Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,468 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import sys
|
| 2 |
+
sys.path.append('SAFMN')
|
| 3 |
+
|
| 4 |
+
import os
|
| 5 |
+
import cv2
|
| 6 |
+
import argparse
|
| 7 |
+
import glob
|
| 8 |
+
import numpy as np
|
| 9 |
+
import os
|
| 10 |
+
import torch
|
| 11 |
+
import torch.nn as nn
|
| 12 |
+
import torch.nn.functional as F
|
| 13 |
+
import gradio as gr
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
########################################## Wavelet colorfix ###################################
|
| 17 |
+
from PIL import Image
|
| 18 |
+
from torch import Tensor
|
| 19 |
+
from torchvision.transforms import ToTensor, ToPILImage
|
| 20 |
+
def adain_color_fix(target: Image, source: Image):
|
| 21 |
+
# Convert images to tensors
|
| 22 |
+
to_tensor = ToTensor()
|
| 23 |
+
target_tensor = to_tensor(target).unsqueeze(0)
|
| 24 |
+
source_tensor = to_tensor(source).unsqueeze(0)
|
| 25 |
+
|
| 26 |
+
# Apply adaptive instance normalization
|
| 27 |
+
result_tensor = adaptive_instance_normalization(target_tensor, source_tensor)
|
| 28 |
+
|
| 29 |
+
# Convert tensor back to image
|
| 30 |
+
to_image = ToPILImage()
|
| 31 |
+
result_image = to_image(result_tensor.squeeze(0).clamp_(0.0, 1.0))
|
| 32 |
+
|
| 33 |
+
return result_image
|
| 34 |
+
|
| 35 |
+
def wavelet_color_fix(target: Image, source: Image):
|
| 36 |
+
if target.size() != source.size():
|
| 37 |
+
source = source.resize((target.size()[-2], target.size()[-1]), Image.LANCZOS)
|
| 38 |
+
# Convert images to tensors
|
| 39 |
+
to_tensor = ToTensor()
|
| 40 |
+
target_tensor = to_tensor(target).unsqueeze(0)
|
| 41 |
+
source_tensor = to_tensor(source).unsqueeze(0)
|
| 42 |
+
|
| 43 |
+
# Apply wavelet reconstruction
|
| 44 |
+
result_tensor = wavelet_reconstruction(target_tensor, source_tensor)
|
| 45 |
+
|
| 46 |
+
# Convert tensor back to image
|
| 47 |
+
to_image = ToPILImage()
|
| 48 |
+
result_image = to_image(result_tensor.squeeze(0).clamp_(0.0, 1.0))
|
| 49 |
+
|
| 50 |
+
return result_image
|
| 51 |
+
|
| 52 |
+
def calc_mean_std(feat: Tensor, eps=1e-5):
|
| 53 |
+
"""Calculate mean and std for adaptive_instance_normalization.
|
| 54 |
+
Args:
|
| 55 |
+
feat (Tensor): 4D tensor.
|
| 56 |
+
eps (float): A small value added to the variance to avoid
|
| 57 |
+
divide-by-zero. Default: 1e-5.
|
| 58 |
+
"""
|
| 59 |
+
size = feat.size()
|
| 60 |
+
assert len(size) == 4, 'The input feature should be 4D tensor.'
|
| 61 |
+
b, c = size[:2]
|
| 62 |
+
feat_var = feat.view(b, c, -1).var(dim=2) + eps
|
| 63 |
+
feat_std = feat_var.sqrt().view(b, c, 1, 1)
|
| 64 |
+
feat_mean = feat.view(b, c, -1).mean(dim=2).view(b, c, 1, 1)
|
| 65 |
+
return feat_mean, feat_std
|
| 66 |
+
|
| 67 |
+
def adaptive_instance_normalization(content_feat:Tensor, style_feat:Tensor):
|
| 68 |
+
"""Adaptive instance normalization.
|
| 69 |
+
Adjust the reference features to have the similar color and illuminations
|
| 70 |
+
as those in the degradate features.
|
| 71 |
+
Args:
|
| 72 |
+
content_feat (Tensor): The reference feature.
|
| 73 |
+
style_feat (Tensor): The degradate features.
|
| 74 |
+
"""
|
| 75 |
+
size = content_feat.size()
|
| 76 |
+
style_mean, style_std = calc_mean_std(style_feat)
|
| 77 |
+
content_mean, content_std = calc_mean_std(content_feat)
|
| 78 |
+
normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(size)
|
| 79 |
+
return normalized_feat * style_std.expand(size) + style_mean.expand(size)
|
| 80 |
+
|
| 81 |
+
def wavelet_blur(image: Tensor, radius: int):
|
| 82 |
+
"""
|
| 83 |
+
Apply wavelet blur to the input tensor.
|
| 84 |
+
"""
|
| 85 |
+
# input shape: (1, 3, H, W)
|
| 86 |
+
# convolution kernel
|
| 87 |
+
kernel_vals = [
|
| 88 |
+
[0.0625, 0.125, 0.0625],
|
| 89 |
+
[0.125, 0.25, 0.125],
|
| 90 |
+
[0.0625, 0.125, 0.0625],
|
| 91 |
+
]
|
| 92 |
+
kernel = torch.tensor(kernel_vals, dtype=image.dtype, device=image.device)
|
| 93 |
+
# add channel dimensions to the kernel to make it a 4D tensor
|
| 94 |
+
kernel = kernel[None, None]
|
| 95 |
+
# repeat the kernel across all input channels
|
| 96 |
+
kernel = kernel.repeat(3, 1, 1, 1)
|
| 97 |
+
image = F.pad(image, (radius, radius, radius, radius), mode='replicate')
|
| 98 |
+
# apply convolution
|
| 99 |
+
output = F.conv2d(image, kernel, groups=3, dilation=radius)
|
| 100 |
+
return output
|
| 101 |
+
|
| 102 |
+
def wavelet_decomposition(image: Tensor, levels=5):
|
| 103 |
+
"""
|
| 104 |
+
Apply wavelet decomposition to the input tensor.
|
| 105 |
+
This function only returns the low frequency & the high frequency.
|
| 106 |
+
"""
|
| 107 |
+
high_freq = torch.zeros_like(image)
|
| 108 |
+
for i in range(levels):
|
| 109 |
+
radius = 2 ** i
|
| 110 |
+
low_freq = wavelet_blur(image, radius)
|
| 111 |
+
high_freq += (image - low_freq)
|
| 112 |
+
image = low_freq
|
| 113 |
+
|
| 114 |
+
return high_freq, low_freq
|
| 115 |
+
|
| 116 |
+
def wavelet_reconstruction(content_feat:Tensor, style_feat:Tensor):
|
| 117 |
+
"""
|
| 118 |
+
Apply wavelet decomposition, so that the content will have the same color as the style.
|
| 119 |
+
"""
|
| 120 |
+
# calculate the wavelet decomposition of the content feature
|
| 121 |
+
content_high_freq, content_low_freq = wavelet_decomposition(content_feat)
|
| 122 |
+
del content_low_freq
|
| 123 |
+
# calculate the wavelet decomposition of the style feature
|
| 124 |
+
style_high_freq, style_low_freq = wavelet_decomposition(style_feat)
|
| 125 |
+
del style_high_freq
|
| 126 |
+
# reconstruct the content feature with the style's high frequency
|
| 127 |
+
return content_high_freq + style_low_freq
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
########################################## URL Load ###################################
|
| 131 |
+
from torch.hub import download_url_to_file, get_dir
|
| 132 |
+
from urllib.parse import urlparse
|
| 133 |
+
|
| 134 |
+
def load_file_from_url(url, model_dir=None, progress=True, file_name=None):
|
| 135 |
+
"""Load file form http url, will download models if necessary.
|
| 136 |
+
|
| 137 |
+
Ref:https://github.com/1adrianb/face-alignment/blob/master/face_alignment/utils.py
|
| 138 |
+
|
| 139 |
+
Args:
|
| 140 |
+
url (str): URL to be downloaded.
|
| 141 |
+
model_dir (str): The path to save the downloaded model. Should be a full path. If None, use pytorch hub_dir.
|
| 142 |
+
Default: None.
|
| 143 |
+
progress (bool): Whether to show the download progress. Default: True.
|
| 144 |
+
file_name (str): The downloaded file name. If None, use the file name in the url. Default: None.
|
| 145 |
+
|
| 146 |
+
Returns:
|
| 147 |
+
str: The path to the downloaded file.
|
| 148 |
+
"""
|
| 149 |
+
if model_dir is None: # use the pytorch hub_dir
|
| 150 |
+
hub_dir = get_dir()
|
| 151 |
+
model_dir = os.path.join(hub_dir, 'checkpoints')
|
| 152 |
+
|
| 153 |
+
os.makedirs(model_dir, exist_ok=True)
|
| 154 |
+
|
| 155 |
+
parts = urlparse(url)
|
| 156 |
+
filename = os.path.basename(parts.path)
|
| 157 |
+
if file_name is not None:
|
| 158 |
+
filename = file_name
|
| 159 |
+
cached_file = os.path.abspath(os.path.join(model_dir, filename))
|
| 160 |
+
if not os.path.exists(cached_file):
|
| 161 |
+
print(f'Downloading: "{url}" to {cached_file}\n')
|
| 162 |
+
download_url_to_file(url, cached_file, hash_prefix=None, progress=progress)
|
| 163 |
+
return cached_file
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
########################################## Model Define ###################################
|
| 167 |
+
# Layer Norm
|
| 168 |
+
class LayerNorm(nn.Module):
|
| 169 |
+
def __init__(self, normalized_shape, eps=1e-6, data_format="channels_first"):
|
| 170 |
+
super().__init__()
|
| 171 |
+
self.weight = nn.Parameter(torch.ones(normalized_shape))
|
| 172 |
+
self.bias = nn.Parameter(torch.zeros(normalized_shape))
|
| 173 |
+
self.eps = eps
|
| 174 |
+
self.data_format = data_format
|
| 175 |
+
if self.data_format not in ["channels_last", "channels_first"]:
|
| 176 |
+
raise NotImplementedError
|
| 177 |
+
self.normalized_shape = (normalized_shape, )
|
| 178 |
+
|
| 179 |
+
def forward(self, x):
|
| 180 |
+
if self.data_format == "channels_last":
|
| 181 |
+
return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
|
| 182 |
+
elif self.data_format == "channels_first":
|
| 183 |
+
u = x.mean(1, keepdim=True)
|
| 184 |
+
s = (x - u).pow(2).mean(1, keepdim=True)
|
| 185 |
+
x = (x - u) / torch.sqrt(s + self.eps)
|
| 186 |
+
x = self.weight[:, None, None] * x + self.bias[:, None, None]
|
| 187 |
+
return x
|
| 188 |
+
|
| 189 |
+
# CCM
|
| 190 |
+
class CCM(nn.Module):
|
| 191 |
+
def __init__(self, dim, growth_rate=2.0):
|
| 192 |
+
super().__init__()
|
| 193 |
+
hidden_dim = int(dim * growth_rate)
|
| 194 |
+
|
| 195 |
+
self.ccm = nn.Sequential(
|
| 196 |
+
nn.Conv2d(dim, hidden_dim, 3, 1, 1),
|
| 197 |
+
nn.GELU(),
|
| 198 |
+
nn.Conv2d(hidden_dim, dim, 1, 1, 0)
|
| 199 |
+
)
|
| 200 |
+
|
| 201 |
+
def forward(self, x):
|
| 202 |
+
return self.ccm(x)
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
# SAFM
|
| 206 |
+
class SAFM(nn.Module):
|
| 207 |
+
def __init__(self, dim, n_levels=4):
|
| 208 |
+
super().__init__()
|
| 209 |
+
self.n_levels = n_levels
|
| 210 |
+
chunk_dim = dim // n_levels
|
| 211 |
+
|
| 212 |
+
# Spatial Weighting
|
| 213 |
+
self.mfr = nn.ModuleList([nn.Conv2d(chunk_dim, chunk_dim, 3, 1, 1, groups=chunk_dim) for i in range(self.n_levels)])
|
| 214 |
+
|
| 215 |
+
# # Feature Aggregation
|
| 216 |
+
self.aggr = nn.Conv2d(dim, dim, 1, 1, 0)
|
| 217 |
+
|
| 218 |
+
# Activation
|
| 219 |
+
self.act = nn.GELU()
|
| 220 |
+
|
| 221 |
+
def forward(self, x):
|
| 222 |
+
h, w = x.size()[-2:]
|
| 223 |
+
|
| 224 |
+
xc = x.chunk(self.n_levels, dim=1)
|
| 225 |
+
out = []
|
| 226 |
+
