图像修复是一个从低质量图像中恢复高质量图像的任务。以往通过GAN等方式的图像修复大多是以卷积网络进行的，本文讲解一个基于Swin-Transformer的SwinIR模型，包括：浅层特征提取、深层特征提取和高质量图像重建。
好处有两个：(1)Transformer 可以有效捕捉长距离依赖，Swin Transformer 将自注意力计算限制在分割的不重叠窗口内从而降低计算量；(2) 使用 CNN 在 Transformer Layer 后避免原论文中的层级结构，实现即插即用，同时在 Transformer 中 CNN 可以稳定训练过程与融合特征。
模型细节：
主要由Swin transformer Bolck和RSTB块组成：
对于Swin transformer Block：
class Mlp(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
def window_partition(x, window_size):
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
return windows
def window_reverse(windows, window_size, H, W):
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class WindowAttention(nn.Module):
def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter(
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
trunc_normal_(self.relative_position_bias_table, std=.02)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
B_, N, C = x.shape
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
q = q * self.scale
attn = (q @ k.transpose(-2, -1))
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(-1, self.num_heads, N, N)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
def extra_repr(self) -> str:
return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class SwinTransformerBlock(nn.Module):
def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
self.norm1 = norm_layer(dim)
self.attn = WindowAttention(
dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
if self.shift_size > 0:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
h_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
w_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
cnt = 0
for h in h_slices:
for w in w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
return attn_mask
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
# assert L == H * W, "input feature has wrong size"
shortcut = x
x = self.norm1(x)
x = x.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
# partition windows
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
else:
attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
else:
x = shifted_x
x = x.view(B, H * W, C)
# FFN
x = shortcut + self.drop_path(x)
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
def flops(self):
flops = 0
H, W = self.input_resolution
# norm1
flops += self.dim * H * W
# W-MSA/SW-MSA
nW = H * W / self.window_size / self.window_size
flops += nW * self.attn.flops(self.window_size * self.window_size)
# mlp
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# norm2
flops += self.dim * H * W
return flops
class PatchMerging(nn.Module):
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
x = self.norm(x)
x = self.reduction(x)
return x
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
class BasicLayer(nn.Module):
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList([
SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
num_heads=num_heads, window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
norm_layer=norm_layer)
for i in range(depth)])
# patch merging layer
if downsample is not None:
self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
else:
self.downsample = None
def forward(self, x, x_size):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
if self.downsample is not None:
x = self.downsample(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
def flops(self):
flops = 0
for blk in self.blocks:
flops += blk.flops()
if self.downsample is not None:
flops += self.downsample.flops()
return flops然后是 RSTB：
class RSTB(nn.Module):
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
img_size=224, patch_size=4, resi_connection='1conv'):
super(RSTB, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = BasicLayer(dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint)
if resi_connection == '1conv':
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1))
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
def forward(self, x, x_size):
return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
def flops(self):
flops = 0
flops += self.residual_group.flops()
H, W = self.input_resolution
flops += H * W * self.dim * self.dim * 9
flops += self.patch_embed.flops()
flops += self.patch_unembed.flops()
return flops
class PatchEmbed(nn.Module):
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
flops = 0
H, W = self.img_size
if self.norm is not None:
flops += H * W * self.embed_dim
return flops
class PatchUnEmbed(nn.Module):
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
class Upsample(nn.Sequential):
def __init__(self, scale, num_feat):
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
super(Upsample, self).__init__(*m)
class UpsampleOneStep(nn.Sequential):
def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
self.num_feat = num_feat
self.input_resolution = input_resolution
m = []
m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
super(UpsampleOneStep, self).__init__(*m)
def flops(self):
H, W = self.input_resolution
flops = H * W * self.num_feat * 3 * 9
return flops最后实现SwinIR：
class SwinIR(nn.Module):
def __init__(self, img_size=64, patch_size=1, in_chans=3,
embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
**kwargs):
super(SwinIR, self).__init__()
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.img_range = img_range
if in_chans == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
self.upscale = upscale
self.upsampler = upsampler
self.window_size = window_size
self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
trunc_normal_(self.absolute_pos_embed, std=.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
# build Residual Swin Transformer blocks (RSTB)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(dim=embed_dim,
input_resolution=(patches_resolution[0],
patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection
)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
# build the last conv layer in deep feature extraction
if resi_connection == '1conv':
self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
if self.upsampler == 'pixelshuffle':
# for classical SR
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.upsample = Upsample(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
(patches_resolution[0], patches_resolution[1]))
elif self.upsampler == 'nearest+conv':
# for real-world SR (less artifacts)
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
if self.upscale == 4:
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'absolute_pos_embed'}
@torch.jit.ignore
def no_weight_decay_keywords(self):
return {'relative_position_bias_table'}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
x = self.check_image_size(x)
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == 'pixelshuffle':
# for classical SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
elif self.upsampler == 'nearest+conv':
# for real-world SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
if self.upscale == 4:
x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
x = x + self.conv_last(res)
x = x / self.img_range + self.mean
return x[:, :, :H*self.upscale, :W*self.upscale]
def flops(self):
flops = 0
H, W = self.patches_resolution
flops += H * W * 3 * self.embed_dim * 9
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops()
flops += H * W * 3 * self.embed_dim * self.embed_dim
flops += self.upsample.flops()
return flops训练代码：
def main(json_path='options/train_msrresnet_psnr.json'):
'''
# ----------------------------------------
# Step--1 (prepare opt)
# ----------------------------------------
'''
parser = argparse.ArgumentParser()
parser.add_argument('--opt', type=str, default=json_path, help='Path to option JSON file.')
