Skip to content

Fix running with pytorch-cuda. #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
veluca93 wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Fix running with pytorch-cuda. #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
104 changes: 63 additions & 41 deletions IQA_pytorch/DISTS.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,43 +2,53 @@
import os
import sys
import torch
from torchvision import models,transforms
from torchvision import models, transforms
import torch.nn as nn
import torch.nn.functional as F
import inspect
from .utils import downsample


class L2pooling(nn.Module):
def __init__(self, filter_size=5, stride=2, channels=None, pad_off=0):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
super(L2pooling, self).__init__()
self.padding = (filter_size - 2 )//2
self.padding = (filter_size - 2) // 2
self.stride = stride
self.channels = channels
a = np.hanning(filter_size)[1:-1]
# a = torch.hann_window(5,periodic=False)
g = torch.Tensor(a[:,None]*a[None,:])
g = g/torch.sum(g)
self.register_buffer('filter', g[None,None,:,:].repeat((self.channels,1,1,1)))
g = torch.Tensor(a[:, None] * a[None, :]).to(device)
g = g / torch.sum(g)
self.register_buffer(
'filter', g[None, None, :, :].repeat((self.channels, 1, 1, 1)))

def forward(self, input):
input = input**2
out = F.conv2d(input, self.filter, stride=self.stride, padding=self.padding, groups=input.shape[1])
return (out+1e-12).sqrt()
out = F.conv2d(input,
self.filter,
stride=self.stride,
padding=self.padding,
groups=input.shape[1])
return (out + 1e-12).sqrt()


class DISTS(torch.nn.Module):
'''
Refer to https://github.com/dingkeyan93/DISTS
'''
def __init__(self, channels=3, load_weights=True):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
assert channels == 3
super(DISTS, self).__init__()
vgg_pretrained_features = models.vgg16(pretrained=True).features
vgg_pretrained_features = models.vgg16(
pretrained=True).features.to(device)
self.stage1 = torch.nn.Sequential()
self.stage2 = torch.nn.Sequential()
self.stage3 = torch.nn.Sequential()
self.stage4 = torch.nn.Sequential()
self.stage5 = torch.nn.Sequential()
for x in range(0,4):
for x in range(0, 4):
self.stage1.add_module(str(x), vgg_pretrained_features[x])
self.stage2.add_module(str(4), L2pooling(channels=64))
for x in range(5, 9):
Expand All @@ -52,25 +62,36 @@ def __init__(self, channels=3, load_weights=True):
self.stage5.add_module(str(23), L2pooling(channels=512))
for x in range(24, 30):
self.stage5.add_module(str(x), vgg_pretrained_features[x])

for param in self.parameters():
param.requires_grad = False

self.register_buffer("mean", torch.tensor([0.485, 0.456, 0.406]).view(1,-1,1,1))
self.register_buffer("std", torch.tensor([0.229, 0.224, 0.225]).view(1,-1,1,1))

self.chns = [3,64,128,256,512,512]
self.register_parameter("alpha", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))
self.register_parameter("beta", nn.Parameter(torch.randn(1, sum(self.chns),1,1)))
self.alpha.data.normal_(0.1,0.01)
self.beta.data.normal_(0.1,0.01)
self.register_buffer(
"mean",
torch.tensor([0.485, 0.456, 0.406]).to(device).view(1, -1, 1, 1))
self.register_buffer(
"std",
torch.tensor([0.229, 0.224, 0.225]).to(device).view(1, -1, 1, 1))

self.chns = [3, 64, 128, 256, 512, 512]
self.register_parameter(
"alpha",
nn.Parameter(torch.randn(1, sum(self.chns), 1, 1).to(device)))
self.register_parameter(
"beta",
nn.Parameter(torch.randn(1, sum(self.chns), 1, 1).to(device)))
self.alpha.data.normal_(0.1, 0.01)
self.beta.data.normal_(0.1, 0.01)
if load_weights:
weights = torch.load(os.path.abspath(os.path.join(inspect.getfile(DISTS),'..','weights/DISTS.pt')))
weights = torch.load(
os.path.abspath(
os.path.join(inspect.getfile(DISTS), '..',
'weights/DISTS.pt')))
self.alpha.data = weights['alpha']
self.beta.data = weights['beta']

def forward_once(self, x):
h = (x-self.mean)/self.std
h = (x - self.mean) / self.std
h = self.stage1(h)
h_relu1_2 = h
h = self.stage2(h)
Expand All @@ -81,39 +102,41 @@ def forward_once(self, x):
h_relu4_3 = h
h = self.stage5(h)
h_relu5_3 = h
return [x,h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3]
return [x, h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3, h_relu5_3]