for i in range(self.n_levels):
|
| 227 |
+
if i > 0:
|
| 228 |
+
p_size = (h//2**i, w//2**i)
|
| 229 |
+
s = F.adaptive_max_pool2d(xc[i], p_size)
|
| 230 |
+
s = self.mfr[i](s)
|
| 231 |
+
s = F.interpolate(s, size=(h, w), mode='nearest')
|
| 232 |
+
else:
|
| 233 |
+
s = self.mfr[i](xc[i])
|
| 234 |
+
out.append(s)
|
| 235 |
+
|
| 236 |
+
out = self.aggr(torch.cat(out, dim=1))
|
| 237 |
+
out = self.act(out) * x
|
| 238 |
+
return out
|
| 239 |
+
|
| 240 |
+
class AttBlock(nn.Module):
|
| 241 |
+
def __init__(self, dim, ffn_scale=2.0):
|
| 242 |
+
super().__init__()
|
| 243 |
+
|
| 244 |
+
self.norm1 = LayerNorm(dim)
|
| 245 |
+
self.norm2 = LayerNorm(dim)
|
| 246 |
+
|
| 247 |
+
# Multiscale Block
|
| 248 |
+
self.safm = SAFM(dim)
|
| 249 |
+
# Feedforward layer
|
| 250 |
+
self.ccm = CCM(dim, ffn_scale)
|
| 251 |
+
|
| 252 |
+
def forward(self, x):
|
| 253 |
+
x = self.safm(self.norm1(x)) + x
|
| 254 |
+
x = self.ccm(self.norm2(x)) + x
|
| 255 |
+
return x
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
class SAFMN(nn.Module):
|
| 259 |
+
def __init__(self, dim, n_blocks=8, ffn_scale=2.0, upscaling_factor=4):
|
| 260 |
+
super().__init__()
|
| 261 |
+
self.to_feat = nn.Conv2d(3, dim, 3, 1, 1)
|
| 262 |
+
|
| 263 |
+
self.feats = nn.Sequential(*[AttBlock(dim, ffn_scale) for _ in range(n_blocks)])
|
| 264 |
+
|
| 265 |
+
self.to_img = nn.Sequential(
|
| 266 |
+
nn.Conv2d(dim, 3 * upscaling_factor**2, 3, 1, 1),
|
| 267 |
+
nn.PixelShuffle(upscaling_factor)
|
| 268 |
+
)
|
| 269 |
+
|
| 270 |
+
def forward(self, x):
|
| 271 |
+
x = self.to_feat(x)
|
| 272 |
+
x = self.feats(x) + x
|
| 273 |
+
x = self.to_img(x)
|
| 274 |
+
return x
|
| 275 |
+
|
| 276 |
+
########################################## Gradio inference ###################################
|
| 277 |
+
pretrain_model_url = {
|
| 278 |
+
'safmn_x2': 'https://github.com/sunny2109/SAFMN/releases/download/v0.1.0/SAFMN_L_Real_LSDIR_x2-v2.pth',
|
| 279 |
+
'safmn_x4': 'https://github.com/sunny2109/SAFMN/releases/download/v0.1.0/SAFMN_L_Real_LSDIR_x4-v2.pth',
|
| 280 |
+
}
|
| 281 |
+
|
| 282 |
+
|
| 283 |
+
# download weights
|
| 284 |
+
if not os.path.exists('./experiments/pretrained_models/SAFMN_L_Real_LSDIR_x2-v2.pth'):
|
| 285 |
+
load_file_from_url(url=pretrain_model_url['safmn_x2'], model_dir='./experiments/pretrained_models/', progress=True, file_name=None)
|
| 286 |
+
|
| 287 |
+
if not os.path.exists('./experiments/pretrained_models/SAFMN_L_Real_LSDIR_x4-v2.pth'):
|
| 288 |
+
load_file_from_url(url=pretrain_model_url['safmn_x4'], model_dir='./experiments/pretrained_models/', progress=True, file_name=None)
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
| 292 |
+
|
| 293 |
+
def set_safmn(upscale):
|
| 294 |
+
model = SAFMN(dim=128, n_blocks=16, ffn_scale=2.0, upscaling_factor=upscale)
|
| 295 |
+
if upscale == 2:
|
| 296 |
+
model_path = './experiments/pretrained_models/SAFMN_L_Real_LSDIR_x2.pth'
|
| 297 |
+
elif upscale == 4:
|
| 298 |
+
model_path = './experiments/pretrained_models/SAFMN_L_Real_LSDIR_x4-v2.pth'
|
| 299 |
+
else:
|
| 300 |
+
raise NotImplementedError('Only support x2/x4 upscaling!')
