# INIT ENV
env, observation_space, action_space = build_env_wrapper(
params["env_name"], env_type=params["env_type"])
# LOGGING
writer = SummaryWriter(comment="-" + params["run_name"] + "-basic")
# NETWORK
net = dqn_mlp_net.Network(observation_space,
action_space,
hidden_layer_size=64).to(device)
tgt_net = agents.TargetNetwork(net)
# AGENT
selector = actions.EpsilonGreedyActionSelector(
epsilon=params["epsilon_start"])
epsilon_tracker = logger.EpsilonTracker(selector, params)
agent = agents.DQNAgent(net, selector, device=device)
# RUNNER
exp_source = runner.RunnerSourceFirstLast(env,
agent,
gamma=params["gamma"],
steps_count=1)
buffer = ExperienceReplayBuffer(exp_source,
buffer_size=params["replay_size"])
optimizer = optim.Adam(net.parameters(), lr=params["learning_rate"])
frame_idx = 0
done = False
net = models.SimpleFFDQN(env.get_obs_len(), env.get_action_space_size())
# load the network
if RL_options['load_net'] is True:
with open(
os.path.join(RL_options['net_saved_path'],
RL_options['net_file']), "rb") as f:
checkpoint = torch.load(f)
net = models.SimpleFFDQN(env.get_obs_len(),
env.get_action_space_size())
net.load_state_dict(checkpoint['state_dict'])
# create buffer
net.to(torch.device("cuda")) # pass into gpu
selector = actions.EpsilonGreedyActionSelector(RL_options['epsilon_start'])
agent = agents.DQNAgent(net, selector)
# agent = agents.Supervised_DQNAgent(net, selector, sample_sheet, assistance_ratio=0.2)
exp_source = experience.ExperienceSourceFirstLast(
env,
agent,
RL_options['gamma'],
steps_count=RL_options['reward_steps'])
buffer = experience.ExperienceReplayBuffer(exp_source,
RL_options['replay_size'])
# create optimizer
optimizer = optim.Adam(net.parameters(), lr=RL_options['lr'])
# create net pre-processor
net_processor = common.netPreprocessor(net, agent.target_model)