avg_pool = nn.AdaptiveAvgPool2d(1) |
|
self.fc = nn.Sequential( |
|
nn.Linear(channel, channel // reduction, bias=False), |
|
nn.ReLU(inplace=False), |
|
nn.Linear(channel // reduction, channel, bias=False), |
|
nn.Sigmoid() |
|
) |
|
|
|
def forward(self, x): |
|
b, c, _, _ = x.size() |
|
y = self.avg_pool(x).view(b, c) |
|
y = self.fc(y).view(b, c, 1, 1) |
|
res = x * y.expand_as(x) |
|
return res |
|
|
|
|
|
class FourierUnit(nn.Module): |
|
|
|
def __init__(self, in_channels, out_channels, groups=1, spatial_scale_factor=None, spatial_scale_mode='bilinear', |
|
spectral_pos_encoding=False, use_se=False, se_kwargs=None, ffc3d=False, fft_norm='ortho'): |
|
|
|
super(FourierUnit, self).__init__() |
|
self.groups = groups |
|
|
|
self.conv_layer = torch.nn.Conv2d(in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0), |
|
out_channels=out_channels * 2, |
|
kernel_size=1, stride=1, padding=0, groups=self.groups, bias=False) |
|
self.relu = torch.nn.ReLU(inplace=False) |
|
|
|
|
|
self.use_se = use_se |
|
if use_se: |
|
if se_kwargs is None: |
|
se_kwargs = {} |
|
self.se = SELayer(self.conv_layer.in_channels, **se_kwargs) |
|
|
|
self.spatial_scale_factor = spatial_scale_factor |
|
self.spatial_scale_mode = spatial_scale_mode |
|
self.spectral_pos_encoding = spectral_pos_encoding |
|
self.ffc3d = ffc3d |
|
self.fft_norm = fft_norm |
|
|
|
def forward(self, x): |
|
batch = x.shape[0] |
|
|
|
if self.spatial_scale_factor is not None: |
|
orig_size = x.shape[-2:] |
|
x = F.interpolate(x, scale_factor=self.spatial_scale_factor, mode=self.spatial_scale_mode, align_corners=False) |
|
|
|
r_size = x.size() |
|
|
|
fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1) |
|
ffted = fft.rfftn(x, dim=fft_dim, norm=self.fft_norm) |
|
ffted = torch.stack((ffted.real, ffted.imag), dim=-1) |
|
ffted = ffted.permute(0, 1, 4, 2, 3).contiguous() |
|
ffted = ffted.view((batch, -1,) + ffted.size()[3:]) |
|
|
|
if self.spectral_pos_encoding: |
|
height, width = ffted.shape[-2:] |
|
coords_vert = torch.linspace(0, 1, height)[None, None, :, None].expand(batch, 1, height, width).to(ffted) |
|
coords_hor = torch.linspace(0, 1, width)[None, None, None, :].expand(batch, 1, height, width).to(ffted) |
|
ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1) |
|
|
|
if self.use_se: |
|
ffted = self.se(ffted) |
|
|
|
ffted = self.conv_layer(ffted) |
|
ffted = self.relu(ffted) |
|
|
|
ffted = ffted.view((batch, -1, 2,) + ffted.size()[2:]).permute( |
|
0, 1, 3, 4, 2).contiguous() |
|
ffted = torch.complex(ffted[..., 0], ffted[..., 1]) |
|
|
|
ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:] |
|
output = torch.fft.irfftn(ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm) |
|
|
|
if self.spatial_scale_factor is not None: |
|
output = F.interpolate(output, size=orig_size, mode=self.spatial_scale_mode, align_corners=False) |
|
|
|
return output |
|
|
|
|
|
class SpectralTransform(nn.Module): |
|
|
|
def __init__(self, in_channels, out_channels, stride=1, groups=1, enable_lfu=True, **fu_kwargs): |
|
|
|
super(SpectralTransform, self).__init__() |
|
self.enable_lfu = enable_lfu |
|
if stride == 2: |
|
self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2) |
|
else: |
|
self.downsample = nn.Identity() |
|
|
|
self.stride = stride |
|
self.conv1 = nn.Sequential( |
|
nn.Conv2d(in_channels, out_channels // |
|
2, kernel_size=1, groups=groups, bias=False), |
|
|
|
nn.ReLU(inplace=True) |
|
) |
|
self.fu = FourierUnit( |
|
out_channels // 2, out_channels // 2, groups, **fu_kwargs) |
|
if self.enable_lfu: |
|
self.lfu = FourierUnit( |
|
out_channels // 2, out_channels // 2, groups) |
|
self.conv2 = torch.nn.Conv2d( |
|
out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False) |
|
|
|
def forward(self, x): |
|
|
|
x = self.downsample(x) |
|
x = self.conv1(x) |
|
output = self.fu(x) |
|
|
|
if self.enable_lfu: |
|
n, c, h, w = x.shape |
|
split_no = 2 |
|
split_s = h // split_no |
|
xs = torch.cat(torch.split( |
|
x[:, :c // 4], split_s, dim=-2), dim=1).contiguous() |
|
xs = torch.cat(torch.split(xs, split_s, dim=-1), |
|
dim=1).contiguous() |
|
xs = self.lfu(xs) |
|
xs = xs.repeat(1, 1, split_no, split_no).contiguous() |
|
else: |
|
xs = 0 |
|
|
|
output = self.conv2(x + output + xs) |
|
|
|
return output |
|
|
|
class FFC(nn.Module): |
|
|
|
def __init__(self, in_channels, out_channels, kernel_size, |
|
ratio_gin, ratio_gout, stride=1, padding=0, |
|
dilation=1, groups=1, bias=False, enable_lfu=True, |
|
padding_type='reflect', gated=False, **spectral_kwargs): |
|
super(FFC, self).__init__() |
|
|
|
assert stride == 1 or stride == 2, "Stride should be 1 or 2." |
|
self.stride = stride |
|
|
|
