Implementation of S_m^k?

[BrandonHanx]

DCML/architectures/multifeature_resnet50.py

Lines 61 to 70 in 01a7220

	for i in range(self.pars.compos_num):
	prob.append(torch.nn.functional.softmax(self.compos_layer[i](detach_embedding), dim = 1))
	prob0 = prob[i]
	log_prob = -torch.log(prob0)
	index = prob0.max(dim = 1, keepdim = True)[1]
	supervise = torch.zeros_like(prob0).scatter_(1, index, 1.0)
	supervise_loss = supervise_loss + torch.sum(log_prob*supervise)/x.size(0)
	for i in range(self.pars.compos_num):
	mod_x.append(soft_select(x_embedding, prob[i], self.pars.ensemble_num, self.pars.embed_dim))

There is a difference between your paper and your code.
According to this code block, it seems like you didn't implement S_m^k. This causes C_m^k directly equals to T_m^k.

Please point me out if I were wrong.