def __init__(self, arglist):
self.env_name = arglist.game
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.vae.load_json(arglist.vae_file)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
self.rnn.load_json(arglist.rnn_file)
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 2)
self.bias_output = np.random.randn(2)
self.param_count = ((self.input_size + 1) *
self.hidden_size) + (self.hidden_size * 2 + 2)
else:
self.weight = np.random.randn(self.input_size, 2)
self.bias = np.random.randn(2)
self.param_count = (self.input_size) * 2 + 2
self.render_mode = False
def __init__(self):
self.env_name = "carracing"
self.vae = ConvVAE(batch_size=1, gpu_mode=False, is_training=False, reuse=True)
self.vae.load_json('vae/vae.json')
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size, self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 3)
self.bias_output = np.random.randn(3)
self.param_count = ((self.input_size+1)*self.hidden_size) + (self.hidden_size*3+3)
else:
self.weight = np.random.randn(self.input_size, 3)
self.bias = np.random.randn(3)
self.param_count = (self.input_size)*3+3
self.render_mode = False
def __init__(self, load_model=True):
# For Mac
# self.env_name = "/Users/intuinno/codegit/pushBlock/app/mac/VisualPushBlockContinuous"
# For linux
self.env_name = "/home/intuinno/codegit/pushblock/app/linux/pushblock.x86_64"
self.vae = ConvVAE(batch_size=1, gpu_mode=False, is_training=False, reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size, self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 3)
self.bias_output = np.random.randn(3)
self.param_count = ((self.input_size+1)*self.hidden_size) + (self.hidden_size*3+3)
else:
self.weight = np.random.randn(self.input_size, 3)
self.bias = np.random.randn(3)
self.param_count = (self.input_size)*3+3
self.render_mode = False
def __init__(self, load_model=True):
self.env_name = './VisualPushBlock_withBlock_z_info.x86_64' #'./VisualPushBlock.x86_64'
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = z_size
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer ###CHANGE is made here
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, ACTION_SIZE)
self.bias_output = np.random.randn(ACTION_SIZE)
self.param_count = ((self.input_size + 1) * self.hidden_size) + (
self.hidden_size * ACTION_SIZE + ACTION_SIZE)
else:
self.weight = np.random.randn(self.input_size, ACTION_SIZE)
self.bias = np.random.randn(ACTION_SIZE)
self.param_count = (self.input_size) * ACTION_SIZE + ACTION_SIZE
self.render_mode = False
def __init__(self, arglist, action_space, scope, load_model=True):
self.action_space = action_space
self.arglist = arglist
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
hps_sample = hps_model._replace(
batch_size=1,
input_seq_width=32 + arglist.action_space +
(arglist.agent_num - 1) * arglist.action_space * arglist.timestep,
max_seq_len=1,
use_recurrent_dropout=0,
is_training=0)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json(arglist.vae_model_dir)
self.rnn.load_json(arglist.rnn_model_dir)
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
if arglist.inference:
self.input_size = rnn_output_size(
EXP_MODE) + (arglist.agent_num - 1) * arglist.action_space
else:
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
# action trajectories recording
self.act_traj = [
collections.deque(np.zeros(
(arglist.timestep, arglist.action_space)),
maxlen=arglist.timestep)
] * (arglist.agent_num - 1)
self.oppo_model = Oppo_Model(arglist.agent_num, arglist.timestep,
arglist.action_space,
arglist.action_space,
"oppo_model_{}".format(scope))
self.inference = arglist.inference
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size,
self.action_space)
self.bias_output = np.random.randn(self.action_space)
self.param_count = ((self.input_size + 1) * self.hidden_size) + (
self.hidden_size * self.action_space + self.action_space)
else:
self.weight = np.random.randn(self.input_size, self.action_space)
self.bias = np.random.randn(self.action_space)
self.param_count = (
self.input_size) * self.action_space + self.action_space
class ModelMCTS(Model):
def __init__(self, load_model=True):
self.env_name = "carracing"
self.env = make_env(self.env_name,
seed=SEED,
render_mode=render_mode,
full_episode=False)
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('../vae/vae.json')
self.rnn.load_json('../rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 3)
self.bias_output = np.random.randn(3)
self.param_count = ((self.input_size + 1) *
self.hidden_size) + (self.hidden_size * 3 + 3)
else:
self.weight = np.random.randn(self.input_size, 3)
self.bias = np.random.randn(3)
self.param_count = (self.input_size) * 3 + 3
self.render_mode = False
self.mct = None
def get_action(self, z):
a = random_linear_sample(-1, 1)
b = random_linear_sample(0, 1)
c = random_linear_sample(0, 1)
actions = dp(a, b, c)
action, self.mct = mcts.mcts(z,
self.env,
actions,
old_tree=self.mct,
tree_depth=6,
simulate_depth=200)
self.state = rnn_next_state(self.rnn, z, action, self.state)
return action
def __init__(self, full_episode=False):
super(VAERacing, self).__init__()
self._internal_counter = 0
self.z_size = games['vae_racing'].input_size
self.vae = ConvVAE(batch_size=1, z_size=self.z_size, gpu_mode=False, is_training=False, reuse=True)
self.vae.load_json('vae/vae_'+str(self.z_size)+'.json')
self.full_episode = full_episode
high = np.array([np.inf] * self.z_size)
self.observation_space = Box(-high, high)
self._has_rendered = False
self.real_frame = None
def __init__(self,
env,
batchsize=64,
input_size=(64,64),
num_frame_stack=4,
gamma=0.95,
frame_skip=1,
train_freq=4,
initial_epsilon=1.0,
min_epsilon=0.1,
render=True,
epsilon_decay_steps=int(1e6),
min_experience_size=int(1e3),
experience_capacity=int(1e5),
network_update_freq=5000,
regularization=1e-6,
optimizer_params=None,
action_map=None
):
self.vae = ConvVAE(batch_size=batchsize, gpu_mode=False, is_training=False, reuse=True)
self.vae.load_json('vae/vae.json')
if action_map is not None:
self.dim_actions = len(action_map)
else:
self.dim_actions = env.action_space.n
self.network_update_freq = network_update_freq
self.action_map = action_map
self.env = env
self.batchsize = batchsize
self.num_frame_stack = num_frame_stack
self.gamma = gamma
self.frame_skip = frame_skip
self.train_freq = train_freq
self.initial_epsilon = initial_epsilon
self.min_epsilon = min_epsilon
self.epsilon_decay_steps = epsilon_decay_steps
self.render = render
self.min_experience_size = min_experience_size
self.input_size = input_size
self.regularization = regularization
self.optimizer_params = optimizer_params or dict(learning_rate=0.0004, epsilon=1e-7)
self.do_training = True
self.playing_epsilon = 0.0
self.session = None
self.state_size = (self.num_frame_stack,) + self.input_size
self.global_counter = 0
self.episode_counter =0
def __init__(self, full_episode=False, discrete_mode=False):
super(VAERacingStack, self).__init__()
self._internal_counter = 0
self.z_size = games['vae_racing_stack'].input_size
self.vae = ConvVAE(batch_size=1,
z_size=self.z_size,
num_channel=FRAME_STACK,
gpu_mode=False,
is_training=False,
reuse=True)
self.vae.load_json('vae/vae_stack_' + str(FRAME_STACK) + '.json')
self.full_episode = full_episode
high = np.array([np.inf] * self.z_size)
self.observation_space = Box(-high, high)
self.cumulative_frames = None
self._has_rendered = False
self.discrete_mode = discrete_mode
def __init__(self,
model_name='',
load_model=True,
load_full_model=False,
full_model_path=''):
self.model_name = model_name
self.env_name = "carracing"
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_full_model:
self.vae.load_json(os.path.join(full_model_path, 'vae.json'))
self.rnn.load_json(os.path.join(full_model_path, 'rnn.json'))
elif load_model:
self.vae.load_json(
os.path.join(vae_path, self.model_name + '_vae.json'))
self.rnn.load_json(
os.path.join(rnn_path, self.model_name + '_rnn.json'))
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 3)
self.bias_output = np.random.randn(3)
self.param_count = ((self.input_size + 1) *
self.hidden_size) + (self.hidden_size * 3 + 3)
else:
self.weight = np.random.randn(self.input_size, 3)
self.bias = np.random.randn(3)
self.param_count = (self.input_size) * 3 + 3
self.render_mode = False
def __init__(self, type="CarRacing", history_pick=4, seed=None, detect_edges=False, detect_grass=False, flip=False):
self.name = type + str(time.time())
random.seed(30)
self.env = make_env('CarRacing-v0', random.randint(1,10000000), render_mode = False, full_episode = True)
self.image_dimension = [64,64]
self.history_pick = history_pick
self.state_space_size = history_pick * np.prod(self.image_dimension)
self.action_space_size = 5
self.state_shape = [None, self.history_pick] + list(self.image_dimension)
self.history = []
self.action_dict = {0: [-1, 0, 0], 1: [1, 0, 0], 2: [0, 1, 0], 3: [0, 0, 0.8], 4: [0, 0, 0]}
self.seed = seed
self.detect_edges = detect_edges
self.detect_grass = detect_grass
self.flip = flip
self.flip_episode = False
self.vae = ConvVAE(batch_size=1, gpu_mode=False, is_training=False, reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
def __init__(self,
sess=None,
summary_writer=tf.summary.FileWriter("logs/"),
rl_training=False,
reuse=False,
cluster=None,
index=0,
device='/gpu:0',
ppo_load_path=None,
ppo_save_path=None,
load_worldmodel=True,
ntype='worldmodel'):
self.policy_model_path_load = ppo_load_path + ntype
self.policy_model_path_save = ppo_save_path + ntype
self.rl_training = rl_training
self.use_norm = True
self.reuse = reuse
self.sess = sess
self.cluster = cluster
self.index = index
self.device = device
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_worldmodel:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.input_size = rnn_output_size(EXP_MODE)
self._create_graph()
self.rl_saver = tf.train.Saver()
self.summary_writer = summary_writer
class CarRacingMDNRNN(CarRacingWrapper):
def __init__(self, load_model=True, full_episode=False):
super(CarRacingMDNRNN, self).__init__(full_episode=full_episode)
self.vae = CVAE(batch_size=1)
self.rnn = MDNRNN(hps_sample)
if load_model:
self.vae.load_json('tf_vae/vae.json')
self.rnn.load_json('tf_rnn/rnn.json')
self.rnn_states = rnn_init_state(self.rnn)
self.full_episode = False
self.observation_space = Box(low=np.NINF, high=np.Inf, shape=(32+256))
def encode_obs(self, obs):
# convert raw obs to z, mu, logvar
result = np.copy(obs).astype(np.float)/255.0
result = result.reshape(1, 64, 64, 3)
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar/2.0) * np.random.randn(*s)
return z, mu, logvar
def reset(self):
self.rnn_states = rnn_init_state(self.rnn)
z_h = super(CarRacingWrapper, self).reset() # calls step
return z_h
def _step(self, action):
obs, reward, done, _ = super(CarRacingMDNRNN, self)._step(action)
z, _, _ = self.encode_obs(obs)
h = tf.squeeze(self.rnn_states[0])
z_h = tf.concat([z, h], axis=-1)
if action is not None: # don't compute state on reset
self.rnn_states = rnn_next_state(self.rnn, z, action, self.rnn_states)
return z_h, reward, done, {}
def __init__(self, load_model=True):
self.env_name = "Pong"
self._make_env()
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
hps_sample_dynamic = hps_sample._replace(num_actions=self.num_actions)
self.rnn = MDNRNN(hps_sample_dynamic, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
raise Exception("not ported for atari")
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size,
self.num_actions)
self.bias_output = np.random.randn(self.num_actions)
self.param_count = ((self.input_size + 1) * self.hidden_size) + (
(self.hidden_size + 1) * self.num_actions)
else:
# TODO: Not known until env.action_space is queried...
self.weight = np.random.randn(self.input_size, self.num_actions)
self.bias = np.random.randn(self.num_actions)
self.param_count = (self.input_size + 1) * self.num_actions
self.render_mode = False
def __init__(self, load_model=True, full_episode=False):
super(CarRacingMDNRNN, self).__init__(full_episode=full_episode)
self.vae = CVAE(batch_size=1)
self.rnn = MDNRNN(hps_sample)
if load_model:
self.vae.load_json('tf_vae/vae.json')
self.rnn.load_json('tf_rnn/rnn.json')
self.rnn_states = rnn_init_state(self.rnn)
self.full_episode = False
self.observation_space = Box(low=np.NINF, high=np.Inf, shape=(32+256))
def __init__(self, load_model=True, env_name="Pong-v0", render_mode=False):
self.env_name = env_name
self.make_env()
self.z_size = 32
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
hps_atari = hps_sample._replace(input_seq_width=self.z_size + self.na)
self.rnn = MDNRNN(hps_atari, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.init_controller()
self.render_mode = False
def sample_vae2(args):
""" For vae from https://github.com/hardmaru/WorldModelsExperiments.git
"""
z_size = 32
batch_size = args.count
learning_rate = 0.0001
kl_tolerance = 0.5
model_path_name = "tf_vae"
reset_graph()
vae = ConvVAE(
