def main():
logger.configure()
env = make_atari('PongNoFrameskip-v4')
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=True,
lr=1e-4,
total_timesteps=int(1e7),
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
)
model.save('pong_model.pkl')
env.close()
def main(name, size):
model_location = ("logs/"+name+"NoFrameskip-v4_"+str(size)+"/model.pkl")
env_name = name+"NoFrameskip-v4"
#env = gym.make(name+"NoFrameskip-v4")
env = make_atari(env_name)
env = deepq.wrap_atari_dqn(env)
act = deepq.load(model_location)
episodeRewards = []
for i in range(100):
obs, done = env.reset(), False
episode_rew = 0
while not done:
# env.render()
action = act(obs[None])[0]
obs, rew, done, _ = env.step(action)
episode_rew += rew
print(episode_rew)
episodeRewards.append(episode_rew)
output = name+","+str(size)
for r in episodeRewards:
output+=","+str(r)
output+="\n"
with open("validationStats.csv", 'a') as myfile:
myfile.write(output)
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
act = deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized)
)
# act.save("pong_model.pkl") XXX
env.close()
def test_deepq():
"""
test DeepQ on atari
"""
clear_tf_session()
logger.configure()
set_global_seeds(SEED)
env = make_atari(ENV_ID)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=True)
deepq.learn(env,
q_func=model,
learning_rate=1e-4,
max_timesteps=NUM_TIMESTEPS,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
checkpoint_freq=10000)
env.close()
def main():
exp_dir = './runs/pong'
# by default CSV logs will be created in OS temp directory
logger.configure(dir=exp_dir,
format_strs=['stdout','log','csv','tensorboard'], log_suffix=None)
# create Atari environment, use no-op reset, max pool last two frames
env = make_atari('PongNoFrameskip-v4')
# by default monitor will log episod reward and log
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
learn_params = defaults.atari()
learn_params['checkpoint_path'] = exp_dir
learn_params['checkpoint_freq'] = 100000
learn_params['print_freq'] = 10
model = deepq.learn(
env,
# below are defaults
#convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
#hiddens=[256],
total_timesteps=int(1e7),
**learn_params
)
model.save('pong_model.pkl')
env.close()
def main():
logger.configure(dir=game + "_train_log")
env = make_atari(name + 'NoFrameskip-v4')
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[512],
dueling=False,
lr=0.00025,
total_timesteps=int(5e7),
buffer_size=1000000,
exploration_fraction=0.02,
exploration_final_eps=0.1,
train_freq=4,
learning_starts=50000,
target_network_update_freq=10000,
gamma=0.99,
print_freq=1000,
checkpoint_path=game + "_checkpoints")
model.save(game + '_model.pkl')
env.close()
def main():
logger.configure()
env = make_atari('PongNoFrameskip-v4')
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=True,
lr=1e-4,
total_timesteps=int(1e7),
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
)
model.save('pong_model.pkl')
env.close()
def get_player(self, train=False):
if self.env:
return env
if self.config['ENV_TYPE'] == 'Classic':
env = gym.make(self.config['ENV_NAME'])
elif self.config['ENV_TYPE'] == 'Atari':
if train:
env = make_atari(self.config['ENV_NAME'])
env = bench.Monitor(env, self.logger.get_dir())
env = deepq.wrap_atari_dqn(env)
else:
env = gym.make(self.config['ENV_NAME'])
env = deepq.wrap_atari_dqn(env)
else:
raise Exception('Environment Type %s - Not Supported' % self.config['ENV_TYPE'])
return env
def make_env(game_name):
env = gym.make(game_name + "NoFrameskip-v4")
monitored_env = bench.Monitor(
env, logger.get_dir()
) # puts rewards and number of steps in info, before environment is wrapped
env = deepq.wrap_atari_dqn(
monitored_env
) # applies a bunch of modification to simplify the observation space (downsample, make b/w)
return env, monitored_env
def train():
logger.configure()
set_global_seeds(args.seed)
directory = os.path.join(args.log_dir, '_'.join([args.env, datetime.datetime.now().strftime("%m%d%H%M")]))
if not os.path.exists(directory):
os.makedirs(directory)
else:
ValueError("The directory already exists...", directory)
json.dump(vars(args), open(os.path.join(directory, 'learning_prop.json'), 'w'))
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
nb_test_steps = args.nb_test_steps if args.nb_test_steps > 0 else None
if args.record == 1:
env = Monitor(env, directory=args.log_dir)
with tf.device(args.device):
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
act, records = deepq.learn(
env,
q_func=model,
lr=args.learning_rate,
max_timesteps=args.nb_train_steps,
buffer_size=args.buffer_size,
exploration_fraction=args.eps_fraction,
exploration_final_eps=args.eps_min,
train_freq=4,
print_freq=1000,
checkpoint_freq=int(args.nb_train_steps/10),
learning_starts=args.nb_warmup_steps,
target_network_update_freq=args.target_update_freq,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
epoch_steps = args.nb_epoch_steps,
gpu_memory = args.gpu_memory,
double_q = args.double_q,
directory=directory,
nb_test_steps = nb_test_steps,
scope = args.scope,
)
print("Saving model to model.pkl")
act.save(os.path.join(directory,"model.pkl"))
env.close()
plot(records, directory)
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--checkpoint-freq', type=int, default=10000)
# parser.add_argument('--checkpoint-path', type=str, default=None)
parser.add_argument(
'--checkpoint-path',
type=str,
default=
'/home/yangxu/PycharmProjects/ros_inwork/baselines/deepq/experiments/save'
)
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
# 这里使用了3个卷积层,两个全连接层（第一个全连接层为num_convs*hiddens,第二层为hiddens*num_action）
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
checkpoint_freq=args.checkpoint_freq,
checkpoint_path=args.checkpoint_path,
)
env.close()
def main():
env = gym.make("PongNoFrameskip-v4")
env = deepq.wrap_atari_dqn(env)
act = deepq.load("pong_model.pkl")
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='Breakout')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('experiment_id')
args = parser.parse_args()
logging_directory = Path('./experiments/{}--{}'.format(args.experiment_id, args.env))
if not logging_directory.exists():
logging_directory.mkdir(parents=True)
logger.configure(str(logging_directory), ['stdout', 'tensorboard', 'json'])
model_directory = logging_directory / 'models'
if not model_directory.exists():
model_directory.mkdir(parents=True)
set_global_seeds(args.seed)
env_name = args.env + "NoFrameskip-v4"
env = make_atari(env_name)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
)
exploration_schedule = PiecewiseSchedule(
endpoints=[(0, 1), (1e6, 0.1), (5 * 1e6, 0.01)], outside_value=0.01)
act = learn(
env,
q_func=model,
beta1=0.9,
beta2=0.99,
epsilon=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=1000000,
exploration_schedule=exploration_schedule,
start_lr=1e-4,
end_lr=5 * 1e-5,
start_step=1e6,
end_step=5 * 1e6,
train_freq=4,
print_freq=10,
batch_size=32,
learning_starts=50000,
target_network_update_freq=10000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
model_directory=model_directory
)
act.save(str(model_directory / "act_model.pkl"))
env.close()
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("--game_latency", type=int, default=0,help="Input-Latency used in game run.")
parser.add_argument("--trained_with_latency", type=int, default=0, help="Input-Latency the model was trained with.")
parser.add_argument("--trained_with_all_latency_mode", type=int, default=0, help="Input-Latency-All-Mode the model was trained with.")
parser.add_argument("--game_runs", type=int, default=25, help="How offen the game should be ran.")
args = parser.parse_args()
prepath = ""
if args.trained_with_all_latency_mode == 0:
modelfile = "data.L" + str(args.trained_with_latency) + "/PongNoFrameskip-v4.L" + str(args.trained_with_latency) + ".pkl"
csvfile = "data.L" + str(args.trained_with_latency)+ "/PongNoFrameskip-v4.L" + str(args.trained_with_latency) + ".pkl.on-L" + str(args.game_latency) + ".csv"
else:
modelfile = "data.LM" + str(args.trained_with_all_latency_mode)+ "/PongNoFrameskip-v4.LM" + str(args.trained_with_all_latency_mode) + ".pkl"
csvfile = "data.LM" + str(args.trained_with_all_latency_mode) + "/PongNoFrameskip-v4.LM" + str(args.trained_with_all_latency_mode) + ".pkl.on-L" + str(args.game_latency) + ".csv"
modelfile = prepath + modelfile
csvfile = prepath + csvfile
env = make_atari("PongNoFrameskip-v4")
env = deepq.wrap_atari_dqn(env)
act = deepq.load(modelfile)
with open(csvfile, 'w', newline='') as f:
writer = csv.writer(f)
writer.writerows([["Reward", "Time", "Frames"]])
i = 0
while i < args.game_runs:
i+=1
obs, done = env.reset(), False
timeslice_start = datetime.datetime.now()
frames = 0
q_input = deque([0] * args.game_latency)
episode_rew = 0
while not done:
# uncomment to analyse what the game is doing by your eye
# env.render()
# time.sleep(0.01)
q_input.append(act(obs[None])[0])
obs, rew, done, _ = env.step(q_input.popleft())
frames+=1
episode_rew += rew
timespent = (datetime.datetime.now() - timeslice_start)
writer.writerows([[episode_rew, timespent.total_seconds(), frames]])
f.flush()
print(str(episode_rew) + ", " + str(timespent.total_seconds()) + ", " + str(frames))
def main():
"""
run the atari test
"""
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--checkpoint-freq', type=int, default=10000)
parser.add_argument('--checkpoint-path', type=str, default=None)
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
with tf.Session():
deepq.learn(
env,
q_func=model,
learning_rate=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
checkpoint_freq=args.checkpoint_freq,
checkpoint_path=args.checkpoint_path,
)
env.close()
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--train-with-latency', type=int, default=0)
parser.add_argument('--train-with-all-latency-mode', type=int, default=0)
args = parser.parse_args()
loggerid = "L" + (("M" + str(args.train_with_all_latency_mode)) if
(args.train_with_all_latency_mode != 0) else str(
