def main():
parser = argparse.ArgumentParser(description='Run DQN on Atari Breakout')
parser.add_argument('--env', default='Enduro-v0', help='Atari env name')
parser.add_argument('-o',
'--output',
default='atari-v0',
help='Directory to save data to')
parser.add_argument('--seed', default=0, type=int, help='Random seed')
args = parser.parse_args()
# setup parameters
# create the model and the agent
# run fit method.
env = gym.make(args.env)
num_actions = env.action_space.n
replay = True # apply replay memory or not
double = False # apply double network or not
model_type = 'deep'
model1 = create_model((105, 80), num_actions, model_type=model_type)
model2 = create_model((105, 80), num_actions, model_type=model_type)
preprocessor = core.Preprocessor()
if replay:
mem_size = 40000 #1000000
window = 4
mem = core.ReplayMemory(mem_size, window)
else:
mem = list()
p = policy.LinearDecayGreedyEpsilonPolicy(1, 0.01, 500, num_actions)
agent = Agent(model_type=model_type,
q_network=model1,
preprocessor=preprocessor,
memory=mem,
policy=p,
gamma=0.99,
target_update_freq=1000,
num_burn_in=4000,
train_freq=10,
batch_size=32,
replay=replay,
target_network=model2,
double=double)
agent.compile()
agent.fit(env, num_iterations=400)
agent.q_network.save_weights('last-weights')
alpha = 0.001 #Learning Rate
num_iteration = 300000
#ATTENTION TO THE FOLLOWING VARIABLE:
Max_TimeStep = 10000#this is T described in the reference 2
#Make Max_TimeStep to be small so that I can debug on my PC
#Change that to big value when upload to the super computer
num_update_target = 10000 #target update after this number of iterations
#Again, make it small first to debug locally and then make big when submit!
update_counter = 0 #used for debug only, will not affect the algorithm
batch_size = 32
env = gym.make('SpaceInvaders-v0')
#if we use SpaceInvaders as suggested by the paper, we should use 3,
#while all the other games are required that we use 4
num_frame_skip = 3
output_num = env.action_space.n
LinearPolicy = policy.LinearDecayGreedyEpsilonPolicy(output_num, 1, 0.05, 100000)
experience = Experience()
#============BUILD NETWORK===========================
#define session to run:
sess = tf.Session()
#define placeholders for state, action, reward, nextstate, terminal
index1 = tf.placeholder(tf.int32, shape = [None, 2], name = 'index1')
index2 = tf.placeholder(tf.int32, shape = [None, 2], name = 'index2')
action = tf.placeholder(tf.int32, shape = [None, 2], name = 'action')
terminal = tf.placeholder(tf.float32, shape = [None, 1], name = 'terminal')
r = tf.placeholder(tf.float32, shape = [None, 1], name = 'r')
x = tf.placeholder(tf.float32, shape = [None, num_frame_skip*84*84], name = 'x')
next_x = tf.placeholder(tf.float32, shape = [None, num_frame_skip*84*84], name = 'next_x')
#weights for the Q1 linear Q network
#ATTENTION TO THE FOLLOWING VARIABLE:
rewardOneEpisode = 0 # sum the reward obtained from 1 episode
Max_TimeStep = 10000 #this is T described in the reference 2
#Make Max_TimeStep to be small so that I can debug on my PC
#Change that to big value when upload to the super computer!!!
num_update_target = 10000 #target update after this number of iterations
#Again, make it small first to debug locally and then make big when submit!
update_counter = 0 #used for debug only, will not affect the algorithm
batch_size = 32
env = gym.make('SpaceInvaders-v0')
#if we use SpaceInvaders as suggested by the paper, we should use 3,
#while all the other games are required that we use 4
num_frame_skip = 3
#==========Neural Nets Parameters are here=================
output_num_final = env.action_space.n
LinearPolicy = policy.LinearDecayGreedyEpsilonPolicy(output_num_final, 1, 0.05,
200000)
out_dim1 = 32 #number of filters for convolutional layer1
out_dim2 = 64 #number of filters for convolutional layer2
out_dim3 = 64 #number of filters for convolutional layer3
experience = Experience()
#============BUILD NETWORK===========================
# the data fomat used here is "NCHW"!!!
#define session to run:
sess = tf.Session()
#define placeholders for state, action, reward, nextstate, terminal
action = tf.placeholder(tf.int32, shape=[None, 2], name='action')
terminal = tf.placeholder(tf.float32, shape=[None, 1], name='terminal')
r = tf.placeholder(tf.float32, shape=[None, 1], name='r')
x = tf.placeholder(tf.float32,
def main():
parser = argparse.ArgumentParser(description = 'Run Atari')
parser.add_argument('--env',default='SpaceInvaders-v0',help ='Atari Env name')
parser.add_argument('-o', '--output', default='atari-v0', help='Directory to save data to')
parser.add_argument('--seed', default=0, type=int, help='Random seed')
parser.add_argument('--gpu', default='0', type=str, help='GPU to use.')