z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
is_training=False,
reuse=False,
gpu_mode=False) # use GPU on batchsize of 1000 -> much faster
vae.load_json(os.path.join(model_path_name, 'vae.json'))
z = np.random.normal(size=(args.count, z_size))
samples = vae.decode(z)
input_dim = samples.shape[1:]
n = args.count
plt.figure(figsize=(20, 4))
plt.title('VAE samples')
for i in range(n):
ax = plt.subplot(2, n, i + 1)
plt.imshow(samples[i].reshape(input_dim[0], input_dim[1],
input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
#plt.savefig( image_path )
plt.show()
def __init__(self, full_episode=False, pure_world=False):
super(VAERacingWorld, self).__init__()
self._internal_counter = 0
self.z_size = games['vae_racing'].input_size
self.vae = ConvVAE(batch_size=1,
z_size=self.z_size,
gpu_mode=False,
is_training=False,
reuse=True)
self.vae.load_json('vae/vae_' + str(self.z_size) + '.json')
self.full_episode = full_episode
if pure_world:
high = np.array([np.inf] * 10)
else:
high = np.array([np.inf] * (self.z_size + 10))
self.observation_space = Box(-high, high)
self._has_rendered = False
self.real_frame = None
self.world_model = SimpleWorldModel(obs_size=16,
action_size=3,
hidden_size=10)
world_model_path = "./log/learn_vae_racing.cma.4.64.best.json"
self.world_model.load_model(world_model_path)
self.pure_world_mode = pure_world
class Model:
''' simple one layer model for car racing '''
def __init__(self):
self.env_name = "carracing"
self.vae = ConvVAE(batch_size=1, gpu_mode=False, is_training=False, reuse=True)
self.vae.load_json('vae/vae.json')
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size, self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, 3)
self.bias_output = np.random.randn(3)
self.param_count = ((self.input_size+1)*self.hidden_size) + (self.hidden_size*3+3)
else:
self.weight = np.random.randn(self.input_size, 3)
self.bias = np.random.randn(3)
self.param_count = (self.input_size)*3+3
self.render_mode = False
def make_env(self, seed=-1, render_mode=False):
self.render_mode = render_mode
self.env = make_env(self.env_name, seed=seed, render_mode=render_mode)
def reset(self):
self.state = rnn_init_state(self.rnn)
def encode_obs(self, obs):
# convert raw obs to z, mu, logvar
result = np.copy(obs).astype(np.float)/255.0
result = result.reshape(1, 64, 64, 3)
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar/2.0) * np.random.randn(*s)
return z, mu, logvar
def decode_obs(self, z):
# decode the latent vector
img = self.vae.decode(z.reshape(1, self.z_size)) * 255.
img = np.round(img).astype(np.uint8)
img = img.reshape(64, 64, 3)
return img
def get_action(self, z):
h = rnn_output(self.state, z, EXP_MODE)
'''
action = np.dot(h, self.weight) + self.bias
action[0] = np.tanh(action[0])
action[1] = sigmoid(action[1])
action[2] = clip(np.tanh(action[2]))
'''
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
h = np.tanh(np.dot(h, self.weight_hidden) + self.bias_hidden)
action = np.tanh(np.dot(h, self.weight_output) + self.bias_output)
else:
action = np.tanh(np.dot(h, self.weight) + self.bias)
action[1] = (action[1]+1.0) / 2.0
action[2] = clip(action[2])
self.state = rnn_next_state(self.rnn, z, action, self.state)
return action
def set_model_params(self, model_params):
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
params = np.array(model_params)
cut_off = (self.input_size+1)*self.hidden_size
params_1 = params[:cut_off]
params_2 = params[cut_off:]
self.bias_hidden = params_1[:self.hidden_size]
self.weight_hidden = params_1[self.hidden_size:].reshape(self.input_size, self.hidden_size)
self.bias_output = params_2[:3]
self.weight_output = params_2[3:].reshape(self.hidden_size, 3)
else:
self.bias = np.array(model_params[:3])
self.weight = np.array(model_params[3:]).reshape(self.input_size, 3)
def load_model(self, filename):
with open(filename) as f:
data = json.load(f)
print('loading file %s' % (filename))
self.data = data
model_params = np.array(data[0]) # assuming other stuff is in data
self.set_model_params(model_params)
def get_random_model_params(self, stdev=0.1):
return np.random.randn(self.param_count)*stdev
if not os.path.exists(arglist.series_dir):
os.makedirs(arglist.series_dir)
filelist = os.listdir(arglist.data_dir)
filelist.sort()
filelist = filelist[0:10000]
dataset, action_dataset, oppo_action_dataset = load_raw_data_list(
filelist, arglist)
reset_graph()
if arglist.use_vae:
vae = ConvVAE(
z_size=arglist.z_size,
batch_size=arglist.batch_size,
learning_rate=arglist.lr,
kl_tolerance=arglist.kl_tolerance,
is_training=False,
reuse=False,
gpu_mode=True) # use GPU on batchsize of 1000 -> much faster
vae.load_json(os.path.join(arglist.vae_path, 'vae.json'))
mu_dataset = []
logvar_dataset = []
action_dataset_real = []
oppo_action_dataset_real = []
for i in range(len(dataset)):
data_batch = dataset[i]
if len(data_batch) <= arglist.batch_size:
continue
else:
if arglist.use_image:
dataset = create_dataset_with_image(filelist, arglist.data_dir)
else:
dataset = create_dataset(filelist, arglist)
# split into batches:
total_length = len(dataset)
num_batches = int(np.floor(total_length/arglist.batch_size))
print("num_batches", num_batches)
reset_graph()
vae = ConvVAE(z_size=arglist.z_size,
batch_size=arglist.batch_size,
learning_rate=arglist.lr,
kl_tolerance=arglist.kl_tolerance,
is_training=True,
reuse=False,
gpu_mode=True)
# train loop:
print("train", "step", "loss", "recon_loss", "kl_loss")
for epoch in range(arglist.epoch):
np.random.shuffle(dataset)
for idx in range(num_batches):
batch = dataset[idx*arglist.batch_size:(idx+1)*arglist.batch_size]
obs = np.array(batch).astype(np.float)/255.0
feed = {vae.x: obs,}
filelist.sort()
filelist = filelist[0:10000]
#print("check total number of images:", count_length_of_filelist(filelist))
dataset = create_dataset(filelist)
# split into batches:
total_length = len(dataset)
num_batches = int(np.floor(total_length/batch_size))
print("num_batches", num_batches)
reset_graph()
vae = ConvVAE(z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
is_training=True,
reuse=False,
gpu_mode=True)
# train loop:
print("train", "step", "loss", "recon_loss", "kl_loss")
for epoch in range(NUM_EPOCH):
np.random.shuffle(dataset)
for idx in range(num_batches):
batch = dataset[idx*batch_size:(idx+1)*batch_size]
obs = batch.astype(np.float)
feed = {vae.x: obs,}
batch_size = 1000 # treat every episode as a batch of 1000!
learning_rate = 0.0001
kl_tolerance = 0.5
filelist = os.listdir(DATA_DIR)
filelist.sort()
filelist = filelist[0:1000]
dataset, action_dataset = load_raw_data_list(filelist)
reset_graph()
vae = ConvVAE(z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
is_training=False,
reuse=False,
gpu_mode=True) # use GPU on batchsize of 1000 -> much faster
vae.load_json(os.path.join(model_path_name, 'vae.json'))
mu_dataset = []
logvar_dataset = []
for i in range(len(dataset)):
data_batch = dataset[i]
mu, logvar, z = encode_batch(data_batch)
mu_dataset.append(mu.astype(np.float16))
logvar_dataset.append(logvar.astype(np.float16))
if ((i + 1) % 100 == 0):
print(i + 1)
filelist.sort()
filelist = filelist[:NUM_DATA]
#print("check total number of images:", count_length_of_filelist(filelist))
dataset = create_dataset(filelist, N=NUM_DATA, M=NUM_FRAMES)
# split into batches:
total_length = len(dataset)
num_batches = int(np.floor(total_length / batch_size))
print("num_batches", num_batches)
reset_graph()
vae = ConvVAE(z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
beta=beta,
is_training=True,
reuse=False,
gpu_mode=True)
print('Training beta-VAE with beta={:.1f}'.format(beta))
# train loop:
print("train", "step", "loss", "recon_loss", "kl_loss")
for epoch in range(NUM_EPOCH):
np.random.shuffle(dataset)
for idx in range(num_batches):
batch = dataset[idx * batch_size:(idx + 1) * batch_size]
obs = batch.astype(np.float) / 255.0
feed = {
def main( dirs, z_size=32, batch_size=100, learning_rate=0.0001, kl_tolerance=0.5, epochs=100, save_model=False, verbose=True, optimizer="Adam" ):
if save_model:
model_save_path = "tf_vae"
if not os.path.exists(model_save_path):
os.makedirs(model_save_path)
gen = DriveDataGenerator(dirs, image_size=(64,64), batch_size=batch_size, shuffle=True, max_load=10000, images_only=True )
num_batches = len(gen)
reset_graph()
vae = ConvVAE(z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
is_training=True,
reuse=False,
gpu_mode=True,
optimizer=optimizer)
early = EarlyStopping(monitor='loss', min_delta=0.1, patience=5, verbose=verbose, mode='auto')
early.set_model(vae)
early.on_train_begin()
best_loss = sys.maxsize
if verbose:
print("epoch\tstep\tloss\trecon_loss\tkl_loss")
for epoch in range(epochs):
for idx in range(num_batches):
batch = gen[idx]
obs = batch.astype(np.float)/255.0
feed = {vae.x: obs,}
(train_loss, r_loss, kl_loss, train_step, _) = vae.sess.run([
vae.loss, vae.r_loss, vae.kl_loss, vae.global_step, vae.train_op
], feed)
if train_loss < best_loss:
best_loss = train_loss
if save_model:
if ((train_step+1) % 5000 == 0):
vae.save_json("tf_vae/vae.json")
if verbose:
print("{} of {}\t{}\t{:.2f}\t{:.2f}\t{:.2f}".format( epoch, epochs, (train_step+1), train_loss, r_loss, kl_loss) )
gen.on_epoch_end()
early.on_epoch_end(epoch, logs={"loss": train_loss})
if vae.stop_training:
break
early.on_train_end()
# finished, final model:
if save_model:
vae.save_json("tf_vae/vae.json")
return best_loss
def sample_vae2(args):
""" For vae from https://github.com/hardmaru/WorldModelsExperiments.git
"""
z_size = 64 # This needs to match the size of the trained vae
batch_size = args.count
learning_rate = 0.0001
kl_tolerance = 0.5
model_path_name = "tf_vae"
reset_graph()
vae = ConvVAE(
z_size=z_size,
batch_size=batch_size,
learning_rate=learning_rate,
kl_tolerance=kl_tolerance,
is_training=False,
reuse=False,
gpu_mode=False) # use GPU on batchsize of 1000 -> much faster
vae.load_json(os.path.join(model_path_name, 'vae.json'))
z = np.random.normal(size=(args.count, z_size))
samples = vae.decode(z)
input_dim = samples.shape[1:]
gen = DriveDataGenerator(args.dirs,
image_size=(64, 64),
batch_size=args.count,
shuffle=True,
max_load=10000,
images_only=True)
orig = gen[0].astype(np.float) / 255.0
#mu, logvar = vae.encode_mu_logvar(orig)
#recon = vae.decode( mu )
recon = vae.decode(vae.encode(orig))
n = args.count
plt.figure(figsize=(20, 6), tight_layout=False)
plt.title('VAE samples')
for i in range(n):
ax = plt.subplot(3, n, i + 1)
plt.imshow(samples[i].reshape(input_dim[0], input_dim[1],
input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Random")
for i in range(n):
ax = plt.subplot(3, n, n + i + 1)
plt.imshow(orig[i].reshape(input_dim[0], input_dim[1], input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Real")
ax = plt.subplot(3, n, (2 * n) + i + 1)
plt.imshow(recon[i].reshape(input_dim[0], input_dim[1], input_dim[2]))
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
if 0 == i:
ax.set_title("Reconstructed")
plt.savefig("samples_vae.png")
plt.show()
class Model:
''' simple one layer model for translating game state to actions'''
def __init__(self, load_model=True):
self.env_name = "Pong"
self._make_env()
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
hps_sample_dynamic = hps_sample._replace(num_actions=self.num_actions)
self.rnn = MDNRNN(hps_sample_dynamic, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = 32
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
raise Exception("not ported for atari")
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size,
self.num_actions)
self.bias_output = np.random.randn(self.num_actions)
self.param_count = ((self.input_size + 1) * self.hidden_size) + (
(self.hidden_size + 1) * self.num_actions)
else:
# TODO: Not known until env.action_space is queried...
self.weight = np.random.randn(self.input_size, self.num_actions)
self.bias = np.random.randn(self.num_actions)
self.param_count = (self.input_size + 1) * self.num_actions
self.render_mode = False
def _make_env(self):
self.render_mode = render_mode
self.env = make_env(self.env_name)
self.num_actions = self.env.action_space.n
def make_env(self):
pass #TODO (Chazzz): eventually remove
def reset(self):
self.state = rnn_init_state(self.rnn)
def encode_obs(self, obs):
# convert raw obs to z, mu, logvar
result = np.copy(obs).astype(np.float) / 255.0
result = result.reshape(1, 64, 64, 1)
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar / 2.0) * np.random.randn(*s)
return z, mu, logvar
def get_action(self, z):
h = rnn_output(self.state, z, EXP_MODE)
# print(len(h), " h:", h) #TODO: 256+32 (the 32 comes first)
# So we could have 288*2*18 params, or 288*2*environment.action_space.n (6 for Pong)
'''
action = np.dot(h, self.weight) + self.bias
action[0] = np.tanh(action[0])
action[1] = sigmoid(action[1])
action[2] = clip(np.tanh(action[2]))
'''
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
raise Exception("Not ported to atari")
