Add files via upload

Author: ltkong218

benchmark/Vimeo90K.py

@@ -0,0 +1,55 @@
+import os
+import sys
+sys.path.append('.')
+import cv2
+import math
+import torch
+import numpy as np
+from torch.nn import functional as F
+from benchmark.pytorch_msssim import ssim_matlab as SSIM
+from models.WaveletVFI import WaveletVFI
+from thop import profile
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def convert(param):
+    return {k.replace("module.", ""): v for k, v in param.items() if "module." in k}
+
+def PSNR(img_pred, img_gt):
+    psnr = -10 * torch.log10(((img_pred - img_gt) * (img_pred - img_gt)).mean())
+    return psnr
+
+model = WaveletVFI()
+model.load_state_dict(convert(torch.load('./models/waveletvfi_latest.pth', map_location='cpu')))
+model.eval()
+model.to(device)
+
+th = None
+
+path = '/youtu_action_data/NTIRE/vimeo_triplet/'
+f = open(path + 'tri_testlist.txt', 'r')
+psnr_list = []
+ssim_list = []
+flops_list = []
+for i in f:
+    name = str(i).strip()
+    if(len(name) <= 1):
+        continue
+    print(path + 'sequences/' + name + '/im1.png')
+    I0 = cv2.imread(path + 'sequences/' + name + '/im1.png')
+    I1 = cv2.imread(path + 'sequences/' + name + '/im2.png')
+    I2 = cv2.imread(path + 'sequences/' + name + '/im3.png')
+    I0 = (torch.tensor(I0.transpose(2, 0, 1)).to(device) / 255.).unsqueeze(0)
+    I1 = (torch.tensor(I1.transpose(2, 0, 1)).to(device) / 255.).unsqueeze(0)
+    I2 = (torch.tensor(I2.transpose(2, 0, 1)).to(device) / 255.).unsqueeze(0)
+
+    macs, params, outputs = profile(model, inputs=(I0, I2, I1, False, th), verbose=False, output=True)
+    I1_pred = outputs[0]
+
+    psnr = PSNR(I1_pred, I1).detach().cpu().numpy()
+    ssim = SSIM(I1_pred, I1).detach().cpu().numpy()
+
+    psnr_list.append(psnr)
+    ssim_list.append(ssim)
+    flops_list.append(macs / 1e9)
+    print("Avg PSNR: {:.3f} SSIM: {:.4f} FLOPs(G): {:.3f}".format(np.mean(psnr_list), np.mean(ssim_list), np.mean(flops_list)))

