About ohem

[IssacCyj]

loss_c = log_sum_exp(batch_conf) - batch_conf.gather(1, conf_t.view(-1, 1))

I'm confused with this line of code. Why not using conf_data to select hard examples? What is this loss_c stand for?

[YangQun1]

@IssacCyj I'm confused with this line of code too. Do you understand why use loss_c instead of conf_data here now?

[JingyuanHu]

I'm confused about this process, too. Maybe someone can explain this.

[DonghoonPark12]

I'm also confused.

I don't understand why author minus below term. - batch_conf.gather(1, conf_t.view(-1, 1))

[lujiazho]

loss_c = log_sum_exp(batch_conf) - batch_conf.gather(1, conf_t.view(-1, 1))

I guess this author considers log_sum_exp part as an approximation to maximum function, which represents predicted confidence. And batch_conf.gather(1, conf_t.view(-1, 1)) just pick out target classes' confidence.

It's just my guess, but you can change it to torch.nn.functional.cross_entropy(batch_conf, conf_t.view(-1), reduction='none') or the like. Although the eventual rank would be a little different.