def forward(self, x, y, as_loss=True, resize = True):
def forward(self, x, y, as_loss=True, resize=True):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
assert x.shape == y.shape
if resize:
x, y = downsample(x, y)
if as_loss:
feats0 = self.forward_once(x)
feats1 = self.forward_once(y)
feats1 = self.forward_once(y)
else:
with torch.no_grad():
feats0 = self.forward_once(x)
feats1 = self.forward_once(y)
dist1 = 0
dist2 = 0
feats1 = self.forward_once(y)
dist1 = 0
dist2 = 0
c1 = 1e-6
c2 = 1e-6
w_sum = self.alpha.sum() + self.beta.sum()
alpha = torch.split(self.alpha/w_sum, self.chns, dim=1)
beta = torch.split(self.beta/w_sum, self.chns, dim=1)
alpha = torch.split(self.alpha / w_sum, self.chns, dim=1)
beta = torch.split(self.beta / w_sum, self.chns, dim=1)
for k in range(len(self.chns)):
x_mean = feats0[k].mean([2,3], keepdim=True)
y_mean = feats1[k].mean([2,3], keepdim=True)
S1 = (2*x_mean*y_mean+c1)/(x_mean**2+y_mean**2+c1)
dist1 = dist1+(alpha[k]*S1).sum(1,keepdim=True)

x_var = ((feats0[k]-x_mean)**2).mean([2,3], keepdim=True)
y_var = ((feats1[k]-y_mean)**2).mean([2,3], keepdim=True)
xy_cov = (feats0[k]*feats1[k]).mean([2,3],keepdim=True) - x_mean*y_mean
S2 = (2*xy_cov+c2)/(x_var+y_var+c2)
dist2 = dist2+(beta[k]*S2).sum(1,keepdim=True)

score = 1 - (dist1+dist2).squeeze()
x_mean = feats0[k].mean([2, 3], keepdim=True)
y_mean = feats1[k].mean([2, 3], keepdim=True)
S1 = (2 * x_mean * y_mean + c1) / (x_mean**2 + y_mean**2 + c1)
dist1 = dist1 + (alpha[k].to(device) * S1).sum(1, keepdim=True)

x_var = ((feats0[k] - x_mean)**2).mean([2, 3], keepdim=True)
y_var = ((feats1[k] - y_mean)**2).mean([2, 3], keepdim=True)
xy_cov = (feats0[k] * feats1[k]).mean(
[2, 3], keepdim=True) - x_mean * y_mean
S2 = (2 * xy_cov + c2) / (x_var + y_var + c2)
dist2 = dist2 + (beta[k].to(device) * S2).sum(1, keepdim=True)

score = 1 - (dist1 + dist2).squeeze()
if as_loss:
return score.mean()
else:
Expand Down Expand Up @@ -141,4 +164,3 @@ def forward(self, x, y, as_loss=True, resize = True):
score = model(ref, dist, as_loss=False)
print('score: %.4f' % score.item())
# score: 0.3347

46 changes: 30 additions & 16 deletions IQA_pytorch/GMSD.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,34 +4,48 @@
import numpy as np
from torchvision import transforms


class GMSD(nn.Module):
# Refer to http://www4.comp.polyu.edu.hk/~cslzhang/IQA/GMSD/GMSD.htm

def __init__(self, channels=3):
super(GMSD, self).__init__()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.channels = channels
dx = (torch.Tensor([[1,0,-1],[1,0,-1],[1,0,-1]])/3.).unsqueeze(0).unsqueeze(0).repeat(channels,1,1,1)
dy = (torch.Tensor([[1,1,1],[0,0,0],[-1,-1,-1]])/3.).unsqueeze(0).unsqueeze(0).repeat(channels,1,1,1)
dx = (torch.Tensor([[1, 0, -1], [1, 0, -1], [1, 0, -1]]).to(device) /
3.).unsqueeze(0).unsqueeze(0).repeat(channels, 1, 1, 1)
dy = (torch.Tensor([[1, 1, 1], [0, 0, 0], [-1, -1, -1]]).to(device) /
3.).unsqueeze(0).unsqueeze(0).repeat(channels, 1, 1, 1)
self.dx = nn.Parameter(dx, requires_grad=False)
self.dy = nn.Parameter(dy, requires_grad=False)
self.aveKernel = nn.Parameter(torch.ones(channels,1,2,2)/4., requires_grad=False)
self.aveKernel = nn.Parameter(
torch.ones(channels, 1, 2, 2).to(device) / 4., requires_grad=False)

def gmsd(self, img1, img2, T=170):
Y1 = F.conv2d(img1, self.aveKernel, stride=2, padding =0, groups = self.channels)
Y2 = F.conv2d(img2, self.aveKernel, stride=2, padding =0, groups = self.channels)
Y1 = F.conv2d(img1,
self.aveKernel,
stride=2,
padding=0,
groups=self.channels)
Y2 = F.conv2d(img2,
self.aveKernel,
stride=2,
padding=0,
groups=self.channels)

IxY1 = F.conv2d(Y1, self.dx, stride=1, padding=1, groups=self.channels)
IyY1 = F.conv2d(Y1, self.dy, stride=1, padding=1, groups=self.channels)
gradientMap1 = torch.sqrt(IxY1**2 + IyY1**2 + 1e-12)

IxY1 = F.conv2d(Y1, self.dx, stride=1, padding =1, groups = self.channels)
IyY1 = F.conv2d(Y1, self.dy, stride=1, padding =1, groups = self.channels)
gradientMap1 = torch.sqrt(IxY1**2 + IyY1**2+1e-12)
IxY2 = F.conv2d(Y2, self.dx, stride=1, padding=1, groups=self.channels)
IyY2 = F.conv2d(Y2, self.dy, stride=1, padding=1, groups=self.channels)
gradientMap2 = torch.sqrt(IxY2**2 + IyY2**2 + 1e-12)

IxY2 = F.conv2d(Y2, self.dx, stride=1, padding =1, groups = self.channels)
IyY2 = F.conv2d(Y2, self.dy, stride=1, padding =1, groups = self.channels)
gradientMap2 = torch.sqrt(IxY2**2 + IyY2**2+1e-12)

quality_map = (2*gradientMap1*gradientMap2 + T)/(gradientMap1**2+gradientMap2**2 + T)
score = torch.std(quality_map.view(quality_map.shape[0],-1),dim=1)
quality_map = (2 * gradientMap1 * gradientMap2 +
T) / (gradientMap1**2 + gradientMap2**2 + T)
score = torch.std(quality_map.view(quality_map.shape[0], -1), dim=1)
return score

def forward(self, y, x, as_loss=True):
assert x.shape == y.shape
x = x * 255
Expand All @@ -44,6 +58,7 @@ def forward(self, y, x, as_loss=True):
score = self.gmsd(x, y)
return score


if __name__ == '__main__':
from PIL import Image
import argparse
Expand All @@ -64,4 +79,3 @@ def forward(self, y, x, as_loss=True):
score = model(ref, dist, as_loss=False)
print('score: %.4f' % score.item())
# score: 0.1907

41 changes: 25 additions & 16 deletions IQA_pytorch/LPIPSvgg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@
import os
import sys
import torch
from torchvision import models,transforms
from torchvision import models, transforms
import torch.nn as nn
import torch.nn.functional as F
import inspect
Expand All @@ -11,16 +11,17 @@
class LPIPSvgg(torch.nn.Module):
def __init__(self, channels=3):
# Refer to https://github.com/richzhang/PerceptualSimilarity

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
assert channels == 3
super(LPIPSvgg, self).__init__()
vgg_pretrained_features = models.vgg16(pretrained=True).features
vgg_pretrained_features = models.vgg16(
pretrained=True).features.to(device)
self.stage1 = torch.nn.Sequential()
self.stage2 = torch.nn.Sequential()
self.stage3 = torch.nn.Sequential()
self.stage4 = torch.nn.Sequential()
self.stage5 = torch.nn.Sequential()
for x in range(0,4):
for x in range(0, 4):
self.stage1.add_module(str(x), vgg_pretrained_features[x])
for x in range(4, 9):
self.stage2.add_module(str(x), vgg_pretrained_features[x])
Expand All @@ -30,19 +31,26 @@ def __init__(self, channels=3):
self.stage4.add_module(str(x), vgg_pretrained_features[x])
for x in range(23, 30):
self.stage5.add_module(str(x), vgg_pretrained_features[x])

for param in self.parameters():
param.requires_grad = False

self.register_buffer("mean", torch.tensor([0.485, 0.456, 0.406]).view(1,-1,1,1))
self.register_buffer("std", torch.tensor([0.229, 0.224, 0.225]).view(1,-1,1,1))
self.register_buffer(
"mean",
torch.tensor([0.485, 0.456, 0.406]).to(device).view(1, -1, 1, 1))
self.register_buffer(
"std",
torch.tensor([0.229, 0.224, 0.225]).to(device).view(1, -1, 1, 1))

self.chns = [64,128,256,512,512]
self.weights = torch.load(os.path.abspath(os.path.join(inspect.getfile(LPIPSvgg),'..','weights/LPIPSvgg.pt')))
self.chns = [64, 128, 256, 512, 512]
self.weights = torch.load(
os.path.abspath(
os.path.join(inspect.getfile(LPIPSvgg), '..',
'weights/LPIPSvgg.pt')))
self.weights = list(self.weights.items())

def forward_once(self, x):
h = (x-self.mean)/self.std
h = (x - self.mean) / self.std
h = self.stage1(h)
h_relu1_2 = h
h = self.stage2(h)
Expand All @@ -62,19 +70,21 @@ def forward(self, x, y, as_loss=True):
assert x.shape == y.shape
if as_loss:
feats0 = self.forward_once(x)
feats1 = self.forward_once(y)
feats1 = self.forward_once(y)
else:
with torch.no_grad():
feats0 = self.forward_once(x)
feats1 = self.forward_once(y)
score = 0
feats1 = self.forward_once(y)
score = 0
for k in range(len(self.chns)):
score = score + (self.weights[k][1]*(feats0[k]-feats1[k])**2).mean([2,3]).sum(1)
score = score + (self.weights[k][1] *
(feats0[k] - feats1[k])**2).mean([2, 3]).sum(1)
if as_loss:
return score.mean()
else:
return score


if __name__ == '__main__':
from PIL import Image
import argparse
Expand All @@ -95,4 +105,3 @@ def forward(self, x, y, as_loss=True):
score = model(ref, dist, as_loss=False)
print('score: %.4f' % score.item())
# score: 0.5435

Loading