Updated RIFE-CUDA to RIFE 2.0

Code/IO/IOUtils.cs

@@ -624,6 +624,7 @@ namespace Flowframes.IO
 				Directory.CreateDirectory(path);
 				return false;
 			}
+
 			return true;
         }
 

Pkgs/rife-cuda/model/IFNet_HD.py

@@ -91,9 +91,12 @@ class IFNet(nn.Module):
         self.block2 = IFBlock(8, scale=2, c=96)
         self.block3 = IFBlock(8, scale=1, c=48)
 
-    def forward(self, x):
-        x = F.interpolate(x, scale_factor=0.5, mode="bilinear",
-                          align_corners=False)
+    def forward(self, x, UHD=False):
+        if UHD:
+            x = F.interpolate(x, scale_factor=0.25, mode="bilinear", align_corners=False)
+        else:
+            x = F.interpolate(x, scale_factor=0.5, mode="bilinear",
+                              align_corners=False)
         flow0 = self.block0(x)
         F1 = flow0
         warped_img0 = warp(x[:, :3], F1)

Pkgs/rife-cuda/model/RIFE.py

@@ -204,7 +204,7 @@ class Model:
     def inference(self, img0, img1):
         imgs = torch.cat((img0, img1), 1)
         flow, _ = self.flownet(imgs)
-        return self.predict(imgs, flow, training=False).detach()
+        return self.predict(imgs, flow, training=False)
 
     def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
         for param_group in self.optimG.param_groups:

Pkgs/rife-cuda/model/RIFE2F.py

@@ -204,7 +204,7 @@ class Model:
         with torch.no_grad():
             imgs = torch.cat((img0, img1), 1)
             flow, _ = self.flownet(imgs)
-        return self.predict(imgs, flow, training=False).detach()
+        return self.predict(imgs, flow, training=False)
 
     def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
         for param_group in self.optimG.param_groups:

Pkgs/rife-cuda/model/RIFE_HD.py

@@ -209,10 +209,10 @@ class Model:
         else:
             return pred
 
-    def inference(self, img0, img1, K=False):
+    def inference(self, img0, img1, UHD=False):
         imgs = torch.cat((img0, img1), 1)
-        flow, _ = self.flownet(imgs)
-        return self.predict(imgs, flow, training=False).detach()
+        flow, _ = self.flownet(imgs, UHD)
+        return self.predict(imgs, flow, training=False, UHD=UHD)
 
     def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
         for param_group in self.optimG.param_groups:

pkgs/rife-cuda/model/IFNet2F15C.py

@@ -0,0 +1,115 @@
+import torch
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+from model.warplayer import warp
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def conv_wo_act(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=False),
+        nn.BatchNorm2d(out_planes),
+    )
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=False),
+        nn.BatchNorm2d(out_planes),
+        nn.PReLU(out_planes)
+    )
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=1):
+        super(ResBlock, self).__init__()
+        if in_planes == out_planes and stride == 1:
+            self.conv0 = nn.Identity()
+        else:
+            self.conv0 = nn.Conv2d(in_planes, out_planes,
+                                   3, stride, 1, bias=False)
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv_wo_act(out_planes, out_planes, 3, 1, 1)
+        self.relu1 = nn.PReLU(1)
+        self.relu2 = nn.PReLU(out_planes)
+        self.fc1 = nn.Conv2d(out_planes, 16, kernel_size=1, bias=False)
+        self.fc2 = nn.Conv2d(16, out_planes, kernel_size=1, bias=False)
+
+    def forward(self, x):
+        y = self.conv0(x)
+        x = self.conv1(x)
+        x = self.conv2(x)
+        w = x.mean(3, True).mean(2, True)
+        w = self.relu1(self.fc1(w))
+        w = torch.sigmoid(self.fc2(w))
+        x = self.relu2(x * w + y)
+        return x
+
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, scale=1, c=64):
+        super(IFBlock, self).__init__()
+        self.scale = scale
+        self.conv0 = conv(in_planes, c, 3, 1, 1)
+        self.res0 = ResBlock(c, c)
+        self.res1 = ResBlock(c, c)
+        self.res2 = ResBlock(c, c)
+        self.res3 = ResBlock(c, c)
+        self.res4 = ResBlock(c, c)
+        self.res5 = ResBlock(c, c)
+        self.conv1 = nn.Conv2d(c, 2, 3, 1, 1)
+        self.up = nn.PixelShuffle(2)
+
+    def forward(self, x):
+        if self.scale != 1:
+            x = F.interpolate(x, scale_factor=1. / self.scale, mode="bilinear",
+                              align_corners=False)
+        x = self.conv0(x)
+        x = self.res0(x)
+        x = self.res1(x)
+        x = self.res2(x)
+        x = self.res3(x)
+        x = self.res4(x)
+        x = self.res5(x)
+        x = self.conv1(x)
+        flow = x # self.up(x)
+        if self.scale != 1:
+            flow = F.interpolate(flow, scale_factor=self.scale, mode="bilinear",
+                                 align_corners=False)
+        return flow
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(6, scale=4, c=288)
+        self.block1 = IFBlock(8, scale=2, c=192)
+        self.block2 = IFBlock(8, scale=1, c=96)
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=0.5, mode="bilinear",
+                          align_corners=False)
+        flow0 = self.block0(x)
+        F1 = flow0
+        warped_img0 = warp(x[:, :3], F1)
+        warped_img1 = warp(x[:, 3:], -F1)
+        flow1 = self.block1(torch.cat((warped_img0, warped_img1, F1), 1))
+        F2 = (flow0 + flow1)
+        warped_img0 = warp(x[:, :3], F2)
+        warped_img1 = warp(x[:, 3:], -F2)
+        flow2 = self.block2(torch.cat((warped_img0, warped_img1, F2), 1))
+        F3 = (flow0 + flow1 + flow2)
+        return F3, [F1, F2, F3]
+
+if __name__ == '__main__':
+    img0 = torch.zeros(3, 3, 256, 256).float().to(device)
+    img1 = torch.tensor(np.random.normal(
+        0, 1, (3, 3, 256, 256))).float().to(device)
+    imgs = torch.cat((img0, img1), 1)
+    flownet = IFNet()
+    flow, _ = flownet(imgs)
+    print(flow.shape)

pkgs/rife-cuda/model/IFNet_HDv2.py

@@ -0,0 +1,93 @@
+import torch
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+from model.warplayer import warp
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def conv_wo_act(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+    )
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+        nn.PReLU(out_planes)
+    )
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, scale=1, c=64):
+        super(IFBlock, self).__init__()
+        self.scale = scale
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c, 3, 2, 1),
+            conv(c, 2*c, 3, 2, 1),
+            )
+        self.convblock = nn.Sequential(
+            conv(2*c, 2*c),
+            conv(2*c, 2*c),
+            conv(2*c, 2*c),
+            conv(2*c, 2*c),
+            conv(2*c, 2*c),
+            conv(2*c, 2*c),
+        )        
+        self.conv1 = nn.ConvTranspose2d(2*c, 4, 4, 2, 1)
+                    
+    def forward(self, x):
+        if self.scale != 1:
+            x = F.interpolate(x, scale_factor=1. / self.scale, mode="bilinear",
+                              align_corners=False)
+        x = self.conv0(x)
+        x = self.convblock(x)
+        x = self.conv1(x)
+        flow = x
+        if self.scale != 1:
+            flow = F.interpolate(flow, scale_factor=self.scale, mode="bilinear",
+                                 align_corners=False)
+        return flow
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(6, scale=8, c=192)
+        self.block1 = IFBlock(10, scale=4, c=128)
+        self.block2 = IFBlock(10, scale=2, c=96)
+        self.block3 = IFBlock(10, scale=1, c=48)
+
+    def forward(self, x, UHD=False):
+        if UHD:
+            x = F.interpolate(x, scale_factor=0.5, mode="bilinear", align_corners=False)
+        flow0 = self.block0(x)
+        F1 = flow0
+        F1_large = F.interpolate(F1, scale_factor=2.0, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 2.0
+        warped_img0 = warp(x[:, :3], F1_large[:, :2])
+        warped_img1 = warp(x[:, 3:], F1_large[:, 2:4])
+        flow1 = self.block1(torch.cat((warped_img0, warped_img1, F1_large), 1))
+        F2 = (flow0 + flow1)
+        F2_large = F.interpolate(F2, scale_factor=2.0, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 2.0
+        warped_img0 = warp(x[:, :3], F2_large[:, :2])
+        warped_img1 = warp(x[:, 3:], F2_large[:, 2:4])
+        flow2 = self.block2(torch.cat((warped_img0, warped_img1, F2_large), 1))
+        F3 = (flow0 + flow1 + flow2)
+        F3_large = F.interpolate(F3, scale_factor=2.0, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 2.0
+        warped_img0 = warp(x[:, :3], F3_large[:, :2])
+        warped_img1 = warp(x[:, 3:], F3_large[:, 2:4])
+        flow3 = self.block3(torch.cat((warped_img0, warped_img1, F3_large), 1))
+        F4 = (flow0 + flow1 + flow2 + flow3)
+        return F4, [F1, F2, F3, F4]
+
+if __name__ == '__main__':
+    img0 = torch.zeros(3, 3, 256, 256).float().to(device)
+    img1 = torch.tensor(np.random.normal(
+        0, 1, (3, 3, 256, 256))).float().to(device)
+    imgs = torch.cat((img0, img1), 1)
+    flownet = IFNet()
+    flow, _ = flownet(imgs)
+    print(flow.shape)

