eb110220d9a39a294479433cefc274e42506737e,main.py,,main,#,75

Before Change

            returns[step] = returns[step + 1] * \
                args.gamma * masks[step] + rewards[step]

        value_loss = (values[:-1] - Variable(returns[:-1])).pow().mean()

        advantages = returns[:-1] - values[:-1].data
        action_loss = -(Variable(advantages) * action_log_probs).mean()

After Change

        dist_entropy = -(log_probs * probs).sum(-1).mean()

        advantages = Variable(returns[:-1]) - values
        value_loss = advantages.pow(2).mean()

        action_loss = -(Variable(advantages.data) * action_log_probs).mean()

        optimizer.zero_grad()

Instances