in_cg = int(in_channels * ratio_gin) |
|
in_cl = in_channels - in_cg |
|
out_cg = int(out_channels * ratio_gout) |
|
out_cl = out_channels - out_cg |
|
|
|
|
|
|
|
self.ratio_gin = ratio_gin |
|
self.ratio_gout = ratio_gout |
|
self.global_in_num = in_cg |
|
|
|
module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d |
|
self.convl2l = module(in_cl, out_cl, kernel_size, |
|
stride, padding, dilation, groups, bias, padding_mode=padding_type) |
|
module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d |
|
self.convl2g = module(in_cl, out_cg, kernel_size, |
|
stride, padding, dilation, groups, bias, padding_mode=padding_type) |
|
module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d |
|
self.convg2l = module(in_cg, out_cl, kernel_size, |
|
stride, padding, dilation, groups, bias, padding_mode=padding_type) |
|
module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform |
|
self.convg2g = module( |
|
in_cg, out_cg, stride, 1 if groups == 1 else groups // 2, enable_lfu, **spectral_kwargs) |
|
|
|
self.gated = gated |
|
module = nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d |
|
self.gate = module(in_channels, 2, 1) |
|
|
|
def forward(self, x, fname=None): |
|
x_l, x_g = x if type(x) is tuple else (x, 0) |
|
out_xl, out_xg = 0, 0 |
|
|
|
if self.gated: |
|
total_input_parts = [x_l] |
|
if torch.is_tensor(x_g): |
|
total_input_parts.append(x_g) |
|
total_input = torch.cat(total_input_parts, dim=1) |
|
|
|
gates = torch.sigmoid(self.gate(total_input)) |
|
g2l_gate, l2g_gate = gates.chunk(2, dim=1) |
|
else: |
|
g2l_gate, l2g_gate = 1, 1 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
spec_x = self.convg2g(x_g) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if self.ratio_gout != 1: |
|
out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate |
|
if self.ratio_gout != 0: |
|
out_xg = self.convl2g(x_l) * l2g_gate + spec_x |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
return out_xl, out_xg |
|
|
|
class FFC_BN_ACT(nn.Module): |
|
|
|
def __init__(self, in_channels, out_channels, |
|
kernel_size, ratio_gin, ratio_gout, |
|
stride=1, padding=0, dilation=1, groups=1, bias=False, |
|
norm_layer=nn.SyncBatchNorm, activation_layer=nn.Identity, |
|
padding_type='reflect', |
|
enable_lfu=True, **kwargs): |
|
super(FFC_BN_ACT, self).__init__() |
|
self.ffc = FFC(in_channels, out_channels, kernel_size, |
|
ratio_gin, ratio_gout, stride, padding, dilation, |
|
groups, bias, enable_lfu, padding_type=padding_type, **kwargs) |
|
lnorm = nn.Identity if ratio_gout == 1 else norm_layer |
|
gnorm = nn.Identity if ratio_gout == 0 else norm_layer |
|
global_channels = int(out_channels * ratio_gout) |
|
|
|
|
|
|
|
lact = nn.Identity if ratio_gout == 1 else activation_layer |
|
gact = nn.Identity if ratio_gout == 0 else activation_layer |
|
self.act_l = lact(inplace=True) |
|
self.act_g = gact(inplace=True) |
|
|
|
def forward(self, x, fname=None): |
|
x_l, x_g = self.ffc(x, fname=fname,) |
|
x_l = self.act_l(x_l) |
|
x_g = self.act_g(x_g) |
|
return x_l, x_g |
|
|
|
|
|
class FFCResnetBlock(nn.Module): |
|
def __init__(self, dim, padding_type, norm_layer, activation_layer=nn.ReLU, dilation=1, |
|
spatial_transform_kwargs=None, inline=False, ratio_gin=0.75, ratio_gout=0.75): |
|
super().__init__() |
|
self.conv1 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation, |
|
norm_layer=norm_layer, |
|
activation_layer=activation_layer, |
|
padding_type=padding_type, |
|
ratio_gin=ratio_gin, ratio_gout=ratio_gout) |
|
self.conv2 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation, |
|
norm_layer=norm_layer, |
|
activation_layer=activation_layer, |
|
padding_type=padding_type, |
|
ratio_gin=ratio_gin, ratio_gout=ratio_gout) |
|
if spatial_transform_kwargs is not None: |
|
self.conv1 = LearnableSpatialTransformWrapper(self.conv1, **spatial_transform_kwargs) |
|
self.conv2 = LearnableSpatialTransformWrapper(self.conv2, **spatial_transform_kwargs) |
|
self.inline = inline |
|
|
|
def forward(self, x, fname=None): |
|
if self.inline: |
|
x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:] |
|
else: |
|
x_l, x_g = x if type(x) is tuple else (x, 0) |
|
|
|
id_l, id_g = x_l, x_g |
|
|
|
x_l, x_g = self.conv1((x_l, x_g), fname=fname) |
|
x_l, x_g = self.conv2((x_l, x_g), fname=fname) |
|
|
|
x_l, x_g = id_l + x_l, id_g + x_g |
|
out = x_l, x_g |
|
if self.inline: |
|
out = torch.cat(out, dim=1) |
|
return out |
|
|
|
class ConcatTupleLayer(nn.Module): |
|
def forward(self, x): |
|
assert isinstance(x, tuple) |
|
x_l, x_g = x |
|
assert torch.is_tensor(x_l) or torch.is_tensor(x_g) |
|
if not torch.is_tensor(x_g): |
|
return x_l |
|
return torch.cat(x, dim=1) |
|
|