parser.add_argument('--launcher', default='pytorch', help='job launcher')
parser.add_argument('--local_rank', type=int, default=0)
parser.add_argument('--dist', default=False)
opt = option.parse(parser.parse_args().opt, is_train=True)
opt['dist'] = parser.parse_args().dist
# ----------------------------------------
# distributed settings
# ----------------------------------------
if opt['dist']:
init_dist('pytorch')
opt['rank'], opt['world_size'] = get_dist_info()
if opt['rank'] == 0:
util.mkdirs((path for key, path in opt['path'].items() if 'pretrained' not in key))
# ----------------------------------------
# update opt
# ----------------------------------------
# -->-->-->-->-->-->-->-->-->-->-->-->-->-
init_iter_G, init_path_G = option.find_last_checkpoint(opt['path']['models'], net_type='G')
init_iter_E, init_path_E = option.find_last_checkpoint(opt['path']['models'], net_type='E')
opt['path']['pretrained_netG'] = init_path_G
opt['path']['pretrained_netE'] = init_path_E
init_iter_optimizerG, init_path_optimizerG = option.find_last_checkpoint(opt['path']['models'], net_type='optimizerG')
opt['path']['pretrained_optimizerG'] = init_path_optimizerG
current_step = max(init_iter_G, init_iter_E, init_iter_optimizerG)
border = opt['scale']
# --<--<--<--<--<--<--<--<--<--<--<--<--<-
# ----------------------------------------
# save opt to a '../option.json' file
# ----------------------------------------
if opt['rank'] == 0:
option.save(opt)
# ----------------------------------------
# return None for missing key
# ----------------------------------------
opt = option.dict_to_nonedict(opt)
# ----------------------------------------
# configure logger
# ----------------------------------------
if opt['rank'] == 0:
logger_name = 'train'
utils_logger.logger_info(logger_name, os.path.join(opt['path']['log'], logger_name+'.log'))
logger = logging.getLogger(logger_name)
logger.info(option.dict2str(opt))
# ----------------------------------------
# seed
# ----------------------------------------
seed = opt['train']['manual_seed']
if seed is None:
seed = random.randint(1, 10000)
print('Random seed: {}'.format(seed))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
'''
# ----------------------------------------
# Step--2 (creat dataloader)
# ----------------------------------------
'''
# ----------------------------------------
# 1) create_dataset
# 2) creat_dataloader for train and test
# ----------------------------------------
for phase, dataset_opt in opt['datasets'].items():
if phase == 'train':
train_set = define_Dataset(dataset_opt)
train_size = int(math.ceil(len(train_set) / dataset_opt['dataloader_batch_size']))
if opt['rank'] == 0:
logger.info('Number of train images: {:,d}, iters: {:,d}'.format(len(train_set), train_size))
if opt['dist']:
train_sampler = DistributedSampler(train_set, shuffle=dataset_opt['dataloader_shuffle'], drop_last=True, seed=seed)
train_loader = DataLoader(train_set,
batch_size=dataset_opt['dataloader_batch_size']//opt['num_gpu'],
shuffle=False,
num_workers=dataset_opt['dataloader_num_workers']//opt['num_gpu'],
drop_last=True,
pin_memory=True,
sampler=train_sampler)
else:
train_loader = DataLoader(train_set,
batch_size=dataset_opt['dataloader_batch_size'],
shuffle=dataset_opt['dataloader_shuffle'],
num_workers=dataset_opt['dataloader_num_workers'],
drop_last=True,
pin_memory=True)
elif phase == 'test':
test_set = define_Dataset(dataset_opt)
test_loader = DataLoader(test_set, batch_size=1,
shuffle=False, num_workers=1,
drop_last=False, pin_memory=True)
else:
raise NotImplementedError("Phase [%s] is not recognized." % phase)
'''
# ----------------------------------------
# Step--3 (initialize model)
# ----------------------------------------
'''
