opt.train_sample_n is 16?

Hi @ruotianluo. May i ask why you are using train_sample_n = 16 for gen_result and n=1 for greedy in the code below? In transformer_sc.yml, you dont define train_sample_n so it should follow the value in opts which is 16. Shouldn’t this be >1 only when using new self-critical when we want to generate multiple samples?

          self.model.eval()
            with torch.no_grad():
                greedy_res, _ = self.model(fc_feats, att_feats, att_masks,
                    mode='sample',
                    opt={'sample_method': opt.sc_sample_method,
                         'beam_size': opt.sc_beam_size})
            self.model.train()
            gen_result, sample_logprobs = self.model(fc_feats, att_feats, att_masks,
                    opt={'sample_method':opt.train_sample_method,
                        'beam_size':opt.train_beam_size,
                        'sample_n': opt.train_sample_n},
                    mode='sample')

[ruotianluo]

No. It’s because some other loss may need the full probability.

[homelifes]

@ruotianluo alright. Thanks a lot. I have one more question. I realized that in gathering the logprobs:

           it, sampleLogprobs = self.sample_next_word(logprobs, sample_method, temperature)

            # stop when all finished
            if t == 0:
                unfinished = it > 0
            else:
                unfinished = unfinished * (it > 0)
            it = it * unfinished.type_as(it)
            seq[:,t] = it
            seqLogprobs[:,t] = logprobs

You are saving the logprobs and not the sampled ones sampleLogprobs. And then in RewardCriterion you gather them according to it:
input = input.gather(2, seq.unsqueeze(2)).squeeze(2)

May i also know if that contributes to efficient computation performance? Or is it just another way?