env = DtRewardWrapper(env)
# Set seeds
seed(args.seed)
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
policy = DDPG(state_dim, action_dim, max_action, net_type="cnn")
replay_buffer = utils.ReplayBuffer(args.replay_buffer_max_size)
# Evaluate untrained policy
evaluations = [evaluate_policy(env, policy)]
exp.metric("rewards", evaluations[0])
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
episode_reward = None
env_counter = 0
while total_timesteps < args.max_timesteps:
if done:
if total_timesteps != 0:
print(("Total T: %d Episode Num: %d Episode T: %d Reward: %f") %
action_dim,
max_action,
net_type=args.net_type,
args=args)
if args.load_model:
policy.load("TD3_2_best_start_0_best",
"./pytorch_models/TD3_2_best_start_0")
print("load suceed!")
if not args.priority_replay:
replay_buffer = ReplayBuffer(args.replay_buffer_max_size)
else:
replay_buffer = PriReplayMemory(args, args.replay_buffer_max_size)
# Evaluate untrained policy
evaluation = np.mean([evaluate_policy(e, policy) for e in envs])
evaluations = [evaluation]
exp.metric("rewards", evaluations[0])
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
episode_reward = None
env_counter = 0
best_eval_rew = -np.float("Inf")
best_eval_index = 0
env = envs[random.randint(0, len(envs) - 1)]
while total_timesteps < args.max_timesteps: