def forward(self, input, target):
bce = F.binary_cross_entropy(input, target, reduction=None)
pt = F.exp(-bce)
ones = torch.ones(target.shape,dtype=torch.float)
focal_factor = torch.pow(ones-pt,self.gamma)
W = torch.ones(target.shape,dtype=torch.float)
W[target==1] = self.w
pixelwise_loss = W * focal_factor * bce
return torch.mean(pixelwise_loss)
def fstarT(self, v):
if self.divergence_name == "kl":
return torch.exp(v - 1.0)
elif self.divergence_name == "klrev":
return -1.0 - v
# According to Table 4 of the f-GAN paper, we use Pearson Chi-Squared.
# After Pearson Chi-Squared, the next best are KL and then Jensen-Shannon.
elif self.divergence_name == "pearson":
return 0.25 * v * v + v
elif self.divergence_name == "neyman":
return 2.0 - 2.0 * F.exp(0.5 * v)
elif self.divergence_name == "hellinger":
return F.exp(-v) - 1.0
elif self.divergence_name == "jensen":
return F.softplus(v) - math.log(2.0)
elif self.divergence_name == "gan":
return F.softplus(v)
else:
raise ValueError("This is an unknown f-divergence.")
def T(self, v):
if self.divergence_name == "kl":
return v
elif self.divergence_name == "klrev":
return -F.exp(v)
# According to Table 4 of the f-GAN paper, we use Pearson Chi-Squared.
# After Pearson Chi-Squared, the next best are KL and then Jensen-Shannon.
elif self.divergence_name == "pearson":
return v
elif self.divergence_name == "neyman":
return 1.0 - F.exp(v)
elif self.divergence_name == "hellinger":
return 1.0 - F.exp(v)
elif self.divergence_name == "jensen":
return math.log(2.0) - F.softplus(-v)
elif self.divergence_name == "gan":
return -F.softplus(-v)
else:
raise ValueError("Unknown f-divergence.")
def pvh(self, v, h):
'''
get the probability of configuration in binary visual/hidden nodes.
Args:
v (1darray): visual node configuration.
h (1darray): hidden node configuration.
'''
return F.exp(
self.h.mv(self.W.mm(v)) + self.h_bias.dot(h) + self.v_bias.mm(v))
def forward(self, x):
#import pdb; pdb.set_trace()
mu = self.mu(x)
if self.std_to_positive_transform == 'softplus':
sigma = F.softplus(self.sigma_lin(x))
elif self.std_to_positive_transform == 'exp':
sigma = F.exp(self.sigma_lin(x))
else:
raise ValueError("Unknown positive transformation")
return self.reparametrize(mu, sigma, self.num_samples), mu, sigma
def forward(self, x):
aria2 = 1 + ((F.exp(-x) ** self.b) ** (-self.a))
return x * aria2
def forward(self, x):
aria = self.A + (self.k - self.A) / ((self.C + self.Q * F.exp(-x) ** self.B) ** (1/self.v))
return x * aria
def inverse_log_norm(x, min_clip, mean, std, F=torch):
x = F.exp(x * std + mean) + min_clip - 1e-5
return x