pkgs/rife-cuda/model/RIFE2F15C.py

@@ -0,0 +1,250 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from torch.optim import AdamW
+import torch.optim as optim
+import itertools
+from model.warplayer import warp
+from torch.nn.parallel import DistributedDataParallel as DDP
+from model.IFNet2F15C import *
+import torch.nn.functional as F
+from model.loss import *
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+        nn.PReLU(out_planes)
+    )
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes,
+                                 kernel_size=4, stride=2, padding=1, bias=True),
+        nn.PReLU(out_planes)
+    )
+
+def conv_woact(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+    )
+
+class ResBlock(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=2):
+        super(ResBlock, self).__init__()
+        if in_planes == out_planes and stride == 1:
+            self.conv0 = nn.Identity()
+        else:
+            self.conv0 = nn.Conv2d(in_planes, out_planes,
+                                   3, stride, 1, bias=False)
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv_woact(out_planes, out_planes, 3, 1, 1)
+        self.relu1 = nn.PReLU(1)
+        self.relu2 = nn.PReLU(out_planes)
+        self.fc1 = nn.Conv2d(out_planes, 16, kernel_size=1, bias=False)
+        self.fc2 = nn.Conv2d(16, out_planes, kernel_size=1, bias=False)
+
+    def forward(self, x):
+        y = self.conv0(x)
+        x = self.conv1(x)
+        x = self.conv2(x)
+        w = x.mean(3, True).mean(2, True)
+        w = self.relu1(self.fc1(w))
+        w = torch.sigmoid(self.fc2(w))
+        x = self.relu2(x * w + y)
+        return x
+
+c = 24
+
+class ContextNet(nn.Module):
+    def __init__(self):
+        super(ContextNet, self).__init__()
+        self.conv1 = ResBlock(3, c, 1)
+        self.conv2 = ResBlock(c, 2*c)
+        self.conv3 = ResBlock(2*c, 4*c)
+        self.conv4 = ResBlock(4*c, 8*c)
+
+    def forward(self, x, flow):
+        x = self.conv1(x)
+        f1 = warp(x, flow)
+        x = self.conv2(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f2 = warp(x, flow)
+        x = self.conv3(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f3 = warp(x, flow)
+        x = self.conv4(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f4 = warp(x, flow)
+        return [f1, f2, f3, f4]
+
+
+class FusionNet(nn.Module):
+    def __init__(self):
+        super(FusionNet, self).__init__()
+        self.down0 = ResBlock(8, 2*c, 1)
+        self.down1 = ResBlock(4*c, 4*c)
+        self.down2 = ResBlock(8*c, 8*c)
+        self.down3 = ResBlock(16*c, 16*c)
+        self.up0 = deconv(32*c, 8*c)
+        self.up1 = deconv(16*c, 4*c)
+        self.up2 = deconv(8*c, 2*c)
+        self.up3 = deconv(4*c, c)
+        self.conv = nn.Conv2d(c, 4, 3, 2, 1)
+
+    def forward(self, img0, img1, flow, c0, c1, flow_gt):
+        warped_img0 = warp(img0, flow)
+        warped_img1 = warp(img1, -flow)
+        if flow_gt == None:
+            warped_img0_gt, warped_img1_gt = None, None
+        else:
+            warped_img0_gt = warp(img0, flow_gt[:, :2])
+            warped_img1_gt = warp(img1, flow_gt[:, 2:4])
+        s0 = self.down0(torch.cat((warped_img0, warped_img1, flow), 1))
+        s1 = self.down1(torch.cat((s0, c0[0], c1[0]), 1))
+        s2 = self.down2(torch.cat((s1, c0[1], c1[1]), 1))
+        s3 = self.down3(torch.cat((s2, c0[2], c1[2]), 1))
+        x = self.up0(torch.cat((s3, c0[3], c1[3]), 1))
+        x = self.up1(torch.cat((x, s2), 1))
+        x = self.up2(torch.cat((x, s1), 1))
+        x = self.up3(torch.cat((x, s0), 1))
+        x = self.conv(x)
+        return x, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt
+
+
+class Model:
+    def __init__(self, local_rank=-1):
+        self.flownet = IFNet()
+        self.contextnet = ContextNet()
+        self.fusionnet = FusionNet()
+        self.device()
+        self.optimG = AdamW(itertools.chain(
+            self.flownet.parameters(),
+            self.contextnet.parameters(),
+            self.fusionnet.parameters()), lr=1e-6, weight_decay=1e-5)
+        self.schedulerG = optim.lr_scheduler.CyclicLR(
+            self.optimG, base_lr=1e-6, max_lr=1e-3, step_size_up=8000, cycle_momentum=False)
+        self.epe = EPE()
+        self.ter = Ternary()
+        self.sobel = SOBEL()
+        if local_rank != -1:
+            self.flownet = DDP(self.flownet, device_ids=[
+                               local_rank], output_device=local_rank)