parser.add_argument('--optimizer', default='rmsprop', type=str,help='optimizer (rmsprop/adam).')
parser.add_argument('--learning_rate', default=0.00025, type=float,help='Learning rate.')
parser.add_argument('--model', default='convnet', type=str,
help='Type of model to use.')
parser.add_argument('--max_iters', default=100000000, type=int,
help='Max num of iterations to run for.')
parser.add_argument('--checkpoint',
default='',
type=str,
help='Checkpoint to load from.')
parser.add_argument('--render', action='store_true',
default=False,
help='Render what we got, or train?')
parser.add_argument('--render_path', type=str,
default='/dev/null/',
help='Path to store the render in.')
parser.add_argument('--input_image', action='store_true',
default=False,
help='Standardize (-1,1) the image or not.')
parser.add_argument('--part_gpu', action='store_true',
default=True,
help='Use part of GPU.')
parser.add_argument('--train_policy', type=str,
default='anneal',
help='anneal/epgreedy')
parser.add_argument('--exp_id', type=int,
default=None,
help='For assoc between scripts and results, '
'give script number.')
parser.add_argument('--target_update_freq', type=int,
default=10000,
help='Sync the target and live networks.')
parser.add_argument('--train_freq', type=int,
default=4,
help='Number of iters to push into replay before training.')
parser.add_argument('--mem_size', type=int,
default=100000,
help='Size of replay memory, 1M is too large.')
parser.add_argument('--learning_type', type=str,
default='normal',
help='Set normal, or double for DDQN.')
parser.add_argument('--final_eval', action='store_true',
default=False,
help='Perform the final 100 episode evaluation.')
args =parser.parse_args()
args.outout,experiment_id = get_output_folder(args.output,args.env,args.exp_id)
np.random.seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
if args.part_gpu:
from keras.backend.tensorflow_backend import set_session
config = tf.ConfigProto()
config.gpu_options.allow_growth =True
set_session(tf.Session(config=config))
input_size =64
frame_history =4
batch_size = 32
input_shape = (input_size,input_size,frame_history)
mem_size = args.mem_size
gamma =0.99
target_update_freq = args.target_update_freq
num_burn_in = 20000
train_freq = args.train_freq
summary_writer = None
if not args.render and not args.final_eval:
summary_writer= tf.summary.FileWriter(logdir =args.output)
env= gym.make(args.env)
env.seed(args.seed)
env.reset()
model = networks.get_model(args.model, input_shape, env.action_space.n)
target_model = networks.get_model(args.model, input_shape, env.action_space.n)
preproc = preprocessor.SequentialPreprocessor(
[preprocessor.AtariPreprocessor(input_size, args.std_img),
preprocessor.PreservePreprocessor(frame_history)])
if args.train_policy == 'anneal':
pol = policy.LinearDecayGreedyEpsilonPolicy(
1, 0.1, 1000000)
elif args.train_policy == 'epgreedy':
pol = policy.GreedyEpsilonPolicy(0.1)
else:
raise ValueError()
if args.optimizer == 'rmsprop':
optimizer = RMSprop(args.learning_rate, rho=0.95, epsilon=0.01)
elif args.optimizer == 'adam':
optimizer = Adam(args.learning_rate)
mem = memory.BasicReplayMemory(mem_size)
learning_type = dqn_utils.CASE_NORMAL
if args.learning_type == 'double':
learning_type = dqn_utils.CASE_DOUBLE
D = dqn_utils.DQNAgent(
model,
target_model,
preproc,
mem,
pol,
gamma,
target_update_freq,
num_burn_in,
train_freq,
batch_size,
optimizer=optimizer,
loss_func='mse',
summary_writer=summary_writer,
checkpoint_dir=args.output,
experiment_id=experiment_id,
env_name=args.env,
learning_type=learning_type)
if args.checkpoint:
D.load(args.checkpoint)
if args.final_eval:
D.evaluate(env, 100, 0, max_episode_length=1e4, final_eval=True)
elif args.render:
# testing_env = copy.deepcopy(env)
D.evaluate(env, 1, 0, max_episode_length=1e4,
render=True, render_path=args.render_path)
else:
D.fit(env, args.max_iters)