import torch
import torch.nn as nn
import torchvision.models.vgg as vgg
from mmcv.runner import load_checkpoint
from mmedit.utils import get_root_logger
from ..registry import LOSSES
[docs]class PerceptualVGG(nn.Module):
"""VGG network used in calculating perceptual loss.
In this implementation, we allow users to choose whether use normalization
in the input feature and the type of vgg network. Note that the pretrained
path must fit the vgg type.
Args:
layer_name_list (list[str]): According to the name in this list,
forward function will return the corresponding features. This
list contains the name each layer in `vgg.feature`. An example
of this list is ['4', '10'].
vgg_type (str): Set the type of vgg network. Default: 'vgg19'.
use_input_norm (bool): If True, normalize the input image.
Importantly, the input feature must in the range [0, 1].
Default: True.
pretrained (str): Path for pretrained weights. Default:
'torchvision://vgg19'
"""
def __init__(self,
layer_name_list,
vgg_type='vgg19',
use_input_norm=True,
pretrained='torchvision://vgg19'):
super(PerceptualVGG, self).__init__()
if pretrained.startswith('torchvision://'):
assert vgg_type in pretrained
self.layer_name_list = layer_name_list
self.use_input_norm = use_input_norm
# get vgg model and load pretrained vgg weight
# remove _vgg from attributes to avoid `find_unused_parameters` bug
_vgg = getattr(vgg, vgg_type)()
self.init_weights(_vgg, pretrained)
num_layers = max(map(int, layer_name_list)) + 1
assert len(_vgg.features) >= num_layers
# only borrow layers that will be used from _vgg to avoid unused params
self.vgg_layers = _vgg.features[:num_layers]
if self.use_input_norm:
# the mean is for image with range [0, 1]
self.register_buffer(
'mean',
torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
# the std is for image with range [-1, 1]
self.register_buffer(
'std',
torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
for v in self.vgg_layers.parameters():
v.requires_grad = False
[docs] def forward(self, x):
"""Forward function.
Args:
x (Tensor): Input tensor with shape (n, c, h, w).
Returns:
Tensor: Forward results.
"""
if self.use_input_norm:
x = (x - self.mean) / self.std
output = {}
for name, module in self.vgg_layers.named_children():
x = module(x)
if name in self.layer_name_list:
output[name] = x.clone()
return output
[docs] def init_weights(self, model, pretrained):
"""Init weights.
Args:
model (nn.Module): Models to be inited.
pretrained (str): Path for pretrained weights.
"""
logger = get_root_logger()
load_checkpoint(model, pretrained, logger=logger)
[docs]@LOSSES.register_module()
class PerceptualLoss(nn.Module):
"""Perceptual loss with commonly used style loss.
Args:
layers_weights (dict): The weight for each layer of vgg feature.
Here is an example: {'4': 1., '9': 1., '18': 1.}, which means the
5th, 10th and 18th feature layer will be extracted with weight 1.0
in calculting losses.
vgg_type (str): The type of vgg network used as feature extractor.
Default: 'vgg19'.
use_input_norm (bool): If True, normalize the input image in vgg.
Default: True.
perceptual_weight (float): If `perceptual_weight > 0`, the perceptual
loss will be calculated and the loss will multiplied by the
weight. Default: 1.0.
style_weight (float): If `style_weight > 0`, the style loss will be
calculated and the loss will multiplied by the weight.
Default: 1.0.
norm_img (bool): If True, the image will be normed to [0, 1]. Note that
this is different from the `use_input_norm` which norm the input in
in forward function of vgg according to the statistics of dataset.
Importantly, the input image must be in range [-1, 1].
pretrained (str): Path for pretrained weights. Default:
'torchvision://vgg19'
"""
def __init__(self,
layer_weights,
vgg_type='vgg19',
use_input_norm=True,
perceptual_weight=1.0,
style_weight=1.0,
norm_img=True,
pretrained='torchvision://vgg19',
criterion='l1'):
super(PerceptualLoss, self).__init__()
self.norm_img = norm_img
self.perceptual_weight = perceptual_weight
self.style_weight = style_weight
self.layer_weights = layer_weights
self.vgg = PerceptualVGG(
layer_name_list=list(layer_weights.keys()),
vgg_type=vgg_type,
use_input_norm=use_input_norm,
pretrained=pretrained)
if criterion == 'l1':
self.criterion = torch.nn.L1Loss()
else:
raise NotImplementedError(
f'{criterion} criterion has not been supported in'
' this version.')
[docs] def forward(self, x, gt):
"""Forward function.
Args:
x (Tensor): Input tensor with shape (n, c, h, w).
gt (Tensor): Ground-truth tensor with shape (n, c, h, w).
Returns:
Tensor: Forward results.
"""
if self.norm_img:
x = (x + 1.) * 0.5
gt = (gt + 1.) * 0.5
# extract vgg features
x_features = self.vgg(x)
gt_features = self.vgg(gt.detach())
# calculate perceptual loss
if self.perceptual_weight > 0:
percep_loss = 0
for k in x_features.keys():
percep_loss += self.criterion(
x_features[k], gt_features[k]) * self.layer_weights[k]
percep_loss *= self.perceptual_weight
else:
percep_loss = None
# calculate style loss
if self.style_weight > 0:
style_loss = 0
for k in x_features.keys():
style_loss += self.criterion(
self._gram_mat(x_features[k]),
self._gram_mat(gt_features[k])) * self.layer_weights[k]
style_loss *= self.style_weight
else:
style_loss = None
return percep_loss, style_loss
def _gram_mat(self, x):
"""Calculate Gram matrix.
Args:
x (torch.Tensor): Tensor with shape of (n, c, h, w).
Returns:
torch.Tensor: Gram matrix.
"""
(n, c, h, w) = x.size()
features = x.view(n, c, w * h)
features_t = features.transpose(1, 2)
gram = features.bmm(features_t) / (c * h * w)
return gram