+            self.contextnet = DDP(self.contextnet, device_ids=[
+                                  local_rank], output_device=local_rank)
+            self.fusionnet = DDP(self.fusionnet, device_ids=[
+                                 local_rank], output_device=local_rank)
+
+    def train(self):
+        self.flownet.train()
+        self.contextnet.train()
+        self.fusionnet.train()
+
+    def eval(self):
+        self.flownet.eval()
+        self.contextnet.eval()
+        self.fusionnet.eval()
+
+    def device(self):
+        self.flownet.to(device)
+        self.contextnet.to(device)
+        self.fusionnet.to(device)
+
+    def load_model(self, path, rank=0):
+        def convert(param):
+            return {
+                k.replace("module.", ""): v
+                for k, v in param.items()
+                if "module." in k
+            }
+        if rank == 0:
+            self.flownet.load_state_dict(
+                convert(torch.load('{}/flownet.pkl'.format(path), map_location=device)))
+            self.contextnet.load_state_dict(
+                convert(torch.load('{}/contextnet.pkl'.format(path), map_location=device)))
+            self.fusionnet.load_state_dict(
+                convert(torch.load('{}/unet.pkl'.format(path), map_location=device)))
+
+    def save_model(self, path, rank=0):
+        if rank == 0:
+            torch.save(self.flownet.state_dict(),
+                       '{}/flownet.pkl'.format(path))
+            torch.save(self.contextnet.state_dict(),
+                       '{}/contextnet.pkl'.format(path))
+            torch.save(self.fusionnet.state_dict(), '{}/unet.pkl'.format(path))
+
+    def predict(self, imgs, flow, training=True, flow_gt=None):
+        img0 = imgs[:, :3]
+        img1 = imgs[:, 3:]
+        flow = F.interpolate(flow, scale_factor=2.0, mode="bilinear",
+	                     align_corners=False) * 2.0
+        c0 = self.contextnet(img0, flow)
+        c1 = self.contextnet(img1, -flow)
+        refine_output, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt = self.fusionnet(
+            img0, img1, flow, c0, c1, flow_gt)
+        res = torch.sigmoid(refine_output[:, :3]) * 2 - 1
+        mask = torch.sigmoid(refine_output[:, 3:4])
+        merged_img = warped_img0 * mask + warped_img1 * (1 - mask)
+        pred = merged_img + res
+        pred = torch.clamp(pred, 0, 1)
+        if training:
+            return pred, mask, merged_img, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt
+        else:
+            return pred
+
+    def inference(self, img0, img1):
+        with torch.no_grad():
+            imgs = torch.cat((img0, img1), 1)
+            flow, _ = self.flownet(imgs)
+        return self.predict(imgs, flow, training=False)
+
+    def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
+        for param_group in self.optimG.param_groups:
+            param_group['lr'] = learning_rate
+        if training:
+            self.train()
+        else:
+            self.eval()
+        flow, flow_list = self.flownet(imgs)
+        pred, mask, merged_img, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt = self.predict(
+            imgs, flow, flow_gt=flow_gt)
+        loss_ter = self.ter(pred, gt).mean()
+        if training:
+            with torch.no_grad():
+                loss_flow = torch.abs(warped_img0_gt - gt).mean()
+                loss_mask = torch.abs(
+                    merged_img - gt).sum(1, True).float().detach()
+            loss_cons = 0
+            for i in range(3):
+                loss_cons += self.epe(flow_list[i], flow_gt[:, :2], 1)
+                loss_cons += self.epe(-flow_list[i], flow_gt[:, 2:4], 1)
+            loss_cons = loss_cons.mean() * 0.01
+        else:
+            loss_cons = torch.tensor([0])
+            loss_flow = torch.abs(warped_img0 - gt).mean()
+            loss_mask = 1
+        loss_l1 = (((pred - gt) ** 2 + 1e-6) ** 0.5).mean()
+        if training:
+            self.optimG.zero_grad()
+            loss_G = loss_l1 + loss_cons + loss_ter
+            loss_G.backward()
+            self.optimG.step()
+        return pred, merged_img, flow, loss_l1, loss_flow, loss_cons, loss_ter, loss_mask
+
+
+if __name__ == '__main__':
+    img0 = torch.zeros(3, 3, 256, 256).float().to(device)
+    img1 = torch.tensor(np.random.normal(
+        0, 1, (3, 3, 256, 256))).float().to(device)
+    imgs = torch.cat((img0, img1), 1)
+    model = Model()
+    model.eval()
+    print(model.inference(imgs).shape)

pkgs/rife-cuda/model/RIFE_HDv2.py

@@ -0,0 +1,247 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from torch.optim import AdamW
+import torch.optim as optim
+import itertools
+from model.warplayer import warp
+from torch.nn.parallel import DistributedDataParallel as DDP
+from model.IFNet_HDv2 import *
+import torch.nn.functional as F
+from model.loss import *
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+        nn.PReLU(out_planes)
+    )
+
+
+def deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes,
+                                 kernel_size=4, stride=2, padding=1, bias=True),
+        nn.PReLU(out_planes)
+    )
+
+def conv_woact(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),
+    )
+
+class Conv2(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=2):
+        super(Conv2, self).__init__()
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv(out_planes, out_planes, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        return x
+
+c = 32
+
+class ContextNet(nn.Module):
+    def __init__(self):
+        super(ContextNet, self).__init__()
+        self.conv0 = Conv2(3, c)
+        self.conv1 = Conv2(c, c)
+        self.conv2 = Conv2(c, 2*c)
+        self.conv3 = Conv2(2*c, 4*c)
+        self.conv4 = Conv2(4*c, 8*c)
+
+    def forward(self, x, flow):
+        x = self.conv0(x)
+        x = self.conv1(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False) * 0.5
+        f1 = warp(x, flow)
+        x = self.conv2(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f2 = warp(x, flow)
+        x = self.conv3(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f3 = warp(x, flow)
+        x = self.conv4(x)
+        flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear",
+                             align_corners=False) * 0.5
+        f4 = warp(x, flow)
+        return [f1, f2, f3, f4]
+
+
+class FusionNet(nn.Module):
+    def __init__(self):
+        super(FusionNet, self).__init__()
+        self.conv0 = Conv2(10, c)
+        self.down0 = Conv2(c, 2*c)
+        self.down1 = Conv2(4*c, 4*c)
+        self.down2 = Conv2(8*c, 8*c)
+        self.down3 = Conv2(16*c, 16*c)
+        self.up0 = deconv(32*c, 8*c)
+        self.up1 = deconv(16*c, 4*c)
+        self.up2 = deconv(8*c, 2*c)
+        self.up3 = deconv(4*c, c)
+        self.conv = nn.ConvTranspose2d(c, 4, 4, 2, 1)
+
+    def forward(self, img0, img1, flow, c0, c1, flow_gt):
+        warped_img0 = warp(img0, flow[:, :2])
+        warped_img1 = warp(img1, flow[:, 2:4])
+        if flow_gt == None:
+            warped_img0_gt, warped_img1_gt = None, None
+        else:
+            warped_img0_gt = warp(img0, flow_gt[:, :2])
+            warped_img1_gt = warp(img1, flow_gt[:, 2:4])
+        x = self.conv0(torch.cat((warped_img0, warped_img1, flow), 1))
+        s0 = self.down0(x)
+        s1 = self.down1(torch.cat((s0, c0[0], c1[0]), 1))
+        s2 = self.down2(torch.cat((s1, c0[1], c1[1]), 1))
+        s3 = self.down3(torch.cat((s2, c0[2], c1[2]), 1))
+        x = self.up0(torch.cat((s3, c0[3], c1[3]), 1))
+        x = self.up1(torch.cat((x, s2), 1))
+        x = self.up2(torch.cat((x, s1), 1))
+        x = self.up3(torch.cat((x, s0), 1))
+        x = self.conv(x)
+        return x, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt
+
+
+class Model:
+    def __init__(self, local_rank=-1):
+        self.flownet = IFNet()
+        self.contextnet = ContextNet()
+        self.fusionnet = FusionNet()
+        self.device()
+        self.optimG = AdamW(itertools.chain(
+            self.flownet.parameters(),
+            self.contextnet.parameters(),
+            self.fusionnet.parameters()), lr=1e-6, weight_decay=1e-5)
+        self.schedulerG = optim.lr_scheduler.CyclicLR(
+            self.optimG, base_lr=1e-6, max_lr=1e-3, step_size_up=8000, cycle_momentum=False)
+        self.epe = EPE()
+        self.ter = Ternary()
+        self.sobel = SOBEL()
+        if local_rank != -1:
+            self.flownet = DDP(self.flownet, device_ids=[
+                               local_rank], output_device=local_rank)
+            self.contextnet = DDP(self.contextnet, device_ids=[
+                                  local_rank], output_device=local_rank)
+            self.fusionnet = DDP(self.fusionnet, device_ids=[
+                                 local_rank], output_device=local_rank)
+
+    def train(self):
+        self.flownet.train()
+        self.contextnet.train()
+        self.fusionnet.train()
+
+    def eval(self):
+        self.flownet.eval()
+        self.contextnet.eval()
+        self.fusionnet.eval()
+
+    def device(self):
+        self.flownet.to(device)
+        self.contextnet.to(device)
+        self.fusionnet.to(device)
+
+    def load_model(self, path, rank):
+        def convert(param):
+            if rank == -1:
+                return {
+                    k.replace("module.", ""): v
+                    for k, v in param.items()
+                    if "module." in k
+                }