# h = np.tanh(np.dot(h, self.weight_hidden) + self.bias_hidden)
# action = np.tanh(np.dot(h, self.weight_output) + self.bias_output)
else:
# could probabilistically sample from softmax, but greedy
action = np.argmax(np.matmul(h, self.weight) + self.bias)
# action[1] = (action[1]+1.0) / 2.0
# action[2] = clip(action[2])
# print("Action:", action)
action_one_hot = np.zeros(self.num_actions)
action_one_hot[action] = 1
# print("Action hot:", action_one_hot)
self.state = rnn_next_state(self.rnn, z, action_one_hot, self.state)
return action
def set_model_params(self, model_params):
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
params = np.array(model_params)
cut_off = (self.input_size + 1) * self.hidden_size
params_1 = params[:cut_off]
params_2 = params[cut_off:]
self.bias_hidden = params_1[:self.hidden_size]
self.weight_hidden = params_1[self.hidden_size:].reshape(
self.input_size, self.hidden_size)
self.bias_output = params_2[:self.num_actions]
self.weight_output = params_2[self.num_actions:].reshape(
self.hidden_size, self.num_actions)
else:
self.bias = np.array(model_params[:self.num_actions])
self.weight = np.array(model_params[self.num_actions:]).reshape(
self.input_size, self.num_actions)
def load_model(self, filename):
with open(filename) as f:
data = json.load(f)
print('loading file %s' % (filename))
self.data = data
model_params = np.array(data[0]) # assuming other stuff is in data
self.set_model_params(model_params)
def get_random_model_params(self, stdev=0.1):
#return np.random.randn(self.param_count)*stdev
return np.random.standard_cauchy(
self.param_count) * stdev # spice things up
def init_random_model_params(self, stdev=0.1):
params = self.get_random_model_params(stdev=stdev)
self.set_model_params(params)
vae_params = self.vae.get_random_model_params(stdev=stdev)
self.vae.set_model_params(vae_params)
rnn_params = self.rnn.get_random_model_params(stdev=stdev)
self.rnn.set_model_params(rnn_params)
class Model:
''' simple one layer model for car racing '''
def __init__(self, load_model=True):
self.env_name = './VisualPushBlock_withBlock_z_info.x86_64' #'./VisualPushBlock.x86_64'
self.vae = ConvVAE(batch_size=1,
gpu_mode=False,
is_training=False,
reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
if load_model:
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
self.state = rnn_init_state(self.rnn)
self.rnn_mode = True
self.input_size = rnn_output_size(EXP_MODE)
self.z_size = z_size
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer ###CHANGE is made here
self.hidden_size = 40
self.weight_hidden = np.random.randn(self.input_size,
self.hidden_size)
self.bias_hidden = np.random.randn(self.hidden_size)
self.weight_output = np.random.randn(self.hidden_size, ACTION_SIZE)
self.bias_output = np.random.randn(ACTION_SIZE)
self.param_count = ((self.input_size + 1) * self.hidden_size) + (
self.hidden_size * ACTION_SIZE + ACTION_SIZE)
else:
self.weight = np.random.randn(self.input_size, ACTION_SIZE)
self.bias = np.random.randn(ACTION_SIZE)
self.param_count = (self.input_size) * ACTION_SIZE + ACTION_SIZE
self.render_mode = False
def make_env(self,
seed=-1,
render_mode=False,
full_episode=False,
worker_id=0):
self.render_mode = render_mode
self.env = make_env(self.env_name,
seed=seed,
render_mode=render_mode,
full_episode=full_episode,
worker_id=worker_id)
def reset(self):
self.state = rnn_init_state(self.rnn)
def encode_obs(self, obs):
# convert raw obs to z, mu, logvar
#result = np.copy(obs).astype(np.float)/255.0
result = np.copy(obs).astype(np.float)
result = result.reshape(1, IMAGE_W, IMAGE_H, 3)
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar / 2.0) * np.random.randn(*s)
return z, mu, logvar
def get_action(self, z):
h = rnn_output(self.state, z, EXP_MODE)
#print('h', h.shape, h)
'''
action = np.dot(h, self.weight) + self.bias
action[0] = np.tanh(action[0])
action[1] = sigmoid(action[1])
action[2] = clip(np.tanh(action[2]))
'''
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
h = np.tanh(np.dot(h, self.weight_hidden) + self.bias_hidden)
action = np.tanh(np.dot(h, self.weight_output) + self.bias_output)
else:
'''print(h.shape)
print(self.weight.shape)
print(self.bias.shape)'''
action = np.tanh(np.dot(h, self.weight) + self.bias)
'''for i in range(ACTION_SIZE):
action[i] = (action[i]+1.0) / 2.0 #all actions value are in range 0 to 1'''
