RESTORE_PATH = None
if RESTORE_PATH is None:
raise UserWarning(
'Please change the variable `RESTORE_PATH` to where you would like to load the model from. If you haven\'t trained a model, try \'example-save\''
)
# Create environment
game_wrapper = GameWrapper(ENV_NAME, MAX_NOOP_STEPS)
print("The environment has the following {} actions: {}".format(
game_wrapper.env.action_space.n,
game_wrapper.env.unwrapped.get_action_meanings()))
# Create agent
MAIN_DQN = build_q_network(game_wrapper.env.action_space.n,
LEARNING_RATE,
input_shape=INPUT_SHAPE)
TARGET_DQN = build_q_network(game_wrapper.env.action_space.n,
input_shape=INPUT_SHAPE)
replay_buffer = ReplayBuffer(size=MEM_SIZE, input_shape=INPUT_SHAPE)
agent = Agent(MAIN_DQN,
TARGET_DQN,
replay_buffer,
game_wrapper.env.action_space.n,
input_shape=INPUT_SHAPE,
batch_size=BATCH_SIZE)
print('Loading model...')
agent.load(RESTORE_PATH)
print('Loaded')
action_codes = []
inv_action_codes = []
saved_experts = find_saved_experts(env_names)
print(saved_experts)
experts = []
replay_buffers = []
for i in range(num_envs):
game_actions = game_wrappers[i].env.unwrapped.get_action_meanings()
action_code = [all_actions.index(action) for action in game_actions]
action_codes.append(action_code)
inv_action_code = [
game_actions.index(action) for action in all_actions
if action in game_actions
]
inv_action_codes.append(inv_action_code)
MAIN_DQN = build_q_network(len(game_actions), input_shape=INPUT_SHAPE)
TARGET_DQN = build_q_network(len(game_actions),
input_shape=INPUT_SHAPE)
replay_buffer = ReplayBuffer(size=MEM_SIZE,
input_shape=INPUT_SHAPE,
use_per=USE_PER)
expert = Agent(MAIN_DQN,
TARGET_DQN,
replay_buffer,
len(game_actions),
input_shape=INPUT_SHAPE,
batch_size=BATCH_SIZE,
use_per=USE_PER)
expert.load(saved_experts[i], False)
experts.append(expert)
replay_buffers.append(replay_buffer)