[PyTorch] pre-trained VGG16 for perceptual loss. e.g. Style Transfer

vgg.py

"""Modified VGG16 to compute perceptual loss.

This class is mostly copied from pytorch/examples.
See, fast_neural_style in https://github.com/pytorch/examples.
"""
import torch
from torchvision import models


class VGG_OUTPUT(object):

    def __init__(self, relu1_2, relu2_2, relu3_3, relu4_3):
        self.__dict__ = locals()


class VGG16(torch.nn.Module):
    def __init__(self, requires_grad=False):
        super(VGG16, 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(4):
            self.slice1.add_module(str(x), vgg_pretrained_features[x])
        for x in range(4, 9):
            self.slice2.add_module(str(x), vgg_pretrained_features[x])
        for x in range(9, 16):
            self.slice3.add_module(str(x), vgg_pretrained_features[x])
        for x in range(16, 23):
            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_2 = h
        h = self.slice2(h)
        h_relu2_2 = h
        h = self.slice3(h)
        h_relu3_3 = h
        h = self.slice4(h)
        h_relu4_3 = h
        return VGG_OUTPUT(h_relu1_2, h_relu2_2, h_relu3_3, h_relu4_3)

https://github.com/pytorch/examples/blob/master/fast_neural_style/neural_style/vgg.py