benchmark/pytorch_msssim/__init__.py

@@ -0,0 +1,203 @@
+import torch
+import torch.nn.functional as F
+from math import exp
+import numpy as np
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)])
+    return gauss/gauss.sum()
+
+
+def create_window(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0).to(device)
+    window = _2D_window.expand(channel, 1, window_size, window_size).contiguous()
+    return window
+
+
+def create_window_3d(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t())
+    _3D_window = _2D_window.unsqueeze(2) @ (_1D_window.t())
+    window = _3D_window.expand(1, channel, window_size, window_size, window_size).contiguous().to(device)
+    return window
+
+
+def ssim(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, channel, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window(real_size, channel=channel).to(img1.device)
+    
+    # mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
+    # mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
+    mu1 = F.conv2d(F.pad(img1, (5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=channel)
+    mu2 = F.conv2d(F.pad(img2, (5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv2d(F.pad(img1 * img1, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu1_sq
+    sigma2_sq = F.conv2d(F.pad(img2 * img2, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu2_sq
+    sigma12 = F.conv2d(F.pad(img1 * img2, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def ssim_matlab(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, _, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window_3d(real_size, channel=1).to(img1.device)
+        # Channel is set to 1 since we consider color images as volumetric images
+
+    img1 = img1.unsqueeze(1)
+    img2 = img2.unsqueeze(1)
+
+    mu1 = F.conv3d(F.pad(img1, (5, 5, 5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=1)
+    mu2 = F.conv3d(F.pad(img2, (5, 5, 5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=1)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv3d(F.pad(img1 * img1, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu1_sq
+    sigma2_sq = F.conv3d(F.pad(img2 * img2, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu2_sq
+    sigma12 = F.conv3d(F.pad(img1 * img2, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def msssim(img1, img2, window_size=11, size_average=True, val_range=None, normalize=False):
+    device = img1.device
+    weights = torch.FloatTensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333]).to(device)
+    levels = weights.size()[0]
+    mssim = []
+    mcs = []
+    for _ in range(levels):
+        sim, cs = ssim(img1, img2, window_size=window_size, size_average=size_average, full=True, val_range=val_range)
+        mssim.append(sim)
+        mcs.append(cs)
+
+        img1 = F.avg_pool2d(img1, (2, 2))
+        img2 = F.avg_pool2d(img2, (2, 2))
+
+    mssim = torch.stack(mssim)
+    mcs = torch.stack(mcs)
+
+    # Normalize (to avoid NaNs during training unstable models, not compliant with original definition)
+    if normalize:
+        mssim = (mssim + 1) / 2
+        mcs = (mcs + 1) / 2
+
+    pow1 = mcs ** weights
+    pow2 = mssim ** weights
+    # From Matlab implementation https://ece.uwaterloo.ca/~z70wang/research/iwssim/
+    output = torch.prod(pow1[:-1] * pow2[-1])
+    return output
+
+
+# Classes to re-use window
+class SSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, val_range=None):
+        super(SSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.val_range = val_range
+
+        # Assume 3 channel for SSIM
+        self.channel = 3
+        self.window = create_window(window_size, channel=self.channel)
+
+    def forward(self, img1, img2):
+        (_, channel, _, _) = img1.size()
+
+        if channel == self.channel and self.window.dtype == img1.dtype:
+            window = self.window
+        else:
+            window = create_window(self.window_size, channel).to(img1.device).type(img1.dtype)
+            self.window = window
+            self.channel = channel
+
+        _ssim = ssim(img1, img2, window=window, window_size=self.window_size, size_average=self.size_average)
+        dssim = (1 - _ssim) / 2
+        return dssim
+
+
+class MSSSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, channel=3):
+        super(MSSSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = channel
+
+    def forward(self, img1, img2):
+        return msssim(img1, img2, window_size=self.window_size, size_average=self.size_average)