model = define_Model(opt)
model.init_train()
if opt['rank'] == 0:
logger.info(model.info_network())
logger.info(model.info_params())
'''
# ----------------------------------------
# Step--4 (main training)
# ----------------------------------------
'''
for epoch in range(1000000): # keep running
if opt['dist']:
train_sampler.set_epoch(epoch)
for i, train_data in enumerate(train_loader):
current_step += 1
# -------------------------------
# 1) update learning rate
# -------------------------------
model.update_learning_rate(current_step)
# -------------------------------
# 2) feed patch pairs
# -------------------------------
model.feed_data(train_data)
# -------------------------------
# 3) optimize parameters
# -------------------------------
model.optimize_parameters(current_step)
# -------------------------------
# 4) training information
# -------------------------------
if current_step % opt['train']['checkpoint_print'] == 0 and opt['rank'] == 0:
logs = model.current_log() # such as loss
message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format(epoch, current_step, model.current_learning_rate())
for k, v in logs.items(): # merge log information into message
message += '{:s}: {:.3e} '.format(k, v)
logger.info(message)
# -------------------------------
# 5) save model
# -------------------------------
if current_step % opt['train']['checkpoint_save'] == 0 and opt['rank'] == 0:
logger.info('Saving the model.')
model.save(current_step)
# -------------------------------
# 6) testing
# -------------------------------
if current_step % opt['train']['checkpoint_test'] == 0 and opt['rank'] == 0:
avg_psnr = 0.0
idx = 0
for test_data in test_loader:
idx += 1
image_name_ext = os.path.basename(test_data['L_path'][0])
img_name, ext = os.path.splitext(image_name_ext)
img_dir = os.path.join(opt['path']['images'], img_name)
util.mkdir(img_dir)
model.feed_data(test_data)
model.test()
visuals = model.current_visuals()
E_img = util.tensor2uint(visuals['E'])
H_img = util.tensor2uint(visuals['H'])
# -----------------------
# save estimated image E
# -----------------------
save_img_path = os.path.join(img_dir, '{:s}_{:d}.png'.format(img_name, current_step))
util.imsave(E_img, save_img_path)
# -----------------------
# calculate PSNR
# -----------------------
current_psnr = util.calculate_psnr(E_img, H_img, border=border)
logger.info('{:->4d}--> {:>10s} | {:<4.2f}dB'.format(idx, image_name_ext, current_psnr))
avg_psnr += current_psnr
avg_psnr = avg_psnr / idx
# testing log
logger.info('<epoch:{:3d}, iter:{:8,d}, Average PSNR : {:<.2f}dB\n'.format(epoch, current_step, avg_psnr))
if __name__ == '__main__':
main()要训练SwinIR，对于不同的设置，可能需要更改json文件中的dataroot_H, dataroot_L, scale factor, noisel level, JPEG level, G_optimizer_lr, G_scheduler_milestones等。最好参考源代码说明：KAIR/docs/README_SwinIR.md at master · cszn/KAIR · GitHub
小伙伴们也可以直接下载已经训练好的模型。
最后是生成图像的代码，主要是修改task、scale和model_path这些参数：
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--task', type=str, default='real_sr', help='classical_sr, lightweight_sr, real_sr, '
'gray_dn, color_dn, jpeg_car, color_jpeg_car')
parser.add_argument('--scale', type=int, default=4, help='scale factor: 1, 2, 3, 4, 8') # 1 for dn and jpeg car
parser.add_argument('--noise', type=int, default=15, help='noise level: 15, 25, 50')
parser.add_argument('--jpeg', type=int, default=40, help='scale factor: 10, 20, 30, 40')
parser.add_argument('--training_patch_size', type=int, default=128, help='patch size used in training SwinIR. '
'Just used to differentiate two different settings in Table 2 of the paper. '
'Images are NOT tested patch by patch.')
