wrappers.py 文件源码

def __init__(self, env, cmdl):
        super(EvaluationMonitor, self).__init__(env)

        self.freq = cmdl.eval_frequency  # in steps
        self.eval_steps = cmdl.eval_steps
        self.cmdl = cmdl

        if self.cmdl.display_plots:
            import Visdom
            self.vis = Visdom()
            self.plot = self.vis.line(
                Y=np.array([0]), X=np.array([0]),
                opts=dict(
                    title=cmdl.label,
                    caption="Episodic reward per %d steps." % self.eval_steps)
            )

        self.eval_cnt = 0
        self.crt_training_step = 0
        self.step_cnt = 0
        self.ep_cnt = 1
        self.total_rw = 0
        self.max_mean_rw = -1000

        no_of_evals = cmdl.training_steps // cmdl.eval_frequency \
            - (cmdl.eval_start-1) // cmdl.eval_frequency

        self.eval_frame_idx = torch.LongTensor(no_of_evals).fill_(0)
        self.eval_rw_per_episode = torch.FloatTensor(no_of_evals).fill_(0)
        self.eval_rw_per_frame = torch.FloatTensor(no_of_evals).fill_(0)
        self.eval_eps_per_eval = torch.LongTensor(no_of_evals).fill_(0)