args.train_with_latency))
loggerdir = "./data." + loggerid + "/"
logger.configure(dir=loggerdir)
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
act = deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
print_freq=1,
train_with_latency=args.train_with_latency,
train_with_all_latency_mode=args.train_with_all_latency_mode)
act.save(loggerdir + args.env + "." + loggerid + ".pkl")
env.close()
def main():
env = gym.make("PongNoFrameskip-v4")
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=True,
total_timesteps=0)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(model(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
def test():
env = make_atari(args.env)
env = deepq.wrap_atari_dqn(env)
act = deepq.load(os.path.join(args.log_dir, args.log_fname))
if args.record:
env = Monitor(env, directory=args.log_dir)
while True:
obs, done = env.reset(), False
episode_rew = 0
t = 0
while not done:
if not(args.record):
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
t += 1
print("Episode reward %.2f after %d steps"%(episode_rew, t))
def main(envName='BreakoutNoFrameskip-v4', bufferSize=10000, timesteps=3e6):
# parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
# parser.add_argument('--buffer', type=int, default=10000)
# parser.add_argument('--seed', help='RNG seed', type=int, default=0)
# parser.add_argument('--prioritized', type=int, default=1)
# parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
# parser.add_argument('--dueling', type=int, default=1)
# parser.add_argument('--num-timesteps', type=int, default=int(10e6))
# parser.add_argument('--checkpoint-freq', type=int, default=10000)
# parser.add_argument('--checkpoint-path', type=str, default=os.getcwd()+"/logs")
# args = parser.parse_args()
# logger.configure(dir=args.checkpoint_path)
logger.configure(dir=os.getcwd() + "/logs/" + str(envName) + "_" +
str(bufferSize))
set_global_seeds(0)
env = make_atari(envName)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[512],
dueling=bool(1),
)
act = deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=int(timesteps),
buffer_size=bufferSize,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(1),
prioritized_replay_alpha=0.6,
checkpoint_freq=10000,
)
act.save(os.getcwd() + "/logs/" + str(envName) + "_" + str(bufferSize) +
"/model.pkl")
env.close()
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
'--env', help='environment ID', default='BeamRiderNoFrameskip-v4'
) # TODO changed to Beamrider since it gives larger rewards easy to see progress
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int,
default=0) # TODO made it false code was complaining
parser.add_argument('--dueling', type=int,
default=0) # TODO made it false for code simplicity
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
args = parser.parse_args()
logger.configure("./log/BeamRider") # TODO log results under BeamRider
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
nbins=1000, # TODO number of bins
)
act = deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
min_Val=-50, # TODO min value of Q values
max_Val=50, # TODO max value of Q values
nbins=1000 # TODO number of bins
)
# act.save("pong_model.pkl") XXX
env.close()
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--checkpoint-freq', type=int, default=10000)
parser.add_argument('--checkpoint-path', type=str, default=None)
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
checkpoint_freq=args.checkpoint_freq,
checkpoint_path=args.checkpoint_path,
)
env.close()
def main():
env = gym.make("PongNoFrameskip-v4")
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=True,
total_timesteps=0
)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(model(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
def test(env0, act_greedy, nb_itrs=5, nb_test_steps=10000):
env = env0
while (hasattr(env, 'env')):
env = env.env
total_rewards = []
for _ in range(nb_itrs):
if hasattr(env, 'ale'):
from baselines.common.atari_wrappers import make_atari
env_new = make_atari(env.spec.id)
env_new = deepq.wrap_atari_dqn(env_new)
else:
env_new = gym.make(env.spec.id)
obs = env_new.reset()
if nb_test_steps is None:
done = False
episode_reward = 0
while not done:
action = act_greedy(np.array(obs)[None])[0]
obs, rew, done, _ = env_new.step(action)
episode_reward += rew
total_rewards.append(episode_reward)
else:
t = 0
episodes = []
episode_reward = 0
while (t < nb_test_steps):
action = act_greedy(np.array(obs)[None])[0]
obs, rew, done, _ = env_new.step(action)
episode_reward += rew
if done:
episodes.append(episode_reward)
episode_reward = 0
obs = env_new.reset()
t += 1
total_rewards.append(np.mean(episodes))
return np.array(total_rewards, dtype=np.float32)
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID',
default='SeaquestNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
subparsers = parser.add_subparsers(dest='subparser',
help='train model or test existing model')
train_parser = subparsers.add_parser('train')
# Mnih et al (2015) and other DeepMind work usually train for 200e6 frames,
# which is 50e6 time steps with 4 frameskip (introduced by wrapper in
# make_atari.)
train_parser.add_argument('--num-timesteps', type=int, default=int(50e6))
train_parser.add_argument('--out-dir', type=str, default=None,
help='checkpoint directory')
test_parser = subparsers.add_parser('test')
test_parser.add_argument('--fps', type=int, default=int(1e3))
test_parser.add_argument('model', help='model file', type=str)
args = parser.parse_args()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
if args.subparser == 'train':
train(env, args)
elif args.subparser == 'test':
test(env, args)
else:
assert False
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--dir', help='model directory', default='')
args = parser.parse_args()
print("Args env: " + args.env)
env = gym.make(args.env)
env = deepq.wrap_atari_dqn(env)
act = deepq.load(args.dir)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
def make_env(game_name):
env = gym.make(game_name + "NoFrameskip-v4")
env = bench.Monitor(env, None)
env = SimpleMonitor(env)
env = wrap_atari_dqn(env)
return env
def main():
logger.configure(dir="breakout_train_log")
env = make_atari('BreakoutNoFrameskip-v4')
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=False,
total_timesteps=0,
load_path="checkpoints/model.pkl",
)
sess = get_session()
pytorch_network = Net()
variables_names = [v.name for v in tf.trainable_variables()]
values = sess.run(variables_names)
for k, v in zip(variables_names, values):
if (k == 'deepq/q_func/convnet/Conv/weights:0'):
weight1 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv/biases:0'):
bias1 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_1/weights:0'):
weight2 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_1/biases:0'):
bias2 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_2/weights:0'):
weight3 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_2/biases:0'):
bias3 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected/weights:0'):
weight_fc1 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected/biases:0'):
bias_fc1 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected_1/weights:0'):
weight_fc2 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected_1/biases:0'):
bias_fc2 = torch.from_numpy(v)
pytorch_network.conv1.weight = torch.nn.Parameter(weight1.permute(3, 2, 0, 1))
pytorch_network.conv1.bias = torch.nn.Parameter(bias1)
# pytorch_network.conv1.bias = torch.nn.Parameter(np.zeros_like(bias1))
pytorch_network.conv2.weight = torch.nn.Parameter(weight2.permute(3, 2, 0, 1))
pytorch_network.conv2.bias = torch.nn.Parameter(bias2)
pytorch_network.conv3.weight = torch.nn.Parameter(weight3.permute(3, 2, 0, 1))
pytorch_network.conv3.bias = torch.nn.Parameter(bias3)
pytorch_network.fc1.weight = torch.nn.Parameter(weight_fc1.permute(1, 0))
pytorch_network.fc1.bias = torch.nn.Parameter(bias_fc1)
pytorch_network.fc2.weight = torch.nn.Parameter(weight_fc2.permute(1, 0))
pytorch_network.fc2.bias = torch.nn.Parameter(bias_fc2)
torch.save(pytorch_network.state_dict(), 'pytorch_breakout_dqn.pt')
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
# print(torch.tensor(obs[None]).dtype)
# print(pytorch_network(torch.tensor(obs[None], dtype=torch.float).permute(0, 3, 1, 2)))
# print(model(obs[None]))
# print(tf.global_variables())
action = torch.argmax(pytorch_network(torch.tensor(obs[None], dtype=torch.float).permute(0, 3, 1, 2)))
# print(action)
obs, rew, done, _ = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
# model.save('breakout_model.pkl')
env.close()
def make_env(game_name):
env = gym.make(game_name + "NoFrameskip-v4")
env_monitored = bench.Monitor(env, None)
env = deepq.wrap_atari_dqn(env_monitored)