parser.add_argument('--large_model', action='store_true', help='use large model, only provided for real image sr')
parser.add_argument('--model_path', type=str,
default='model_zoo/swinir/003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_GAN.pth')
parser.add_argument('--folder_lq', type=str, default='testsets/RealSRSet+5images', help='input low-quality test image folder')
parser.add_argument('--folder_gt', type=str, default=None, help='input ground-truth test image folder')
parser.add_argument('--tile', type=int, default=None, help='Tile size, None for no tile during testing (testing as a whole)')
parser.add_argument('--tile_overlap', type=int, default=32, help='Overlapping of different tiles')
args = parser.parse_args()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# set up model
if os.path.exists(args.model_path):
print(f'loading model from {args.model_path}')
else:
os.makedirs(os.path.dirname(args.model_path), exist_ok=True)
url = 'https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/{}'.format(os.path.basename(args.model_path))
r = requests.get(url, allow_redirects=True)
print(f'downloading model {args.model_path}')
open(args.model_path, 'wb').write(r.content)
model = define_model(args)
model.eval()
model = model.to(device)
# setup folder and path
folder, save_dir, border, window_size = setup(args)
os.makedirs(save_dir, exist_ok=True)
test_results = OrderedDict()
test_results['psnr'] = []
test_results['ssim'] = []
test_results['psnr_y'] = []
test_results['ssim_y'] = []
test_results['psnrb'] = []
test_results['psnrb_y'] = []
psnr, ssim, psnr_y, ssim_y, psnrb, psnrb_y = 0, 0, 0, 0, 0, 0
for idx, path in enumerate(sorted(glob.glob(os.path.join(folder, '*')))):
# read image
imgname, img_lq, img_gt = get_image_pair(args, path) # image to HWC-BGR, float32
img_lq = np.transpose(img_lq if img_lq.shape[2] == 1 else img_lq[:, :, [2, 1, 0]], (2, 0, 1)) # HCW-BGR to CHW-RGB
img_lq = torch.from_numpy(img_lq).float().unsqueeze(0).to(device) # CHW-RGB to NCHW-RGB
# inference
with torch.no_grad():
# pad input image to be a multiple of window_size
_, _, h_old, w_old = img_lq.size()
h_pad = (h_old // window_size + 1) * window_size - h_old
w_pad = (w_old // window_size + 1) * window_size - w_old
img_lq = torch.cat([img_lq, torch.flip(img_lq, [2])], 2)[:, :, :h_old + h_pad, :]
img_lq = torch.cat([img_lq, torch.flip(img_lq, [3])], 3)[:, :, :, :w_old + w_pad]
output = test(img_lq, model, args, window_size)
output = output[..., :h_old * args.scale, :w_old * args.scale]
# save image
output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
if output.ndim == 3:
output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0)) # CHW-RGB to HCW-BGR
output = (output * 255.0).round().astype(np.uint8) # float32 to uint8
cv2.imwrite(f'{save_dir}/{imgname}_SwinIR.png', output)
# evaluate psnr/ssim/psnr_b
if img_gt is not None:
img_gt = (img_gt * 255.0).round().astype(np.uint8) # float32 to uint8
img_gt = img_gt[:h_old * args.scale, :w_old * args.scale, ...] # crop gt
img_gt = np.squeeze(img_gt)
psnr = util.calculate_psnr(output, img_gt, crop_border=border)
ssim = util.calculate_ssim(output, img_gt, crop_border=border)
test_results['psnr'].append(psnr)
test_results['ssim'].append(ssim)
if img_gt.ndim == 3: # RGB image
psnr_y = util.calculate_psnr(output, img_gt, crop_border=border, test_y_channel=True)
ssim_y = util.calculate_ssim(output, img_gt, crop_border=border, test_y_channel=True)
test_results['psnr_y'].append(psnr_y)
test_results['ssim_y'].append(ssim_y)
if args.task in ['jpeg_car', 'color_jpeg_car']:
psnrb = util.calculate_psnrb(output, img_gt, crop_border=border, test_y_channel=False)
test_results['psnrb'].append(psnrb)
if args.task in ['color_jpeg_car']:
psnrb_y = util.calculate_psnrb(output, img_gt, crop_border=border, test_y_channel=True)
test_results['psnrb_y'].append(psnrb_y)
print('Testing {:d} {:20s} - PSNR: {:.2f} dB; SSIM: {:.4f}; PSNRB: {:.2f} dB;'
'PSNR_Y: {:.2f} dB; SSIM_Y: {:.4f}; PSNRB_Y: {:.2f} dB.'.
