def forward(self, input):
b, c, h, w = input.size()
F = input.view(b, c, h * w)
mean_ = F.mean(dim=2, keepdim=True).detach()
mean = torch.cat(h * w * [mean_], 2)
F = F - mean.detach()
G = torch.bmm(F, F.transpose(1, 2))
G.div_(h * w)
return G.squeeze(0).data, mean_.squeeze().data
def correlation_matrix(F: np.ndarray,
a_min: float = -1.,
a_max: float = 1.) -> np.ndarray:
F_c = F - F.mean(axis=1)[:, np.newaxis]
cov = F_c @ F_c.T
l2_norms = np.linalg.norm(F_c, axis=1) #compute l2-norm across rows
denom = np.outer(l2_norms, l2_norms)
corr_mat = (cov / denom).clip(
min=a_min, max=a_max) #counteract potential rounding errors
return corr_mat
def correlation_matrix(F: np.ndarray,
a_min: float = -1.,
a_max: float = 1.) -> np.ndarray:
"""Compute dissimilarity matrix based on correlation distance (on the matrix-level)."""
F_c = F - F.mean(axis=1)[:, np.newaxis]
cov = F_c @ F_c.T
# compute vector l2-norm across rows
l2_norms = np.linalg.norm(F_c, axis=1)
denom = np.outer(l2_norms, l2_norms)
corr_mat = (cov / denom).clip(min=a_min, max=a_max)
return corr_mat
def loss_dcgan_gen(dis_fake):
loss = F.mean(F.softplus(-dis_fake))
return loss
def bar_entropy(data):
return F.mean(tilde_entropy(data))