def eval_epoch()

in ttw/train/predict_location_discrete.py [0:0]

• 34 lines of code
• 10 McCabe index (conditional complexity)

def eval_epoch(loader, tourist, guide, cuda, t_opt=None, g_opt=None):
    tourist.eval()
    guide.eval()

    correct, total = 0, 0
    for batch in loader:
        # forward
        t_out = tourist(batch)
        if cuda:
            t_out['comms'] = [x.cuda() for x in t_out['comms']]
        g_out = guide(t_out['comms'], batch)

        # acc
        correct += g_out['acc']*len(batch['target'])
        total += len(batch['target'])

        if t_opt and g_opt:
            # train if optimizers are specified
            rewards = -g_out['loss'].unsqueeze(-1)  # tourist reward is log likelihood of correct answer

            t_rl_loss = 0.
            eps = 1e-16

            advantage = Variable((rewards.data - t_out['baseline'].data))
            if cuda:
                advantage = advantage.cuda()
            t_val_loss = ((t_out['baseline'] - Variable(rewards.data)) ** 2).mean()  # mse

            for action, prob in zip(t_out['comms'], t_out['probs']):
                if cuda:
                    action = action.cuda()
                    prob = prob.cuda()

                action_prob = action * prob + (1.0 - action) * (1.0 - prob)

                t_rl_loss -= (torch.log(action_prob + eps) * advantage).sum()

            # backward
            g_opt.zero_grad()
            t_opt.zero_grad()
            g_out['loss'].sum().backward()
            (t_rl_loss + t_val_loss).backward()
            torch.nn.utils.clip_grad_norm(tourist.parameters(), 5)
            torch.nn.utils.clip_grad_norm(guide.parameters(), 5)
            g_opt.step()
            t_opt.step()

    return correct / total