def forward(cls, ctx, input, target):
"""
input (FloatTensor): n x num_classes
target (LongTensor): n, the indices of the target classes
"""
assert_equal(input.shape[0], target.shape[0])
p_star = sparsemax(input, 1)
cls.p_star = p_star.clone().detach()
loss = _omega_sparsemax(p_star)
p_star.scatter_add_(1, target.unsqueeze(1),
torch.full_like(p_star, -1))
loss += torch.einsum("ij,ij->i", p_star, input)
ctx.save_for_backward(p_star)
return loss
def forward(ctx, input, target, k=100):
"""
input (FloatTensor): n x num_classes
target (LongTensor): n, the indices of the target classes
"""
assert_equal(input.shape[0], target.shape[0])
p_star = tsallis15_topk(input, 1, k)
loss = _omega_tsallis15(p_star)
p_star.scatter_add_(1, target.unsqueeze(1),
torch.full_like(p_star, -1))
loss += torch.einsum("ij,ij->i", p_star, input)
ctx.save_for_backward(p_star)
# loss = torch.clamp(loss, min=0.0) # needed?
return loss
def forward(ctx, input, target, n_iter=50):
"""
input (FloatTensor): n x num_classes
target (LongTensor): n, the indices of the target classes
"""
assert_equal(input.shape[0], target.shape[0])
p_star = sparsemax_bisect(input, n_iter)
# this is onw done directly in sparsemax_bisect
# p_star /= p_star.sum(dim=1).unsqueeze(dim=1)
loss = _omega_sparsemax(p_star)
p_star.scatter_add_(1, target.unsqueeze(1),
torch.full_like(p_star, -1))
loss += torch.einsum("ij,ij->i", p_star, input)
ctx.save_for_backward(p_star)
# loss = torch.clamp(loss, min=0.0) # needed?
return loss
