def run(weights_filename, nb_episodes):
# Create environment and model.
environment_name = weights_filename.split("_")[2]
environment = create_environment(environment_name)
observation_shape = environment.observation_space.shape
nb_actions = environment.action_space.n
model = build_model(observation_shape, nb_actions)
# Create memory.
memory = SequentialMemory(limit=1000000, window_length=1) # TODO Why is this necessary?
# Create the processor.
#processor = CarRacingProcessor()
# Create the DQN-Agent.
dqn = DQNAgent(
model=model,
nb_actions=nb_actions,
memory=memory,
enable_dueling_network=True, dueling_type='avg'
#processor=processor,
)
dqn.target_model = dqn.model # TODO Why is this necessary?
dqn.compile(optimizers.Adam(lr=.00025), metrics=['mae']) # TODO Why is this necessary?
# Load the weights.
dqn.load_weights(weights_filename)
print("Testing:", weights_filename)
# Test the agent.
dqn.test(environment, nb_episodes=nb_episodes, visualize=True)
def main():
# Process weights filename.
if len(sys.argv) != 2:
print("Must provide weights file-name.")
exit(0)
weights_filename = sys.argv[1]
# Get the environment and extract the number of actions.
environment_name = weights_filename.split("_")[
1] #"Duckietown-straight_road-v0"
environment = gym.make(environment_name)
environment = CarRacingDiscreteWrapper(environment)
np.random.seed(666)
nb_actions = environment.action_space.n
# Build the model.
model = build_model((WINDOW_LENGTH, ) + INPUT_SHAPE, nb_actions)
print(model.summary())
# Create memory.
memory = SequentialMemory(
limit=1000000,
window_length=WINDOW_LENGTH) # TODO Why is this necessary?
# Create the processor.
processor = CarRacingProcessor()
# Create the DQN-Agent.
dqn = DQNAgent(
model=model,
nb_actions=nb_actions,
memory=memory,
processor=processor,
)
dqn.target_model = dqn.model # TODO Why is this necessary?
dqn.compile(optimizers.Adam(lr=.00025),
metrics=['mae']) # TODO Why is this necessary?
# Load the weights.
dqn.load_weights(weights_filename)
# Test the agent.
dqn.test(environment, nb_episodes=10, visualize=True)
def main():
# Only works when you provide a file that contains trained weights.
if len(sys.argv) != 2:
print("Must provide weights file-name.")
exit(0)
# Get the environment and extract the number of actions.
environment_name = "lawnmower-medium-obstacles-v0"
environment = gym.make(environment_name)
environment.print_description()
nb_actions = environment.action_space.n
# Build the model.
model = build_model((WINDOW_LENGTH, ) + INPUT_SHAPE, nb_actions)
print(model.summary())
# Create memory.
memory = SequentialMemory(
limit=1000000,
window_length=WINDOW_LENGTH) # TODO Why is this necessary?
# Create the processor.
processor = LawnmowerProcessor()
# Create the DQN-Agent.
dqn = DQNAgent(
model=model,
nb_actions=nb_actions,
memory=memory,
processor=processor,
)
dqn.target_model = dqn.model # TODO Why is this necessary?
dqn.compile(optimizers.Adam(lr=.00025),
metrics=['mae']) # TODO Why is this necessary?
# Load the weights.
weights_filename = sys.argv[1]
dqn.load_weights(weights_filename)
# Test the agent.
dqn.test(environment, nb_episodes=10, visualize=True)