def main():
if(len(sys.argv) != 5):
print("usage:{} <env> <model_json> <weights> <directory>".format(sys.argv[0]))
return sys.exit()
env = gym.make(sys.argv[1])
env.frameskip = 1
with open(sys.argv[2]) as json_file:
model = model_from_json(json.load(json_file),{"Eq9":Eq9})
model.load_weights(sys.argv[3])
epsilon = 0.01
input_shape = (84,84)
history_size = 4
eval_size = 1
directory = sys.argv[4]
history_prep = HistoryPreprocessor(history_size)
atari_prep = AtariPreprocessor(input_shape,0,999)
numpy_prep = NumpyPreprocessor()
preprocessors = PreprocessorSequence([atari_prep, history_prep, numpy_prep]) #from left to right
policy = GreedyEpsilonPolicy(epsilon)
agent = DQNAgent(model, preprocessors, None, policy, 0.99, None,None,None,None)
env = gym.wrappers.Monitor(env,directory,force=True)
reward_arr, length_arr = agent.evaluate_detailed(env,eval_size,render=False, verbose=True)
def main():
if (len(sys.argv) != 6):
print("usage:{} <env> <model_json> <weights> <render> <random>".format(
sys.argv[0]))
return sys.exit()
env = gym.make(sys.argv[1])
env.frameskip = 1
with open(sys.argv[2]) as json_file:
model = model_from_json(json.load(json_file), {"Eq9": Eq9})
model.load_weights(sys.argv[3])
epsilon = 0.01
input_shape = (84, 84)
history_size = 4
eval_size = 100
render = (sys.argv[4] == "y")
history_prep = HistoryPreprocessor(history_size)
atari_prep = AtariPreprocessor(input_shape, 0, 999)
numpy_prep = NumpyPreprocessor()
preprocessors = PreprocessorSequence(
[atari_prep, history_prep, numpy_prep]) #from left to right
if (sys.argv[5] == "y"):
print("using random policy")
policy = UniformRandomPolicy(env.action_space.n)
else:
print("using greedy policy")
policy = GreedyEpsilonPolicy(epsilon)
agent = DQNAgent(model, preprocessors, None, policy, 0.99, None, None,
None, None)
agent.add_keras_custom_layers({"Eq9": Eq9})
reward_arr, length_arr = agent.evaluate_detailed(env,
eval_size,
render=render,
verbose=True)
print("\rPlayed {} games, reward:M={}, SD={} length:M={}, SD={}".format(
eval_size, np.mean(reward_arr), np.std(reward_arr),
np.mean(length_arr), np.std(reward_arr)))
print("max:{} min:{}".format(np.max(reward_arr), np.min(reward_arr)))
plt.hist(reward_arr)
plt.show()
def testPerformance(self):
"""
Test to make sure each model(DQN, DDQN, DoubleQN) could be created and compiled
"""
#create a model of the world
env = gym.make("SpaceInvaders-v0")
env.frameskip = 1
#create a fake keras model
input_shape = (84, 84)
window = 4
num_actions = env.action_space.n
model = Sequential(name="test_model")
model.add(
Convolution2D(filters=16,
kernel_size=8,
strides=4,
activation='relu',
input_shape=(input_shape[0], input_shape[1],
window)))
model.add(
Convolution2D(filters=32,
kernel_size=4,
strides=2,
activation='relu'))
model.add(
Convolution2D(filters=64,
kernel_size=3,
strides=1,
activation='relu'))
model.add(Flatten())
model.add(Dense(units=512, activation='relu'))
model.add(Dense(units=num_actions, activation='linear'))
#create loss function & optimizer
optimizer = Adam(lr=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
loss_func = huber_loss
#preprocessors
history_prep = HistoryPreprocessor(4)
atari_prep = AtariPreprocessor(input_shape, 0, 999)
numpy_prep = NumpyPreprocessor()
preprocessors = PreprocessorSequence(
[atari_prep, history_prep, numpy_prep]) #from left to right
memory = ActionReplayMemory(100000, 4)
#policy = LinearDecayGreedyEpsilonPolicy(1, 0.1,100000)
policy = SamePolicy(1)
#agent = DQNAgent(model, preprocessors, memory, policy,0.99, target_update_freq,None,train_freq,batch_size)
dqn_agent = DQNAgent(model, preprocessors, memory, policy, 0.99, 10000,
None, 4, 32)
dqn_agent.compile(optimizer, loss_func)
total_time = 0
times = 50
for i in range(0, times):
start_time = time.time()
dqn_agent.evaluate_detailed(env, 1)
total_time += (time.time() - start_time)
sys.stdout.write('\r{}'.format(i))
sys.stdout.flush()
print("average evaluation time:{} total time:{}".format(
total_time / times, total_time))