|
| 301 |
+
|
| 302 |
+
model.load_state_dict(torch.load(model_path)['params'], strict=True)
|
| 303 |
+
model.eval()
|
| 304 |
+
return model.to(device)
|
| 305 |
+
|
| 306 |
+
|
| 307 |
+
def img2patch(lq, scale=4, crop_size=512):
|
| 308 |
+
b, c, hl, wl = lq.size()
|
| 309 |
+
h, w = hl*scale, wl*scale
|
| 310 |
+
sr_size = (b, c, h, w)
|
| 311 |
+
assert b == 1
|
| 312 |
+
|
| 313 |
+
crop_size_h, crop_size_w = crop_size // scale * scale, crop_size // scale * scale
|
| 314 |
+
|
| 315 |
+
#adaptive step_i, step_j
|
| 316 |
+
num_row = (h - 1) // crop_size_h + 1
|
| 317 |
+
num_col = (w - 1) // crop_size_w + 1
|
| 318 |
+
|
| 319 |
+
import math
|
| 320 |
+
step_j = crop_size_w if num_col == 1 else math.ceil((w - crop_size_w) / (num_col - 1) - 1e-8)
|
| 321 |
+
step_i = crop_size_h if num_row == 1 else math.ceil((h - crop_size_h) / (num_row - 1) - 1e-8)
|
| 322 |
+
|
| 323 |
+
step_i = step_i // scale * scale
|
| 324 |
+
step_j = step_j // scale * scale
|
| 325 |
+
|
| 326 |
+
parts = []
|
| 327 |
+
idxes = []
|
| 328 |
+
|
| 329 |
+
i = 0 # 0~h-1
|
| 330 |
+
last_i = False
|
| 331 |
+
while i < h and not last_i:
|
| 332 |
+
j = 0
|
| 333 |
+
if i + crop_size_h >= h:
|
| 334 |
+
i = h - crop_size_h
|
| 335 |
+
last_i = True
|
| 336 |
+
|
| 337 |
+
last_j = False
|
| 338 |
+
while j < w and not last_j:
|
| 339 |
+
if j + crop_size_w >= w:
|
| 340 |
+
j = w - crop_size_w
|
| 341 |
+
last_j = True
|
| 342 |
+
parts.append(lq[:, :, i // scale :(i + crop_size_h) // scale, j // scale:(j + crop_size_w) // scale])
|
| 343 |
+
idxes.append({'i': i, 'j': j})
|
| 344 |
+
j = j + step_j
|
| 345 |
+
i = i + step_i
|
| 346 |
+
|
| 347 |
+
return torch.cat(parts, dim=0), idxes, sr_size
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def patch2img(outs, idxes, sr_size, scale=4, crop_size=512):
|
| 351 |
+
preds = torch.zeros(sr_size).to(outs.device)
|
| 352 |
+
b, c, h, w = sr_size
|
| 353 |
+
|
| 354 |
+
count_mt = torch.zeros((b, 1, h, w)).to(outs.device)
|
| 355 |
+
crop_size_h, crop_size_w = crop_size // scale * scale, crop_size // scale * scale
|
| 356 |
+
|
| 357 |
+
for cnt, each_idx in enumerate(idxes):
|
| 358 |
+
i = each_idx['i']
|
| 359 |
+
j = each_idx['j']
|
| 360 |
+
preds[0, :, i: i + crop_size_h, j: j + crop_size_w] += outs[cnt]
|
| 361 |
+
count_mt[0, 0, i: i + crop_size_h, j: j + crop_size_w] += 1.
|
| 362 |
+
|
| 363 |
+
return (preds / count_mt).to(outs.device)
|
| 364 |
+
|
| 365 |
+
|
| 366 |
+
os.makedirs('./results', exist_ok=True)
|
| 367 |
+
|
| 368 |
+
def inference(image, upscale, large_input_flag, color_fix):
|
| 369 |
+
upscale = int(upscale) # convert type to int
|
| 370 |
+
if upscale > 4:
|
| 371 |
+
upscale = 4
|
| 372 |
+
if 0 < upscale < 3:
|
| 373 |
+
upscale = 2
|
| 374 |
+
|
| 375 |
+
model = set_safmn(upscale)
|
| 376 |
+
|
| 377 |
+
img = cv2.imread(str(image), cv2.IMREAD_COLOR)
|
| 378 |
+
print(f'input size: {img.shape}')
|
| 379 |
+
|
| 380 |
+
# img2tensor
|
| 381 |
+
img = img.astype(np.float32) / 255.
|
| 382 |
+
img = torch.from_numpy(np.transpose(img[:, :, [2, 1, 0]], (2, 0, 1))).float()
|
| 383 |
+
img = img.unsqueeze(0).to(device)
|
| 384 |
+
|
| 385 |
+
# inference
|
| 386 |
+
if large_input_flag:
|
| 387 |
+
patches, idx, size = img2patch(img, scale=upscale)
|
| 388 |
+
with torch.no_grad():
|
| 389 |
+
n = len(patches)
|
| 390 |
+
outs = []
|
| 391 |
+
m = 1
|
| 392 |
+
i = 0
|
| 393 |
+
while i < n:
|
| 394 |
+
j = i + m
|
| 395 |
+
if j >= n:
|
| 396 |
+
j = n
|
| 397 |
+
pred = output = model(patches[i:j])
|
| 398 |
+
if isinstance(pred, list):
|
| 399 |
+
pred = pred[-1]
|
| 400 |
+
outs.append(pred.detach())
|
| 401 |
+
i = j
|
| 402 |
+
output = torch.cat(outs, dim=0)
|
| 403 |
+
|
| 404 |
+
output = patch2img(output, idx, size, scale=upscale)
|
| 405 |
+
else:
|
| 406 |
+
with torch.no_grad():
|
| 407 |
+
output = model(img)
|
| 408 |
+
|
| 409 |
+
# color fix
|
| 410 |
+
if color_fix:
|
| 411 |
+
img = F.interpolate(img, scale_factor=upscale, mode='bilinear')
|
| 412 |
+
output = wavelet_reconstruction(output, img)
|
| 413 |
+
# tensor2img
|
| 414 |
+
output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
|
| 415 |
+
if output.ndim == 3:
|
| 416 |
+
output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0))
|
| 417 |
+
output = (output * 255.0).round().astype(np.uint8)
|
| 418 |
+
|
| 419 |
+
# save restored img
|
| 420 |
+
save_path = f'results/out.png'
|
| 421 |
+
cv2.imwrite(save_path, output)
|
| 422 |
+
|
| 423 |
+
output = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
|
| 424 |
+
return output, save_path
|
| 425 |
+
|
| 426 |
+
|
| 427 |
+
|
| 428 |
+
title = "Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution"
|
| 429 |
+
description = r"""
|
| 430 |
+
<b>Official Gradio demo</b> for <a href='https://github.com/sunny2109/SAFMN' target='_blank'><b>Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution (ICCV 2023)</b></a>.<br>
|
| 431 |
+
"""
|
| 432 |
+
article = r"""
|
| 433 |
+
If SAFMN is helpful, please help to ⭐ the <a href='https://github.com/sunny2109/SAFMN' target='_blank'>Github Repo</a>. Thanks!
|
| 434 |
+
[](https://github.com/sunny2109/SAFMN)
|
| 435 |
+
|
| 436 |
+
---
|
| 437 |
+
📝 **Citation**
|
| 438 |
+
|
| 439 |
+
If our work is useful for your research, please consider citing:
|
| 440 |
+
```bibtex
|
| 441 |
+
@inproceedings{sun2023safmn,
|
| 442 |
+
title={Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution},
|
| 443 |
+
author={Sun, Long and Dong, Jiangxin and Tang, Jinhui and Pan, Jinshan},
|
| 444 |
+
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
|
| 445 |
+
year={2023}
|
| 446 |
+
}
|
| 447 |
+
```
|
| 448 |
+
|
| 449 |
+
<center><img src='https://visitor-badge.laobi.icu/badge?page_id=sunny2109/SAFMN' alt='visitors'></center>
|
| 450 |
+
"""
|
| 451 |
+
|
| 452 |
+
demo = gr.Interface(
|
| 453 |
+
inference, [
|
| 454 |
+
gr.inputs.Image(type="filepath", label="Input"),
|
| 455 |
+
gr.inputs.Number(default=2, label="Upscaling factor (up to 4)"),
|
| 456 |
+
gr.inputs.Checkbox(default=False, label="Memory-efficient inference"),
|
| 457 |
+
gr.inputs.Checkbox(default=False, label="Color correction"),
|
| 458 |
+
], [
|
| 459 |
+
gr.outputs.Image(type="numpy", label="Output"),
|
| 460 |
+
gr.outputs.File(label="Download the output")
|
| 461 |
+
],
|
| 462 |
+
title=title,
|
| 463 |
+
description=description,
|
| 464 |
+
article=article,
|
| 465 |
+
)
|
| 466 |
+
|
| 467 |
+
demo.queue(concurrency_count=2)
|
| 468 |
+
demo.launch()
|