models/IFNet.py

@@ -0,0 +1,114 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from models.utils import bwarp
+
+
+def centralize(img0, img1):
+    rgb_mean = torch.cat([img0, img1], 2).mean(1, keepdim=True).mean(2, keepdim=True).mean(3, keepdim=True)
+    return img0 - rgb_mean, img1 - rgb_mean, rgb_mean
+
+
+def resize(x, scale_factor):
+    return F.interpolate(x, scale_factor=scale_factor, mode='bilinear', align_corners=False)
+
+
+def convrelu(in_channels, out_channels, kernel_size=3, stride=1, padding=1, dilation=1, groups=1, bias=True):
+    return nn.Sequential(
+        nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias=bias), 
+        nn.LeakyReLU(negative_slope=0.1)
+    )
+
+
+class Decoder(nn.Module):
+    def __init__(self, in_channels, mid_channels, is_bottom=False):
+        super(Decoder, self).__init__()
+        self.is_bottom = is_bottom
+        self.conv1 = convrelu(in_channels, mid_channels, 3, 1)
+        self.conv2 = convrelu(mid_channels, mid_channels, 3, 1)
+        self.conv3 = convrelu(mid_channels, mid_channels, 3, 1)
+        self.conv4 = convrelu(mid_channels, mid_channels, 3, 1)
+        self.conv5 = convrelu(mid_channels, mid_channels, 3, 1)
+        self.conv6 = nn.ConvTranspose2d(mid_channels, 5, 4, 2, 1, bias=True)
+        if self.is_bottom:
+            self.classifier = nn.Sequential(
+                convrelu(mid_channels, mid_channels, 3, 2), 
+                nn.AdaptiveAvgPool2d((1, 1)), 
+                nn.Flatten(1), 
+                nn.Linear(mid_channels, mid_channels, bias=True), 
+                nn.LeakyReLU(negative_slope=0.1), 
+                nn.Linear(mid_channels, 4, bias=True)
+            )
+
+    def forward(self, x):
+        out1 = self.conv1(x)
+        out2 = self.conv2(out1)
+        out3 = self.conv3(out2)
+        out4 = self.conv4(out3)
+        out5 = self.conv5(out4)
+        out = self.conv6(out5)
+        if self.is_bottom:
+            class_prob_ = self.classifier(out5)
+            return out, class_prob_
+        else:
+            return out
+
+
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.pconv1 = nn.Sequential(convrelu(3, 32, 3, 2), convrelu(32, 32, 3, 1))
+        self.pconv2 = nn.Sequential(convrelu(32, 64, 3, 2), convrelu(64, 64, 3, 1))
+        self.pconv3 = nn.Sequential(convrelu(64, 96, 3, 2), convrelu(96, 96, 3, 1))
+        self.pconv4 = nn.Sequential(convrelu(96, 128, 3, 2), convrelu(128, 128, 3, 1))
+
+        self.decoder4 = Decoder(256, 192, True)
+        self.decoder3 = Decoder(197, 160, False)
+        self.decoder2 = Decoder(133, 128, False)
+        self.decoder1 = Decoder(69, 64, False)
+        
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
+                nn.init.kaiming_normal_(m.weight)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, img0, img1):
+        img0, img1, _ = centralize(img0, img1)
+
+        f0_1 = self.pconv1(img0)
+        f1_1 = self.pconv1(img1)
+        f0_2 = self.pconv2(f0_1)
+        f1_2 = self.pconv2(f1_1)
+        f0_3 = self.pconv3(f0_2)
+        f1_3 = self.pconv3(f1_2)
+        f0_4 = self.pconv4(f0_3)
+        f1_4 = self.pconv4(f1_3)
+
+        out4, class_prob_ = self.decoder4(torch.cat([f0_4, f1_4], 1))
+        up_flow_t0_4 = out4[:, 0:2]
+        up_flow_t1_4 = out4[:, 2:4]
+        up_occ_t_4 = out4[:, 4:5]
+
+        f0_3_warp = bwarp(f0_3, up_flow_t0_4)
+        f1_3_warp = bwarp(f1_3, up_flow_t1_4)
+        out3 = self.decoder3(torch.cat([f0_3_warp, f1_3_warp, up_flow_t0_4, up_flow_t1_4, up_occ_t_4], 1))
+        up_flow_t0_3 = out3[:, 0:2] + resize(up_flow_t0_4, 2.0) * 2.0
+        up_flow_t1_3 = out3[:, 2:4] + resize(up_flow_t1_4, 2.0) * 2.0
+        up_occ_t_3 = out3[:, 4:5] + resize(up_occ_t_4, 2.0)
+
+        f0_2_warp = bwarp(f0_2, up_flow_t0_3)
+        f1_2_warp = bwarp(f1_2, up_flow_t1_3)
+        out2 = self.decoder2(torch.cat([f0_2_warp, f1_2_warp, up_flow_t0_3, up_flow_t1_3, up_occ_t_3], 1))
+        up_flow_t0_2 = out2[:, 0:2] + resize(up_flow_t0_3, 2.0) * 2.0
+        up_flow_t1_2 = out2[:, 2:4] + resize(up_flow_t1_3, 2.0) * 2.0
+        up_occ_t_2 = out2[:, 4:5] + resize(up_occ_t_3, 2.0)
+
+        f0_1_warp = bwarp(f0_1, up_flow_t0_2)
+        f1_1_warp = bwarp(f1_1, up_flow_t1_2)
+        out1 = self.decoder1(torch.cat([f0_1_warp, f1_1_warp, up_flow_t0_2, up_flow_t1_2, up_occ_t_2], 1))
+        up_flow_t0_1 = out1[:, 0:2] + resize(up_flow_t0_2, 2.0) * 2.0
+        up_flow_t1_1 = out1[:, 2:4] + resize(up_flow_t1_2, 2.0) * 2.0
+        up_occ_t_1 = torch.sigmoid(out1[:, 4:5] + resize(up_occ_t_2, 2.0))
+
+        return up_flow_t0_1, up_flow_t1_1, up_occ_t_1, class_prob_