# Create environment
game_wrapper = GameWrapper(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(LEARNING_RATE, input_shape=INPUT_SHAPE)
TARGET_DQN = build_q_network(input_shape=INPUT_SHAPE)
replay_buffer = ReplayBuffer(size=MEM_SIZE, input_shape=INPUT_SHAPE)
agent = Agent(MAIN_DQN, TARGET_DQN, replay_buffer, input_shape=INPUT_SHAPE)
print('Loading model...')
# We only want to load the replay buffer when resuming training
agent.load('./saved_models/save-02502048/', load_replay_buffer=False)
print('Loaded.')
terminal = True
eval_rewards = []
evaluate_frame_number = 0
for frame in range(EVAL_LENGTH):
if terminal:
game_wrapper.reset(evaluation=True)
life_lost = True
episode_reward_sum = 0
terminal = False
# Breakout require a "fire" action (action #1) to start the
# game each time a life is lost.
# Build main and target networks
MAIN_DQN = build_q_network(LEARNING_RATE, input_shape=INPUT_SHAPE)
TARGET_DQN = build_q_network(input_shape=INPUT_SHAPE)
replay_buffer = ReplayBuffer(size=MEM_SIZE, input_shape=INPUT_SHAPE)
agent = Agent(MAIN_DQN,
TARGET_DQN,
replay_buffer,
input_shape=INPUT_SHAPE,
batch_size=BATCH_SIZE)
else:
# TODO: LOADING IS A LITTLE BROKEN AT THE MOMENTS!
# Load the agent instead
print('Loading from', LOAD_FROM)
meta = agent.load(LOAD_FROM, LOAD_REPLAY_BUFFER)
# Apply information loaded from meta
frame_number = meta['frame_number']
rewards = meta['rewards']
loss_list = meta['loss_list']
print('Loaded')
# FULL TRAINING LOOP
try:
# Allows us to write to Tensorboard
with writer.as_default():
while frame_number < TOTAL_FRAMES:
epoch_frame = 0