Spaces:
Sleeping
Sleeping
File size: 3,618 Bytes
df60d6b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 |
# Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# This work is licensed under a Creative Commons
# Attribution-NonCommercial-ShareAlike 4.0 International License.
# You should have received a copy of the license along with this
# work. If not, see http://creativecommons.org/licenses/by-nc-sa/4.0/
"""Loss functions used in the paper
"Elucidating the Design Space of Diffusion-Based Generative Models"."""
import torch
from edm.torch_utils import persistence
#----------------------------------------------------------------------------
# Loss function corresponding to the variance preserving (VP) formulation
# from the paper "Score-Based Generative Modeling through Stochastic
# Differential Equations".
@persistence.persistent_class
class VPLoss:
def __init__(self, beta_d=19.9, beta_min=0.1, epsilon_t=1e-5):
self.beta_d = beta_d
self.beta_min = beta_min
self.epsilon_t = epsilon_t
def __call__(self, net, images, labels, augment_pipe=None):
rnd_uniform = torch.rand([images.shape[0], 1, 1, 1], device=images.device)
sigma = self.sigma(1 + rnd_uniform * (self.epsilon_t - 1))
weight = 1 / sigma ** 2
y, augment_labels = augment_pipe(images) if augment_pipe is not None else (images, None)
n = torch.randn_like(y) * sigma
D_yn = net(y + n, sigma, labels, augment_labels=augment_labels)
loss = weight * ((D_yn - y) ** 2)
return loss
def sigma(self, t):
t = torch.as_tensor(t)
return ((0.5 * self.beta_d * (t ** 2) + self.beta_min * t).exp() - 1).sqrt()
#----------------------------------------------------------------------------
# Loss function corresponding to the variance exploding (VE) formulation
# from the paper "Score-Based Generative Modeling through Stochastic
# Differential Equations".
@persistence.persistent_class
class VELoss:
def __init__(self, sigma_min=0.02, sigma_max=100):
self.sigma_min = sigma_min
self.sigma_max = sigma_max
def __call__(self, net, images, labels, augment_pipe=None):
rnd_uniform = torch.rand([images.shape[0], 1, 1, 1], device=images.device)
sigma = self.sigma_min * ((self.sigma_max / self.sigma_min) ** rnd_uniform)
weight = 1 / sigma ** 2
y, augment_labels = augment_pipe(images) if augment_pipe is not None else (images, None)
n = torch.randn_like(y) * sigma
D_yn = net(y + n, sigma, labels, augment_labels=augment_labels)
loss = weight * ((D_yn - y) ** 2)
return loss
#----------------------------------------------------------------------------
# Improved loss function proposed in the paper "Elucidating the Design Space
# of Diffusion-Based Generative Models" (EDM).
@persistence.persistent_class
class EDMLoss:
def __init__(self, P_mean=-1.2, P_std=1.2, sigma_data=0.5):
self.P_mean = P_mean
self.P_std = P_std
self.sigma_data = sigma_data
def __call__(self, net, images, labels=None, augment_pipe=None):
rnd_normal = torch.randn([images.shape[0], 1, 1, 1], device=images.device)
sigma = (rnd_normal * self.P_std + self.P_mean).exp()
weight = (sigma ** 2 + self.sigma_data ** 2) / (sigma * self.sigma_data) ** 2
y, augment_labels = augment_pipe(images) if augment_pipe is not None else (images, None)
n = torch.randn_like(y) * sigma
D_yn = net(y + n, sigma, labels, augment_labels=augment_labels)
loss = weight * ((D_yn - y) ** 2)
return loss
#----------------------------------------------------------------------------
|