+            else:
+                return param
+        if rank <= 0:
+            self.flownet.load_state_dict(
+                convert(torch.load('{}/flownet.pkl'.format(path), map_location=device)))
+            self.contextnet.load_state_dict(
+                convert(torch.load('{}/contextnet.pkl'.format(path), map_location=device)))
+            self.fusionnet.load_state_dict(
+                convert(torch.load('{}/unet.pkl'.format(path), map_location=device)))
+
+    def save_model(self, path, rank):
+        if rank == 0:
+            torch.save(self.flownet.state_dict(), '{}/flownet.pkl'.format(path))
+            torch.save(self.contextnet.state_dict(), '{}/contextnet.pkl'.format(path))
+            torch.save(self.fusionnet.state_dict(), '{}/unet.pkl'.format(path))
+
+    def predict(self, imgs, flow, training=True, flow_gt=None, UHD=False):
+        img0 = imgs[:, :3]
+        img1 = imgs[:, 3:]
+        if UHD:
+            flow = F.interpolate(flow, scale_factor=2.0, mode="bilinear", align_corners=False) * 2.0
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        flow = F.interpolate(flow, scale_factor=2.0, mode="bilinear",
+                             align_corners=False) * 2.0
+        refine_output, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt = self.fusionnet(
+            img0, img1, flow, c0, c1, flow_gt)
+        res = torch.sigmoid(refine_output[:, :3]) * 2 - 1
+        mask = torch.sigmoid(refine_output[:, 3:4])
+        merged_img = warped_img0 * mask + warped_img1 * (1 - mask)
+        pred = merged_img + res
+        pred = torch.clamp(pred, 0, 1)
+        if training:
+            return pred, mask, merged_img, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt
+        else:
+            return pred
+
+    def inference(self, img0, img1, UHD=False):
+        imgs = torch.cat((img0, img1), 1)
+        flow, _ = self.flownet(imgs, UHD)
+        return self.predict(imgs, flow, training=False, UHD=UHD)
+
+    def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
+        for param_group in self.optimG.param_groups:
+            param_group['lr'] = learning_rate
+        if training:
+            self.train()
+        else:
+            self.eval()
+        flow, flow_list = self.flownet(imgs)
+        pred, mask, merged_img, warped_img0, warped_img1, warped_img0_gt, warped_img1_gt = self.predict(
+            imgs, flow, flow_gt=flow_gt)
+        loss_ter = self.ter(pred, gt).mean()
+        if training:
+            with torch.no_grad():
+                loss_flow = torch.abs(warped_img0_gt - gt).mean()
+                loss_mask = torch.abs(
+                    merged_img - gt).sum(1, True).float().detach()
+                loss_mask = F.interpolate(loss_mask, scale_factor=0.5, mode="bilinear",
+                                          align_corners=False).detach()
+                flow_gt = (F.interpolate(flow_gt, scale_factor=0.5, mode="bilinear",
+                                         align_corners=False) * 0.5).detach()
+            loss_cons = 0
+            for i in range(4):
+                loss_cons += self.epe(flow_list[i][:, :2], flow_gt[:, :2], 1)
+                loss_cons += self.epe(flow_list[i][:, 2:4], flow_gt[:, 2:4], 1)
+            loss_cons = loss_cons.mean() * 0.01
+        else:
+            loss_cons = torch.tensor([0])
+            loss_flow = torch.abs(warped_img0 - gt).mean()
+            loss_mask = 1
+        loss_l1 = (((pred - gt) ** 2 + 1e-6) ** 0.5).mean()
+        if training:
+            self.optimG.zero_grad()
+            loss_G = loss_l1 + loss_cons + loss_ter
+            loss_G.backward()
+            self.optimG.step()
+        return pred, merged_img, flow, loss_l1, loss_flow, loss_cons, loss_ter, loss_mask
+
+
+if __name__ == '__main__':
+    img0 = torch.zeros(3, 3, 256, 256).float().to(device)
+    img1 = torch.tensor(np.random.normal(
+        0, 1, (3, 3, 256, 256))).float().to(device)
+    imgs = torch.cat((img0, img1), 1)
+    model = Model()
+    model.eval()
+    print(model.inference(imgs).shape)

pkgs/rife-cuda/models.txt

@@ -1,4 +1,5 @@
 RIFE 1.5 - Optimized for general content
 RIFE 1.6 - Updated general model
-RIFE 1.7 - Optimized for 2D animation (Recommended)
-RIFE 1.8 - Updated 2D animation model (Experimental)
+RIFE 1.7 - Optimized for 2D animation
+RIFE 1.8 - Updated 2D animation model (Experimental)
+RIFE 2.0 - Fastest and highest quality model (Recommended)

pkgs/rife-cuda/rife.py

@@ -45,9 +45,14 @@ parser.add_argument('--exp', dest='exp', type=int, default=1)
 args = parser.parse_args()
 assert (not args.input is None)
 
-from model.RIFE_HD import Model
-model = Model()
-model.load_model(os.path.join(dname, args.model), -1)
+try:
+    from model.RIFE_HD import Model
+    model = Model()
+    model.load_model(os.path.join(dname, args.model), -1)
+except:
+    from model.RIFE_HDv2 import Model
+    model = Model()
+    model.load_model(os.path.join(dname, args.model), -1)
 model.eval()
 model.device()