return env_monitored, env
def main():
game = 'qbert'
env = game.capitalize() + "NoFrameskip-v4"
# env = game + "NoFrameskip-v4"
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default=env)
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=0)
parser.add_argument('--dueling', type=int, default=0)
parser.add_argument('--expected-dueling', type=int, default=0)
parser.add_argument('--double', type=int, default=1)
parser.add_argument('--alpha', type=int, default=1) # Enables alpha-dqn
parser.add_argument('--sample', type=int, default=0)
parser.add_argument('--expected', type=int,
default=1) # Turns on Expected-alpha-dqn
parser.add_argument('--surrogate', type=int,
default=0) # Turns on Surrogate-alpha-dqn
parser.add_argument('--steps', type=int, default=1)
parser.add_argument('--eps-val', type=float, default=0.01)
parser.add_argument('--alpha-val', type=float, default=0.01)
parser.add_argument('--piecewise-schedule', type=int, default=0)
parser.add_argument('--alpha-epsilon', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(50e6))
parser.add_argument('--exploration-fraction', type=int, default=int(1e6))
parser.add_argument('--optimal-test', type=int, default=0)
args = parser.parse_args()
if args.alpha:
assert (bool(args.expected) != bool(args.surrogate)
) # both versions can't run together
if args.surrogate:
surrogate = True
else:
surrogate = False
typename = ''
if args.double: typename += 'D'
typename += "DQN"
typename += '_EPS_{}'.format(args.eps_val)
if args.alpha:
if args.eps_val != args.alpha_val:
typename += '_ALPHA_{}'.format(args.alpha_val)
if surrogate:
typename += '_Surrogate'
else:
typename += '_Expected'
if args.sample:
typename += '_Sample'
if args.piecewise_schedule:
typename += '_Piecewise'
if args.alpha_epsilon:
typename += '_ALPHA=EPS'
typename += '_{}_step'.format(args.steps)
if args.prioritized: typename += '_PRIORITY'
if args.dueling: typename += '_DUEL'
game = args.env[:-14].lower()
directory = 'AlphaGreedy/'
dir = osp.join(
'../experiments/' + directory + game,
datetime.datetime.now().strftime(typename + "-%d-%m-%H-%M-%f"))
logger.configure(dir=dir)
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir(), allow_early_resets=True)
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
act = deepq.learn(
env,
q_func=model,
lr=1e-4,
max_timesteps=args.num_timesteps,
buffer_size=10000,
exploration_fraction=(float(args.exploration_fraction) /
float(args.num_timesteps)),
exploration_final_eps=args.eps_val,
alpha_val=args.alpha_val,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
double=bool(args.double),
epsilon=bool(args.alpha), # turns on alpha-dqn
eps_val=args.eps_val,
q1=surrogate, # q1 is surrogate, else, expected
n_steps=args.steps,
sample=bool(args.sample),
piecewise_schedule=bool(args.piecewise_schedule),
test_agent=1e6)
act.save()
env.close()
def main(game: str, model: str):
name = ''.join([g.capitalize() for g in game.split('_')])
env = make_atari(name + 'NoFrameskip-v4')
env = bench.Monitor(env, logger.get_dir(), allow_early_resets=True)
env = deepq.wrap_atari_dqn(env)
n_actions = env.action_space.n
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
# 1 input image channel, 6 output channels, 5x5 square convolution
# kernel
self.conv1 = nn.Conv2d(4, 32, 8, stride=4, padding=2)
self.relu1 = nn.ReLU()
self.conv2 = nn.Conv2d(32, 64, 4, stride=2, padding=0)
self.relu2 = nn.ReLU()
self.pad2 = nn.ConstantPad2d((1, 2, 1, 2), value=0)
self.conv3 = nn.Conv2d(64, 64, 3, stride=1, padding=0)
self.pad3 = nn.ConstantPad2d((1, 1, 1, 1), value=0)
self.relu3 = nn.ReLU()
# an affine operation: y = Wx + b
self.fc1 = nn.Linear(7744, 512)
self.relu4 = nn.ReLU()
self.fc2 = nn.Linear(512, n_actions)
def forward(self, x):
# Max pooling over a (2, 2) window
x = x / 255.
