def _do_model_step()

in compare_models.py [0:0]

• 32 lines of code
• 15 McCabe index (conditional complexity)

    def _do_model_step(self, scmap, race, inp, targ, game_name, optimize=True):
        ''' Returns:
            ([list of outputs for each model], [list of losses for each model])
            '''
        self.update_number += 1
        for model in self.models:
            if hasattr(model, "init_z"):
                model.init_z(game_name)
            model.hidden = model.init_hidden()

        losses = []
        outputs = []
        for mi, model in enumerate(self.models):
            if self.args.bptt > 0 and model.accepts_bptt:
                cur_losses = []
                cur_outputs = []
                for start, end in self.__generate_bptt_slices(inp.size(0)):
                    output, loss = self.__do_model_step(mi, scmap, race, inp[start:end], targ[start:end], game_name)
                    loss *= end-start if self.args.loss_averaging else 1
                    cur_losses.append(loss.data.cpu())
                    cur_outputs.append([x.data.cpu() for x in output])
                # Need to transpose last_outs and concat across time dim
                outputs.append([th.cat(x) for x in zip(*cur_outputs)])
                losses.append(th.cat(cur_losses).sum())
                if self.args.loss_averaging:
                    losses[mi] /= outputs[mi].size(0)
            elif self.args.bptt == 0 or not model.accepts_bptt:
                output, loss = self.__do_model_step(mi, scmap, race, inp, targ, game_name, optimize)
                outputs.append([x.data.cpu() for x in output])
                assert(loss.numel() == 1)
                losses.append(loss.data.cpu().sum())
            else:
                # TODO bptt sampled with cut ~ Bernouilli(hyperparam)
                self.quit_training("BPTT < 0 not implemented")
        return outputs, losses