Upload denoise_util.py

denoise_util.py

@@ -0,0 +1,410 @@
+import torch
+import torch.nn as nn
+
+# SOURCE: https://github.com/Ascend-Research/CascadedGaze
+# ------------------------------------------------------------------------
+# Modified from NAFNet (https://github.com/megvii-research/NAFNet)
+# ------------------------------------------------------------------------
+import torch.nn.functional as F
+
+class LayerNormFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, weight, bias, eps):
+        ctx.eps = eps
+        N, C, H, W = x.size()
+        mu = x.mean(1, keepdim=True)
+        var = (x - mu).pow(2).mean(1, keepdim=True)
+        y = (x - mu) / (var + eps).sqrt()
+        ctx.save_for_backward(y, var, weight)
+        y = weight.view(1, C, 1, 1) * y + bias.view(1, C, 1, 1)
+        return y
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        eps = ctx.eps
+
+        N, C, H, W = grad_output.size()
+        y, var, weight = ctx.saved_variables
+        g = grad_output * weight.view(1, C, 1, 1)
+        mean_g = g.mean(dim=1, keepdim=True)
+
+        mean_gy = (g * y).mean(dim=1, keepdim=True)
+        gx = 1. / torch.sqrt(var + eps) * (g - y * mean_gy - mean_g)
+        return gx, (grad_output * y).sum(dim=3).sum(dim=2).sum(dim=0), grad_output.sum(dim=3).sum(dim=2).sum(
+            dim=0), None
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, channels, eps=1e-6):
+        super(LayerNorm2d, self).__init__()
+        self.register_parameter('weight', nn.Parameter(torch.ones(channels)))
+        self.register_parameter('bias', nn.Parameter(torch.zeros(channels)))
+        self.eps = eps
+
+    def forward(self, x):
+        return LayerNormFunction.apply(x, self.weight, self.bias, self.eps)
+
+class AvgPool2d(nn.Module):
+    def __init__(self, kernel_size=None, base_size=None, auto_pad=True, fast_imp=False, train_size=None):
+        super().__init__()
+        self.kernel_size = kernel_size
+        self.base_size = base_size
+        self.auto_pad = auto_pad
+
+        # only used for fast implementation
+        self.fast_imp = fast_imp
+        self.rs = [5, 4, 3, 2, 1]
+        self.max_r1 = self.rs[0]
+        self.max_r2 = self.rs[0]
+        self.train_size = train_size
+
+    def extra_repr(self) -> str:
+        return 'kernel_size={}, base_size={}, stride={}, fast_imp={}'.format(
+            self.kernel_size, self.base_size, self.kernel_size, self.fast_imp
+        )
+
+    def forward(self, x):
+        if self.kernel_size is None and self.base_size:
+            train_size = self.train_size
+            if isinstance(self.base_size, int):
+                self.base_size = (self.base_size, self.base_size)
+            self.kernel_size = list(self.base_size)
+            self.kernel_size[0] = x.shape[2] * self.base_size[0] // train_size[-2]
+            self.kernel_size[1] = x.shape[3] * self.base_size[1] // train_size[-1]
+
+            # only used for fast implementation
+            self.max_r1 = max(1, self.rs[0] * x.shape[2] // train_size[-2])
+            self.max_r2 = max(1, self.rs[0] * x.shape[3] // train_size[-1])
+
+        if self.kernel_size[0] >= x.size(-2) and self.kernel_size[1] >= x.size(-1):
+            return F.adaptive_avg_pool2d(x, 1)
+
+        if self.fast_imp:  # Non-equivalent implementation but faster
+            h, w = x.shape[2:]
+            if self.kernel_size[0] >= h and self.kernel_size[1] >= w:
+                out = F.adaptive_avg_pool2d(x, 1)
+            else:
+                r1 = [r for r in self.rs if h % r == 0][0]
+                r2 = [r for r in self.rs if w % r == 0][0]
+                # reduction_constraint
+                r1 = min(self.max_r1, r1)
+                r2 = min(self.max_r2, r2)
+                s = x[:, :, ::r1, ::r2].cumsum(dim=-1).cumsum(dim=-2)
+                n, c, h, w = s.shape
+                k1, k2 = min(h - 1, self.kernel_size[0] // r1), min(w - 1, self.kernel_size[1] // r2)
