def train(args):
# here is where we should start up a session,
# create DQN agent, create the model, etc.
# then we can run our fit method.
env = gym.make(args.env)
num_actions = env.action_space.n
network_model, q_values_func = create_model(
args.stack_frames, (args.cropped_size, args.cropped_size),
num_actions,
model_name='q_network')
preprocessor = PreprocessorSequence(args.stack_frames,
(args.cropped_size, args.cropped_size),
num_actions)
memory = ReplayMemory(args.memsize, args.stack_frames)
policy = {
'init': UniformRandomPolicy(num_actions),
'train': GreedyEpsilonPolicy(num_actions),
'test': GreedyPolicy(),
}
print("Generate Model...")
dqn_agent = DQNAgent(network_model, q_values_func, preprocessor, memory,
policy, args.gamma, args.target_update_freq,
args.num_burn_in, args.train_freq, args.batch_size,
args.output)
print("Compile Model...")
dqn_agent.compile(optimizer=Adam(lr=args.learning_rate),
loss_func=mean_huber_loss)
print("Fit Model...")
sys.stdout.flush()
dqn_agent.fit(env, args.num_iterations, args.max_episode_length)
def test(args):
if not os.path.isfile(args.model_path):
print("The model path: {} doesn't exist in the system.".format(
args.model_path))
print(
"Hints: python dqn_atari.py --mode test --model_path Path_to_your_model_weigths"
)
return
# here is where we should start up a session,
# create DQN agent, create our model, etc.
# then we can run our fit method.
env = gym.make(args.env)
num_actions = env.action_space.n
network_model, q_values_func = create_model(
args.stack_frames, (args.cropped_size, args.cropped_size),
num_actions,
model_name='q_network')
rewards = []
lens = []
tries = 0
while True:
env = gym.make(args.env)
env = wrappers.Monitor(env, 'videos', force=True)
#network_model.load_weights(args.output + '/model_weights_%d.h5' % m)
preprocessor = PreprocessorSequence(
args.stack_frames, (args.cropped_size, args.cropped_size),
num_actions)
memory = ReplayMemory(args.memsize, args.stack_frames)
policy = {
'init': UniformRandomPolicy(num_actions),
'train': GreedyEpsilonPolicy(num_actions),
'test': GreedyPolicy(),
}
dqn_agent = DQNAgent(network_model, q_values_func, preprocessor,
memory, policy, args.gamma,
args.target_update_freq, args.num_burn_in,
args.train_freq, args.batch_size, args.output)
dqn_agent.load_weights(args.model_path)
cumulative_reward, std, average_episode_length = dqn_agent.evaluate(
env, 1, None)
tries += 1
# Sometime the model is not very stable.
if tries > 100 or cumulative_reward > 350:
break
print('average reward = %f, std = %f, average_epis_length = %d' %
(cumulative_reward, std, average_episode_length))
rewards.append(cumulative_reward)
lens.append(average_episode_length)
def train(args):
# need to start a new session
env = gym.make(args.env)
num_actions = env.action_space.n
if args.modeltype == 'dueling_dqn':
network_model, q_values_func = create_model(
args.stack_frames, (args.cropped_size, args.cropped_size),
num_actions,
args.dueling_type,
model_name='dueling_q_network')
else:
network_model, q_values_func = create_model(
args.stack_frames, (args.cropped_size, args.cropped_size),
num_actions,
args.dueling_type,
model_name='q_network')
preprocessor = PreprocessorSequence(args.stack_frames,
(args.cropped_size, args.cropped_size),
num_actions)
memory = ReplayMemory(args.memsize, args.stack_frames)
policy = {
'init': UniformRandomPolicy(num_actions),
'train': GreedyEpsilonPolicy(num_actions),
'test': GreedyPolicy(),
}
print("Generate Model...")
if args.modeltype == 'double_dqn':
ddqn_agent = DQNAgent(network_model, q_values_func, preprocessor,
memory, policy, args.gamma,
args.target_update_freq, args.num_burn_in,
args.train_freq, args.batch_size, args.output,
True, False, args.dueling_type)
dqn_agent = DQNAgent(network_model, q_values_func, preprocessor,
memory, policy, args.gamma,
args.target_update_freq, args.num_burn_in,
args.train_freq, args.batch_size, args.output,
False, False, args.dueling_type)
elif args.modeltype == 'dueling_dqn':
dqn_agent = DQNAgent(network_model, q_values_func, preprocessor,
memory, policy, args.gamma,
args.target_update_freq, args.num_burn_in,
args.train_freq, args.batch_size, args.output,
False, True, args.dueling_type)
else: # natual dqn
dqn_agent = DQNAgent(network_model, q_values_func, preprocessor,
memory, policy, args.gamma,
args.target_update_freq, args.num_burn_in,
args.train_freq, args.batch_size, args.output,
False, False, args.dueling_type)
print("Compiling Model...")
dqn_agent.compile(optimizer=Adam(lr=args.learning_rate),
loss_func=mean_huber_loss)
if args.modeltype == 'double_dqn':
ddqn_agent.compile(optimizer=Adam(lr=args.learning_rate),
loss_func=mean_huber_loss)
print("Fitting model...")
sys.stdout.flush()
dqn_rewards = dqn_agent.fit(env, args.num_iterations,
args.max_episode_length)
if args.modeltype == 'double_dqn':
ddqn_rewards = ddqn_agent.fit(env, args.num_iterations,
args.max_episode_length)
return ddqn_rewards
return dqn_rewards