Spaces:
Running on Zero
Running on Zero
File size: 4,788 Bytes
a95f6c0 | 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 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 | import abc
class SDE():
"""
Definition of the diffusion following the parameterization as in ( Karras et al., "Elucidating...", 2022).
This includes only the utilities needed for training, not for sampling.
"""
def __init__(self,
type,
sde_hp):
self.type = type
self.sde_hp = sde_hp
@abc.abstractmethod
def sample_time_training(self,N):
"""
For training, getting t according to a similar criteria as sampling.
Args:
N (int): batch size
"""
pass
@abc.abstractmethod
def sample_prior(self, shape, *args, **kwargs):
"""
Just sample some gaussian noise, nothing more
Args:
shape (tuple): shape of the noise to sample, something like (B,T)
"""
pass
@abc.abstractmethod
def cskip(self, sigma, *args, **kwargs):
"""
Just one of the preconditioning parameters
"""
pass
@abc.abstractmethod
def cout(self, sigma, *args, **kwargs):
"""
Just one of the preconditioning parameters
Args:
sigma (float): noise level (equal to timestep is sigma=t, which is our default)
"""
pass
@abc.abstractmethod
def cin(self, sigma, *args, **kwargs):
"""
Just one of the preconditioning parameters
"""
pass
@abc.abstractmethod
def cnoise(self, sigma, *args, **kwargs):
"""
preconditioning of the noise embedding
Args:
sigma (float): noise level (equal to timestep is sigma=t, which is our default)
"""
pass
@abc.abstractmethod
def lambda_w(self, sigma, *args, **kwargs):
"""
Score matching loss weighting
"""
pass
@abc.abstractmethod
def _mean(self, *args, **kwargs):
pass
@abc.abstractmethod
def _std(self, *args, **kwargs):
pass
@abc.abstractmethod
def _ode_integrand(self, *args, **kwargs):
pass
@abc.abstractmethod
def Tweedie2score(self, tweedie, xt, t, *args, **kwargs):
pass
@abc.abstractmethod
def score2Tweedie(self, score, xt, t, *args, **kwargs):
pass
def denoiser(self, xn , net, t, *args, **kwargs):
"""
This method does the whole denoising step, which implies applying the model and the preconditioning
Args:
x (Tensor): shape: (B,1,T) Intermediate noisy latent to denoise
model (nn.Module): Model of the denoiser
sigma (float): noise level (equal to timestep is sigma=t, which is our default)
"""
sigma = self._std(t).unsqueeze(-1)
sigma = sigma.view(*sigma.size(), *(1,)*(xn.ndim - sigma.ndim))
cskip = self.cskip(sigma)
cout = self.cout(sigma)
cin = self.cin(sigma)
cnoise = self.cnoise(sigma.squeeze())
#check if cnoise is a scalar, if so, repeat it
if len(cnoise.shape) == 0:
cnoise = cnoise.repeat(xn.shape[0],)
else:
cnoise = cnoise.view(xn.shape[0],)
return cskip * xn + cout * net(cin * xn, cnoise) #this will crash because of broadcasting problems, debug later!
def prepare_train_preconditioning(self, x, t,n=None, *args, **kwargs):
mu, sigma = self._mean(x, t), self._std(t).unsqueeze(-1)
sigma = sigma.view(*sigma.size(), *(1,)*(x.ndim - sigma.ndim))
if n is None:
n=self.sample_prior(x.shape).to(x.device)
x_perturbed = mu + sigma *n
#self.sample_prior(x.shape).to(x.device)
cskip = self.cskip(sigma)
cout = self.cout(sigma)
cin = self.cin(sigma)
cnoise = self.cnoise(sigma.squeeze())
#check if cnoise is a scalar, if so, repeat it
if len(cnoise.shape) == 0:
cnoise = cnoise.repeat(x.shape[0],)
else:
cnoise = cnoise.view(x.shape[0],)
target = 1/cout * (x - cskip * x_perturbed)
return cin * x_perturbed, target, cnoise
def loss_fn(self, net, x,n=None, *args, **kwargs):
"""
Loss function, which is the mean squared error between the denoised latent and the clean latent
Args:
net (nn.Module): Model of the denoiser
x (Tensor): shape: (B,T) Intermediate noisy latent to denoise
sigma (float): noise level (equal to timestep is sigma=t, which is our default)
"""
t = self.sample_time_training(x.shape[0]).to(x.device)
input, target, cnoise = self.prepare_train_preconditioning(x, t, n=n)
estimate = net(input.unsqueeze(1), cnoise).squeeze(1)
error = estimate - target
return error**2, self._std(t)
|