+                out = (s[:, :, :-k1, :-k2] - s[:, :, :-k1, k2:] - s[:, :, k1:, :-k2] + s[:, :, k1:, k2:]) / (k1 * k2)
+                out = torch.nn.functional.interpolate(out, scale_factor=(r1, r2))
+        else:
+            n, c, h, w = x.shape
+            s = x.cumsum(dim=-1).cumsum_(dim=-2)
+            s = torch.nn.functional.pad(s, (1, 0, 1, 0))  # pad 0 for convenience
+            k1, k2 = min(h, self.kernel_size[0]), min(w, self.kernel_size[1])
+            s1, s2, s3, s4 = s[:, :, :-k1, :-k2], s[:, :, :-k1, k2:], s[:, :, k1:, :-k2], s[:, :, k1:, k2:]
+            out = s4 + s1 - s2 - s3
+            out = out / (k1 * k2)
+
+        if self.auto_pad:
+            n, c, h, w = x.shape
+            _h, _w = out.shape[2:]
+            # print(x.shape, self.kernel_size)
+            pad2d = ((w - _w) // 2, (w - _w + 1) // 2, (h - _h) // 2, (h - _h + 1) // 2)
+            out = torch.nn.functional.pad(out, pad2d, mode='replicate')
+
+        return out
+
+def replace_layers(model, base_size, train_size, fast_imp, **kwargs):
+    for n, m in model.named_children():
+        if len(list(m.children())) > 0:
+            ## compound module, go inside it
+            replace_layers(m, base_size, train_size, fast_imp, **kwargs)
+
+        if isinstance(m, nn.AdaptiveAvgPool2d):
+            # print(base_size)
+            pool = AvgPool2d(base_size=base_size, fast_imp=fast_imp, train_size=train_size)
+            assert m.output_size == 1
+            setattr(model, n, pool)
+
+'''
+ref. 
+@article{chu2021tlsc,
+  title={Revisiting Global Statistics Aggregation for Improving Image Restoration},
+  author={Chu, Xiaojie and Chen, Liangyu and and Chen, Chengpeng and Lu, Xin},
+  journal={arXiv preprint arXiv:2112.04491},
+  year={2021}
+}
+'''
+class Local_Base():
+    def convert(self, *args, train_size, **kwargs):
+        replace_layers(self, *args, train_size=train_size, **kwargs)
+        imgs = torch.rand(train_size)
+        with torch.no_grad():
+            self.forward(imgs)
+
+class SimpleGate(nn.Module):
+    def forward(self, x):
+        x1, x2 = x.chunk(2, dim=1)
+        return x1 * x2
+    
+class depthwise_separable_conv(nn.Module):
+    def __init__(self, nin, nout, kernel_size = 3, padding = 0, stide = 1, bias=False):
+        super(depthwise_separable_conv, self).__init__()
+        self.pointwise = nn.Conv2d(nin, nout, kernel_size=1, bias=bias)
+        self.depthwise = nn.Conv2d(nin, nin, kernel_size=kernel_size, stride=stide, padding=padding, groups=nin, bias=bias)
+
+    def forward(self, x):
+        x = self.depthwise(x)
+        x = self.pointwise(x)
+        return x
+    
+
+class UpsampleWithFlops(nn.Upsample):
+    def __init__(self, size=None, scale_factor=None, mode='nearest', align_corners=None):
+        super(UpsampleWithFlops, self).__init__(size, scale_factor, mode, align_corners)
+        self.__flops__ = 0
+
+    def forward(self, input):
+        self.__flops__ += input.numel()
+        return super(UpsampleWithFlops, self).forward(input)
+
+
+class GlobalContextExtractor(nn.Module):
+    def __init__(self, c, kernel_sizes=[3, 3, 5], strides=[3, 3, 5], padding=0, bias=False):
+        super(GlobalContextExtractor, self).__init__()
+
+        self.depthwise_separable_convs = nn.ModuleList([
+            depthwise_separable_conv(c, c, kernel_size, padding, stride, bias)
+            for kernel_size, stride in zip(kernel_sizes, strides)
+        ])
+
+    def forward(self, x):
+        outputs = []
+        for conv in self.depthwise_separable_convs:
+            x = F.gelu(conv(x))
+            outputs.append(x)
+        return outputs
+
+
+class CascadedGazeBlock(nn.Module):
+    def __init__(self, c, GCE_Conv =2, DW_Expand=2, FFN_Expand=2, drop_out_rate=0):
+        super().__init__()
+        self.dw_channel = c * DW_Expand
+        self.GCE_Conv = GCE_Conv
+        self.conv1 = nn.Conv2d(in_channels=c, out_channels=self.dw_channel, kernel_size=1,