x = self.relu1(self.conv1(x))
# If the size is a square you can only specify a single number
x = self.pad2(x)
x = self.relu2(self.conv2(x))
x = self.pad3(x)
x = self.relu3(self.conv3(x))
x = x.permute(0, 2, 3, 1).contiguous()
x = x.view(-1, self.num_flat_features(x))
x = self.relu4(self.fc1(x))
x = self.fc2(x)
return x
def num_flat_features(self, x):
size = x.size()[1:] # all dimensions except the batch dimension
num_features = 1
for s in size:
num_features *= s
return num_features
# logger.configure(dir="breakout_train_log")
if model is None:
model = game + "_model.pkl"
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[512],
dueling=False,
total_timesteps=0,
exploration_final_eps=0.05,
load_path="trained_networks/" + model,
)
sess = get_session()
pytorch_network = Net()
variables_names = [v.name for v in tf.trainable_variables()]
values = sess.run(variables_names)
for k, v in zip(variables_names, values):
if (k == 'deepq/q_func/convnet/Conv/weights:0'):
weight1 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv/biases:0'):
bias1 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_1/weights:0'):
weight2 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_1/biases:0'):
bias2 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_2/weights:0'):
weight3 = torch.from_numpy(v)
if (k == 'deepq/q_func/convnet/Conv_2/biases:0'):
bias3 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected/weights:0'):
weight_fc1 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected/biases:0'):
bias_fc1 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected_1/weights:0'):
weight_fc2 = torch.from_numpy(v)
if (k == 'deepq/q_func/action_value/fully_connected_1/biases:0'):
bias_fc2 = torch.from_numpy(v)
pytorch_network.conv1.weight = torch.nn.Parameter(
weight1.permute(3, 2, 0, 1))
pytorch_network.conv1.bias = torch.nn.Parameter(bias1)
# pytorch_network.conv1.bias = torch.nn.Parameter(np.zeros_like(bias1))
pytorch_network.conv2.weight = torch.nn.Parameter(
weight2.permute(3, 2, 0, 1))
pytorch_network.conv2.bias = torch.nn.Parameter(bias2)
pytorch_network.conv3.weight = torch.nn.Parameter(
weight3.permute(3, 2, 0, 1))
pytorch_network.conv3.bias = torch.nn.Parameter(bias3)
pytorch_network.fc1.weight = torch.nn.Parameter(weight_fc1.permute(1, 0))
pytorch_network.fc1.bias = torch.nn.Parameter(bias_fc1)
pytorch_network.fc2.weight = torch.nn.Parameter(weight_fc2.permute(1, 0))
pytorch_network.fc2.bias = torch.nn.Parameter(bias_fc2)
torch.save(pytorch_network.state_dict(), 'pytorch_' + game + '.pt')
rewards = np.zeros(100)
for episode in range(100):
obs, done = env.reset(), False
episode_rew = 0
while not done:
# env.render()
# print(torch.tensor(obs[None]).dtype)
# print(pytorch_network(torch.tensor(obs[None], dtype=torch.float).permute(0, 3, 1, 2)))
# print(model(obs[None]))
# print(tf.global_variables())
probabilities = policy(
pytorch_network(
torch.tensor(obs[None],
dtype=torch.float).permute(0, 3, 1, 2))[0],
0.05)
action = np.random.choice(np.arange(len(probabilities)),
p=probabilities)
# action = model(obs[None])[0]
obs, rew, done, _ = env.step(action)
episode_rew += rew
rewards[episode] = episode_rew
print("Episode " + str(episode) + " reward", episode_rew)
# model.save('breakout_model.pkl')
env.close()
print("Avg: ", np.mean(rewards))
def main():
args = parse_args()
logdir = args.pop('logdir')
# logger.configure(dir=logdir, enable_std_out=True)
logger.configure(dir=logdir, enable_std_out=False)
with open(os.path.join(logger.get_dir(), "hyparam.yaml"), 'w') as f:
yaml.dump(args, f, default_flow_style=False)
policy_mode = args.pop('policy_mode')
save_array_flag = args.pop('save_array_flag')
use_my_env_wrapper = args.pop('use_my_env_wrapper')
env_id = args.pop('env_id')
if policy_mode == "large_variance":
if args["exploitation_ratio_on_bottleneck"] is None or args[
"bottleneck_threshold_ratio"] is None:
raise AssertionError
if args["exploitation_ratio_on_bottleneck"] is not None:
array_logger = array_logger_getter.get_logger()
array_logger.set_log_dir(logdir, exist_ok=True)
array_logger.set_save_array_flag(save_array_flag)
if use_my_env_wrapper:
env = make_atari_nature(env_id)
else:
env = make_atari(env_id)
env = deepq.wrap_atari_dqn(env)
num_cpu = 1
config = tf.ConfigProto(
allow_soft_placement=True,
inter_op_parallelism_threads=num_cpu,
intra_op_parallelism_threads=num_cpu,
gpu_options=tf.GPUOptions(visible_device_list=args.pop("gpu"),
allow_growth=True),
)
config.gpu_options.allow_growth = True
# nature_set = {'network': 'cnn', 'prioritized_replay': False, 'buffer_size': int(1e5), 'total_time_steps': int(2e6)}
model = deepq.learn(
env,
"conv_only",
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=False,
lr=1e-4,
# total_timesteps=int(1e7),
# total_timesteps=int(2e3)+1,
buffer_size=10000,
# exploration_fraction=0.1,
# exploration_final_eps=0.01,
train_freq=4,
# learning_starts=1000,
# target_network_update_freq=100,
learning_starts=1000,
# target_network_update_freq=1000,
target_network_update_freq=500,
gamma=0.99,
# prioritized_replay=False,
batch_size=64,
# print_freq=1,
# print_freq=200,
print_freq=1000,
config=config,
bottleneck_threshold_update_freq=1000,
**args,
)
model.save(os.path.join(logger.get_dir(), 'Breakout_final_model.pkl'))
env.close()
def main(seed=10086,
time=500,
percentile=99.9,
game='video_pinball',
model='pytorch_video_pinball.pt',
episode=50):
seed = seed
n_examples = 15000
time = time
epsilon = 0.05
percentile = percentile
episode = episode
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
device = 'cuda'
else:
device = 'cpu'
print("device", device)
print("game", game, "episode", episode, "time", time, "seed", seed,
"percentile", percentile)
set_seed(seed)
name = ''.join([g.capitalize() for g in game.split('_')])
env = make_atari(game, max_episode_steps=18000)
env = bench.Monitor(env, logger.get_dir(), allow_early_resets=True)
env = deepq.wrap_atari_dqn(env)
env.seed(seed)
n_actions = env.action_space.n
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
# 1 input image channel, 6 output channels, 5x5 square convolution
# kernel
self.conv1 = nn.Conv2d(4, 32, 8, stride=4, padding=2)
self.relu1 = nn.ReLU()
self.pad2 = nn.ConstantPad2d((1, 2, 1, 2), value=0)
self.conv2 = nn.Conv2d(32, 64, 4, stride=2, padding=0)
self.relu2 = nn.ReLU()
self.pad3 = nn.ConstantPad2d((1, 1, 1, 1), value=0)
self.conv3 = nn.Conv2d(64, 64, 3, stride=1, padding=0)
self.relu3 = nn.ReLU()
#self.perm = Permute((0, 2, 3, 1))
self.perm = Permute((1, 2, 0))
self.fc1 = nn.Linear(7744, 512)
self.relu4 = nn.ReLU()
self.fc2 = nn.Linear(512, n_actions)
def forward(self, x):
x = x / 255.0
x = self.relu1(self.conv1(x))
x = self.pad2(x)
x = self.relu2(self.conv2(x))
x = self.pad3(x)
x = self.relu3(self.conv3(x))
x = x.permute(0, 2, 3, 1).contiguous()
x = x.view(-1, self.num_flat_features(x))
x = self.relu4(self.fc1(x))
x = self.fc2(x)
return x
def show(self, x):
x = x
x = self.relu1(self.conv1(x))
x = self.pad2(x)
x = self.relu2(self.conv2(x))
x = self.pad3(x)
x = self.relu3(self.conv3(x))
x = x.permute(0, 2, 3, 1).contiguous()
x = x.view(-1, self.num_flat_features(x))
x = self.relu4(self.fc1(x))
x = self.fc2(x)
return torch.max(x, 1)[1].data[0]
def num_flat_features(self, x):
size = x.size()[1:] # all dimensions except the batch dimension
num_features = 1
for s in size:
num_features *= s
return num_features
model_path = model
ANN_model = Net()
ANN_model.load_state_dict(torch.load(model_path))
ANN_model.eval()
ANN_model = ANN_model.to(device)
images = []
cnt = 0
for epi in range(1000):
cur_images = []
cur_cnt = 0
obs, done = env.reset(), False
episode_rew = 0
while not done:
#env.render()
image = torch.from_numpy(obs[None]).permute(0, 3, 1, 2)
cur_images.append(image.detach().numpy())
actions_value = ANN_model(image.to(device)).cpu()[0]
probs, best_action = policy(actions_value, epsilon)
action = np.random.choice(np.arange(len(probs)), p=probs)
obs, rew, done, info = env.step(action)
cur_cnt += 1
if info['ale.lives'] == 0:
if cur_cnt + cnt < n_examples:
cnt += cur_cnt
images += cur_images
else:
print("normalization image cnt", cnt)
break
images = torch.from_numpy(np.array(images)).reshape(-1, 4, 84,
84).float() / 255
SNN = ann_to_snn(ANN_model,
input_shape=(4, 84, 84),
data=images.to(device),
percentile=percentile)