format(idx, imgname, psnr, ssim, psnrb, psnr_y, ssim_y, psnrb_y))
else:
print('Testing {:d} {:20s}'.format(idx, imgname))
# summarize psnr/ssim
if img_gt is not None:
ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
ave_ssim = sum(test_results['ssim']) / len(test_results['ssim'])
print('\n{} \n-- Average PSNR/SSIM(RGB): {:.2f} dB; {:.4f}'.format(save_dir, ave_psnr, ave_ssim))
if img_gt.ndim == 3:
ave_psnr_y = sum(test_results['psnr_y']) / len(test_results['psnr_y'])
ave_ssim_y = sum(test_results['ssim_y']) / len(test_results['ssim_y'])
print('-- Average PSNR_Y/SSIM_Y: {:.2f} dB; {:.4f}'.format(ave_psnr_y, ave_ssim_y))
if args.task in ['jpeg_car', 'color_jpeg_car']:
ave_psnrb = sum(test_results['psnrb']) / len(test_results['psnrb'])
print('-- Average PSNRB: {:.2f} dB'.format(ave_psnrb))
if args.task in ['color_jpeg_car']:
ave_psnrb_y = sum(test_results['psnrb_y']) / len(test_results['psnrb_y'])
print('-- Average PSNRB_Y: {:.2f} dB'.format(ave_psnrb_y))
def define_model(args):
# 001 classical image sr
if args.task == 'classical_sr':
model = net(upscale=args.scale, in_chans=3, img_size=args.training_patch_size, window_size=8,
img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='pixelshuffle', resi_connection='1conv')
param_key_g = 'params'
# 002 lightweight image sr
# use 'pixelshuffledirect' to save parameters
elif args.task == 'lightweight_sr':
model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
img_range=1., depths=[6, 6, 6, 6], embed_dim=60, num_heads=[6, 6, 6, 6],
mlp_ratio=2, upsampler='pixelshuffledirect', resi_connection='1conv')
param_key_g = 'params'
# 003 real-world image sr
elif args.task == 'real_sr':
if not args.large_model:
# use 'nearest+conv' to avoid block artifacts
model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='nearest+conv', resi_connection='1conv')
else:
# larger model size; use '3conv' to save parameters and memory; use ema for GAN training
model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
img_range=1., depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], embed_dim=240,
num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
mlp_ratio=2, upsampler='nearest+conv', resi_connection='3conv')
param_key_g = 'params_ema'
# 004 grayscale image denoising
elif args.task == 'gray_dn':
model = net(upscale=1, in_chans=1, img_size=128, window_size=8,
img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='', resi_connection='1conv')
param_key_g = 'params'
# 005 color image denoising
elif args.task == 'color_dn':
model = net(upscale=1, in_chans=3, img_size=128, window_size=8,
img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='', resi_connection='1conv')
param_key_g = 'params'
# 006 grayscale JPEG compression artifact reduction
# use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's sligtly better than 1
elif args.task == 'jpeg_car':
model = net(upscale=1, in_chans=1, img_size=126, window_size=7,
img_range=255., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='', resi_connection='1conv')
param_key_g = 'params'
# 006 color JPEG compression artifact reduction
# use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's sligtly better than 1
elif args.task == 'color_jpeg_car':
model = net(upscale=1, in_chans=3, img_size=126, window_size=7,
img_range=255., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
mlp_ratio=2, upsampler='', resi_connection='1conv')
param_key_g = 'params'
pretrained_model = torch.load(args.model_path)
model.load_state_dict(pretrained_model[param_key_g] if param_key_g in pretrained_model.keys() else pretrained_model, strict=True)
return model
def setup(args):
# 001 classical image sr/ 002 lightweight image sr
if args.task in ['classical_sr', 'lightweight_sr']:
save_dir = f'results/swinir_{args.task}_x{args.scale}'
folder = args.folder_gt
border = args.scale
window_size = 8
# 003 real-world image sr
elif args.task in ['real_sr']:
save_dir = f'results/swinir_{args.task}_x{args.scale}'
if args.large_model:
save_dir += '_large'
folder = args.folder_lq
border = 0
window_size = 8
# 004 grayscale image denoising/ 005 color image denoising
elif args.task in ['gray_dn', 'color_dn']:
save_dir = f'results/swinir_{args.task}_noise{args.noise}'
folder = args.folder_gt
border = 0
window_size = 8
# 006 JPEG compression artifact reduction
elif args.task in ['jpeg_car', 'color_jpeg_car']:
save_dir = f'results/swinir_{args.task}_jpeg{args.jpeg}'
folder = args.folder_gt
border = 0
window_size = 7
return folder, save_dir, border, window_size
def get_image_pair(args, path):
(imgname, imgext) = os.path.splitext(os.path.basename(path))
# 001 classical image sr/ 002 lightweight image sr (load lq-gt image pairs)
if args.task in ['classical_sr', 'lightweight_sr']:
img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
img_lq = cv2.imread(f'{args.folder_lq}/{imgname}x{args.scale}{imgext}', cv2.IMREAD_COLOR).astype(
np.float32) / 255.