+                                padding=0, stride=1, groups=1, bias=True)
+        self.conv2 = nn.Conv2d(in_channels=self.dw_channel, out_channels=self.dw_channel,
+                                kernel_size=3, padding=1, stride=1, groups=self.dw_channel,
+                               bias=True)
+
+        
+        if self.GCE_Conv == 3:
+            self.GCE = GlobalContextExtractor(c=c, kernel_sizes=[3, 3, 5], strides=[2, 3, 4])
+
+            self.project_out = nn.Conv2d(int(self.dw_channel*2.5), c, kernel_size=1)
+
+            self.sca = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(in_channels=int(self.dw_channel*2.5), out_channels=int(self.dw_channel*2.5), kernel_size=1, padding=0, stride=1,
+                        groups=1, bias=True))
+        else:
+            self.GCE = GlobalContextExtractor(c=c, kernel_sizes=[3, 3], strides=[2, 3])
+
+            self.project_out = nn.Conv2d(self.dw_channel*2, c, kernel_size=1)
+
+            self.sca = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(in_channels=self.dw_channel*2, out_channels=self.dw_channel*2, kernel_size=1, padding=0, stride=1,
+                        groups=1, bias=True))
+
+
+        # SimpleGate
+        self.sg = SimpleGate()
+
+        ffn_channel = FFN_Expand * c
+        self.conv4 = nn.Conv2d(in_channels=c, out_channels=ffn_channel, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+        self.conv5 = nn.Conv2d(in_channels=ffn_channel // 2, out_channels=c, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+
+        self.norm1 = LayerNorm2d(c)
+        self.norm2 = LayerNorm2d(c)
+
+        self.dropout1 = nn.Dropout(drop_out_rate) if drop_out_rate > 0. else nn.Identity()
+        self.dropout2 = nn.Dropout(drop_out_rate) if drop_out_rate > 0. else nn.Identity()
+
+        self.beta = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)
+        self.gamma = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)
+
+    def forward(self, inp):
+        x = inp
+        b,c,h,w = x.shape
+        # # Nearest neighbor upsampling as part of the range fusion process
+        self.upsample = UpsampleWithFlops(size=(h,w), mode='nearest')
+
+
+        x = self.norm1(x)
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = F.gelu(x)
+
+
+        # Global Context Extractor + Range fusion
+        x_1 , x_2 = x.chunk(2, dim=1)
+        if self.GCE_Conv == 3:
+            x1, x2, x3 = self.GCE(x_1 + x_2)
+            x = torch.cat([x, self.upsample(x1), self.upsample(x2), self.upsample(x3)], dim = 1)
+        else:
+            x1, x2 = self.GCE(x_1 + x_2)
+            x = torch.cat([x, self.upsample(x1), self.upsample(x2)], dim = 1)
+        x = self.sca(x) * x
+        x = self.project_out(x)
+
+
+        x = self.dropout1(x)
+        #channel-mixing
+        y = inp + x * self.beta
+        x = self.conv4(self.norm2(y))
+        x = self.sg(x)
+        x = self.conv5(x)
+        x = self.dropout2(x)
+
+        return y + x * self.gamma
+
+class NAFBlock0(nn.Module):
+    def __init__(self, c, DW_Expand=2, FFN_Expand=2, drop_out_rate=0.0):
+        super().__init__()
+        dw_channel = c * DW_Expand
+        self.conv1 = nn.Conv2d(in_channels=c, out_channels=dw_channel, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+        self.conv2 = nn.Conv2d(in_channels=dw_channel, out_channels=dw_channel, kernel_size=3, padding=1, stride=1, groups=dw_channel,
+                               bias=True)
+        self.conv3 = nn.Conv2d(in_channels=dw_channel // 2, out_channels=c, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+        
+        # Simplified Channel Attention
+        self.sca = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Conv2d(in_channels=dw_channel // 2, out_channels=dw_channel // 2, kernel_size=1, padding=0, stride=1,
+                      groups=1, bias=True),
+        )
+
+        # SimpleGate
+        self.sg = SimpleGate()