SNN = SNN.to(device)
for l in SNN.layers:
if l != 'Input':
SNN.add_monitor(Monitor(SNN.layers[l],
state_vars=['s', 'v'],
time=time),
name=l)
for c in SNN.connections:
if isinstance(SNN.connections[c], MaxPool2dConnection):
SNN.add_monitor(Monitor(SNN.connections[c],
state_vars=['firing_rates'],
time=time),
name=f'{c[0]}_{c[1]}_rates')
f = open(
"game" + game + "episode" + str(episode) + "time" + str(time) +
"percentile" + str(percentile) + ".csv", 'a')
game_cnt = 0
mix_cnt = 0
spike_cnt = 0
cnt = 0
rewards = np.zeros(episode)
while (game_cnt < episode):
obs, done = env.reset(), False
while not done:
image = torch.from_numpy(obs[None]).permute(0, 3, 1,
2).float() / 255
image = image.to(device)
ANN_action = ANN_model.show(image.to(device))
inpts = {'Input': image.repeat(time, 1, 1, 1, 1)}
SNN.run(inputs=inpts, time=time)
spikes = {
l: SNN.monitors[l].get('s')
for l in SNN.monitors if 's' in SNN.monitors[l].state_vars
}
voltages = {
l: SNN.monitors[l].get('v')
for l in SNN.monitors if 'v' in SNN.monitors[l].state_vars
}
actions_value = spikes['12'].sum(0).cpu() + voltages['12'][
time - 1].cpu()
action = torch.max(actions_value, 1)[1].data.numpy()[0]
spike_actions_value = spikes['12'].sum(0).cpu()
spike_action = torch.max(spike_actions_value, 1)[1].data.numpy()[0]
cnt += 1
if ANN_action == action:
mix_cnt += 1
if ANN_action == spike_action:
spike_cnt += 1
probs, best_action = policy(actions_value[0], epsilon)
action = np.random.choice(np.arange(len(probs)), p=probs)
SNN.reset_state_variables()
obs, rew, done, info = env.step(action)
if info['ale.lives'] == 0:
rewards[game_cnt] = info['episode']['r']
print("Episode " + str(game_cnt) + " reward", rewards[game_cnt])
print("cnt", cnt, "mix", mix_cnt / cnt, "spike", spike_cnt / cnt)
f.write(
str(rewards[game_cnt]) + ", " + str(mix_cnt / cnt) + ", " +
str(spike_cnt / cnt) + "\n")
game_cnt += 1
mix_cnt = 0
spike_cnt = 0
cnt = 0
elif 'TimeLimit.truncated' in info:
if info['TimeLimit.truncated'] == True:
rewards[game_cnt] = info['episode']['r']
print("Episode " + str(game_cnt) + " reward",
rewards[game_cnt])
print("cnt", cnt, "mix", mix_cnt / cnt, "spike",
spike_cnt / cnt)
f.write(
str(rewards[game_cnt]) + ", " + str(mix_cnt / cnt) + ", " +
str(spike_cnt / cnt) + "\n")
game_cnt += 1
mix_cnt = 0
spike_cnt = 0
cnt = 0
env.close()
f.close()
print("Avg: ", np.mean(rewards))
output_str = "Avg: " + str(np.mean(rewards))
def main():
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='PongNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--logdir', type=str, default='~/.tmp/deepq')
# General hyper-parameters
parser.add_argument('--isKfac', type=int, default=0)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--batch_size', type=int, default=160)
parser.add_argument('--target_network_update_freq', type=int, default=1000)
# Kfac parameters
parser.add_argument('--kfac_fisher_metric', type=str, default='gn')
parser.add_argument('--kfac_momentum', type=float, default=0.9)
parser.add_argument('--kfac_clip_kl', type=float, default=0.01)
parser.add_argument('--kfac_epsilon', type=float, default=1e-2)
parser.add_argument('--kfac_stats_decay', type=float, default=0.99)
parser.add_argument('--kfac_cold_iter', type=float, default=10)
args = parser.parse_args()
print('_'.join([str(arg) for arg in vars(args)]))
logdir = osp.join(args.logdir, '_'.join([str(getattr(args, arg)) for arg in vars(args) if arg != 'logdir']))
logger.configure(dir=logdir)
#Get parameters in kfac
kfac_paras = {}
for arg in vars(args):
if arg[:4] == 'kfac':
kfac_paras[arg[5:]] = getattr(args, arg)
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = deepq.wrap_atari_dqn(env)
model = deepq.models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (32, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
act = deepq.learn(
env,
q_func=model,
isKfac=args.isKfac,
kfac_paras=kfac_paras,
lr=args.lr,
max_timesteps=args.num_timesteps,
batch_size=args.batch_size,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=args.target_network_update_freq,
gamma=0.99,
prioritized_replay=bool(args.prioritized)
)
# act.save("pong_model.pkl") XXX
env.close()