perceptual_model.py

import torch
import torch.nn as nn
import torchvision
from torchvision import models

class VGG16_perceptual(torch.nn.Module):
    def __init__(self, requires_grad=False):
        super(VGG16_perceptual, self).__init__()
        vgg_pretrained_features = models.vgg16(pretrained=True).features
        self.slice1 = torch.nn.Sequential()
        self.slice2 = torch.nn.Sequential()
        self.slice3 = torch.nn.Sequential()
        self.slice4 = torch.nn.Sequential()
        for x in range(2):
            self.slice1.add_module(str(x), vgg_pretrained_features[x])
        for x in range(2, 4):
            self.slice2.add_module(str(x), vgg_pretrained_features[x])
        for x in range(4, 14):
            self.slice3.add_module(str(x), vgg_pretrained_features[x])
        for x in range(14, 21):
            self.slice4.add_module(str(x), vgg_pretrained_features[x])
        if not requires_grad:
            for param in self.parameters():
                param.requires_grad = False

    def forward(self, X):
        h = self.slice1(X)
        h_relu1_1 = h
        h = self.slice2(h)
        h_relu1_2 = h
        h = self.slice3(h)
        h_relu3_2 = h
        h = self.slice4(h)
        h_relu4_2 = h
        return h_relu1_1, h_relu1_2, h_relu3_2, h_relu4_2