+
+        ffn_channel = FFN_Expand * c
+        self.conv4 = nn.Conv2d(in_channels=c, out_channels=ffn_channel, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+        self.conv5 = nn.Conv2d(in_channels=ffn_channel // 2, out_channels=c, kernel_size=1, padding=0, stride=1, groups=1, bias=True)
+
+        self.norm1 = LayerNorm2d(c)
+        self.norm2 = LayerNorm2d(c)
+
+        self.dropout1 = nn.Dropout(drop_out_rate) if drop_out_rate > 0. else nn.Identity()
+        self.dropout2 = nn.Dropout(drop_out_rate) if drop_out_rate > 0. else nn.Identity()
+
+        self.beta = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)
+        self.gamma = nn.Parameter(torch.zeros((1, c, 1, 1)), requires_grad=True)
+
+    def forward(self, inp):
+        x = inp
+
+        x = self.norm1(x)
+
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.sg(x)
+        x = x * self.sca(x)
+        x = self.conv3(x)
+
+        x = self.dropout1(x)
+
+        y = inp + x * self.beta
+
+        #Channel Mixing
+        x = self.conv4(self.norm2(y))
+        x = self.sg(x)
+        x = self.conv5(x)
+
+        x = self.dropout2(x)
+
+        return y + x * self.gamma
+    
+
+class CascadedGaze(nn.Module):
+
+    def __init__(self, img_channel=3, width=16, middle_blk_num=1, enc_blk_nums=[], dec_blk_nums=[], GCE_CONVS_nums=[]):
+        super().__init__()
+
+        self.intro = nn.Conv2d(in_channels=img_channel, out_channels=width, kernel_size=3, padding=1, stride=1, groups=1,
+                              bias=True)
+        self.ending = nn.Conv2d(in_channels=width, out_channels=img_channel, kernel_size=3, padding=1, stride=1, groups=1,
+                              bias=True)
+
+        self.encoders = nn.ModuleList()
+        self.decoders = nn.ModuleList()
+        self.middle_blks = nn.ModuleList()
+        self.ups = nn.ModuleList()
+        self.downs = nn.ModuleList()
+
+        chan = width
+        # for num in enc_blk_nums:
+        for i in range(len(enc_blk_nums)):
+            num = enc_blk_nums[i]
+            GCE_Convs = GCE_CONVS_nums[i]
+            self.encoders.append(
+                nn.Sequential(
+                    *[CascadedGazeBlock(chan, GCE_Conv=GCE_Convs) for _ in range(num)]
+                )
+            )
+            self.downs.append(
+                nn.Conv2d(chan, 2*chan, 2, 2)
+            )
+            chan = chan * 2
+
+        self.middle_blks = \
+            nn.Sequential(
+                *[NAFBlock0(chan) for _ in range(middle_blk_num)]
+            )
+
+        for i in range(len(dec_blk_nums)):
+            num = dec_blk_nums[i]
+            self.ups.append(
+                nn.Sequential(
+                    nn.Conv2d(chan, chan * 2, 1, bias=False),
+                    nn.PixelShuffle(2)
+                )
+            )
+            chan = chan // 2
+            self.decoders.append(
+                nn.Sequential(
+                    *[NAFBlock0(chan) for _ in range(num)]
+                )
+            )
+
+        self.padder_size = 2 ** len(self.encoders)
+
+    def forward(self, inp):
+        B, C, H, W = inp.shape
+        inp = self.check_image_size(inp)
+
+        x = self.intro(inp)
+
+        encs = []
+
+        for encoder, down in zip(self.encoders, self.downs):
+            x = encoder(x)
+            encs.append(x)
+            x = down(x)
+
+        x = self.middle_blks(x)
+
+        for decoder, up, enc_skip in zip(self.decoders, self.ups, encs[::-1]):
+            x = up(x)
+            x = x + enc_skip
+            x = decoder(x)
+
+        x = self.ending(x)
+        x = x + inp
+
+        return x[:, :, :H, :W]
+
+    def check_image_size(self, x):
+        _, _, h, w = x.size()
+        mod_pad_h = (self.padder_size - h % self.padder_size) % self.padder_size
+        mod_pad_w = (self.padder_size - w % self.padder_size) % self.padder_size
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h))
+        return x
+