# 003 real-world image sr (load lq image only)
elif args.task in ['real_sr']:
img_gt = None
img_lq = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
# 004 grayscale image denoising (load gt image and generate lq image on-the-fly)
elif args.task in ['gray_dn']:
img_gt = cv2.imread(path, cv2.IMREAD_GRAYSCALE).astype(np.float32) / 255.
np.random.seed(seed=0)
img_lq = img_gt + np.random.normal(0, args.noise / 255., img_gt.shape)
img_gt = np.expand_dims(img_gt, axis=2)
img_lq = np.expand_dims(img_lq, axis=2)
# 005 color image denoising (load gt image and generate lq image on-the-fly)
elif args.task in ['color_dn']:
img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
np.random.seed(seed=0)
img_lq = img_gt + np.random.normal(0, args.noise / 255., img_gt.shape)
# 006 grayscale JPEG compression artifact reduction (load gt image and generate lq image on-the-fly)
elif args.task in ['jpeg_car']:
img_gt = cv2.imread(path, cv2.IMREAD_UNCHANGED)
if img_gt.ndim != 2:
img_gt = util.bgr2ycbcr(img_gt, y_only=True)
result, encimg = cv2.imencode('.jpg', img_gt, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg])
img_lq = cv2.imdecode(encimg, 0)
img_gt = np.expand_dims(img_gt, axis=2).astype(np.float32) / 255.
img_lq = np.expand_dims(img_lq, axis=2).astype(np.float32) / 255.
# 006 JPEG compression artifact reduction (load gt image and generate lq image on-the-fly)
elif args.task in ['color_jpeg_car']:
img_gt = cv2.imread(path)
result, encimg = cv2.imencode('.jpg', img_gt, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg])
img_lq = cv2.imdecode(encimg, 1)
img_gt = img_gt.astype(np.float32)/ 255.
img_lq = img_lq.astype(np.float32)/ 255.
return imgname, img_lq, img_gt
def test(img_lq, model, args, window_size):
if args.tile is None:
# test the image as a whole
output = model(img_lq)
else:
# test the image tile by tile
b, c, h, w = img_lq.size()
tile = min(args.tile, h, w)
assert tile % window_size == 0, "tile size should be a multiple of window_size"
tile_overlap = args.tile_overlap
sf = args.scale
stride = tile - tile_overlap
h_idx_list = list(range(0, h-tile, stride)) + [h-tile]
w_idx_list = list(range(0, w-tile, stride)) + [w-tile]
E = torch.zeros(b, c, h*sf, w*sf).type_as(img_lq)
W = torch.zeros_like(E)
for h_idx in h_idx_list:
for w_idx in w_idx_list:
in_patch = img_lq[..., h_idx:h_idx+tile, w_idx:w_idx+tile]
out_patch = model(in_patch)
out_patch_mask = torch.ones_like(out_patch)
E[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch)
W[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch_mask)
output = E.div_(W)
return output
if __name__ == '__main__':
main()效果展示：
低质量图：
高质量图：
效果还是可以的，适用于超分辨率重建。
以上为全部内容！
550
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
