prob question

[Muran337287]

            preds0, preds1 = MORAN(image, length, text, text_rev, test=True)
            cost = criterion(torch.cat([preds0, preds1], 0), torch.cat([text, text_rev], 0))
            preds0_prob, preds0 = preds0.max(1)
            preds0 = preds0.view(-1)
            preds0_prob = preds0_prob.view(-1)
            sim_preds0 = converter.decode(preds0.data, length.data)
            preds1_prob, preds1 = preds1.max(1)
            preds1 = preds1.view(-1)
            preds1_prob = preds1_prob.view(-1)
            sim_preds1 = converter.decode(preds1.data, length.data)
            sim_preds = []
            for j in range(cpu_images.size(0)):
                text_begin = 0 if j == 0 else length.data[:j].sum()
                if torch.mean(preds0_prob[text_begin:text_begin+len(sim_preds0[j].split('$')[0]+'$')]).data[0] >\
                 torch.mean(preds1_prob[text_begin:text_begin+len(sim_preds1[j].split('$')[0]+'$')]).data[0]:
                    sim_preds.append(sim_preds0[j].split('$')[0]+'$')
                else:
                    sim_preds.append(sim_preds1[j].split('$')[0][-1::-1]+'$')

This code is to obtain the max probability, but it seems to torch.mean(softmax(preds0_prob)) should be calculated, not torch.mean(preds0_prob). So preds0.max(1) should be preds0.softmax(dim=1).max(1).

[Canjie-Luo]

Yes, you are right. Thanks.