#action[2] = clip(action[2])
self.state = rnn_next_state(self.rnn, z, action,
self.state) #update weights of MDN-RNN
return action
def set_model_params(self, model_params):
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
params = np.array(model_params)
cut_off = (self.input_size + 1) * self.hidden_size
params_1 = params[:cut_off]
params_2 = params[cut_off:]
self.bias_hidden = params_1[:self.hidden_size]
self.weight_hidden = params_1[self.hidden_size:].reshape(
self.input_size, self.hidden_size)
self.bias_output = params_2[:ACTION_SIZE]
self.weight_output = params_2[ACTION_SIZE:].reshape(
self.hidden_size, ACTION_SIZE)
else:
self.bias = np.array(model_params[:ACTION_SIZE])
self.weight = np.array(model_params[ACTION_SIZE:]).reshape(
self.input_size, ACTION_SIZE)
def load_model(self, filename):
with open(filename) as f:
data = json.load(f)
print('loading file %s' % (filename))
self.data = data
model_params = np.array(data[0]) # assuming other stuff is in data
self.set_model_params(model_params)
def get_random_model_params(self, stdev=0.1):
#return np.random.randn(self.param_count)*stdev
return np.random.standard_cauchy(
self.param_count) * stdev # spice things up
def init_random_model_params(self, stdev=0.1):
params = self.get_random_model_params(stdev=stdev)
self.set_model_params(params)
vae_params = self.vae.get_random_model_params(stdev=stdev)
self.vae.set_model_params(vae_params)
rnn_params = self.rnn.get_random_model_params(stdev=stdev)
self.rnn.set_model_params(rnn_params)
class VAERacing(CarRacing):
def __init__(self, full_episode=False):
super(VAERacing, self).__init__()
self._internal_counter = 0
self.z_size = games['vae_racing'].input_size
#print("vae_racing.py z", self.z_size)
self.vae = ConvVAE(batch_size=1, z_size=self.z_size, gpu_mode=True, is_training=False, reuse=True)
#print("vae_racing.py vae", self.vae)
self.vae.load_json('vae/vae_'+str(self.z_size)+'.json')
self.full_episode = full_episode
high = np.array([np.inf] * self.z_size)
self.observation_space = Box(-high, high)
self._has_rendered = False
self.real_frame = None
def reset(self):
self._internal_counter = 0
self._has_rendered = False
self.real_frame = None
obs = super(VAERacing, self).reset()
result = np.copy(_process_frame(obs)).astype(np.float)/255.0
result = result.reshape(1, 64, 64, 3)
self.real_frame = result
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar/2.0) * np.random.randn(*s)
if MU_MODE:
return mu
return z
def render(self, mode='human', close=False):
if mode == 'human' or mode == 'rgb_array':
self._has_rendered = True
return super(VAERacing, self).render(mode=mode)
def step(self, action):
#print("action", action)
if not self._has_rendered:
self.render("rgb_array")
self._has_rendered = False
if action is not None:
action[0] = _clip(action[0], lo=-1.0, hi=+1.0)
action[1] = _clip(action[1], lo=-1.0, hi=+1.0)
action[1] = (action[1]+1.0) / 2.0
action[2] = _clip(action[2])
obs, reward, done, _ = super(VAERacing, self).step(action)
result = np.copy(_process_frame(obs)).astype(np.float)/255.0
result = result.reshape(1, 64, 64, 3)
self.real_frame = result
#z = self.vae.encode(result).flatten()
mu, logvar = self.vae.encode_mu_logvar(result)
mu = mu[0]
logvar = logvar[0]
s = logvar.shape
z = mu + np.exp(logvar/2.0) * np.random.randn(*s)
if self.full_episode:
if MU_MODE:
return mu, reward, False, {}
else:
return z, reward, False, {}
self._internal_counter += 1
if self._internal_counter > TIME_LIMIT:
done = True
if MU_MODE:
#print("mu", mu)
return mu, reward, done, {}
return z, reward, done, {}
output_dir = "vae_test_result"
z_size=32
filelist = os.listdir(DATA_DIR)
filelist = [f for f in filelist if '.npz' in f]
obs = np.load(os.path.join(DATA_DIR, random.choice(filelist)))["obs"]
obs = np.expand_dims(obs, axis=-1)
obs = obs.astype(np.float32)/255.0
n = len(obs)
vae = ConvVAE(z_size=z_size,
batch_size=1,
is_training=False,
reuse=False,
gpu_mode=False)
vae.load_json(os.path.join(model_path_name, 'vae.json'))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
print(n, "images loaded")
for i in range(n):
frame = obs[i].reshape(1, 64, 64, 1)
batch_z = vae.encode(frame)
reconstruct = vae.decode(batch_z)
imsave(output_dir+'/%s.png' % pad_num(i), 255.*frame[0].reshape(64, 64))
imsave(output_dir+'/%s_vae.png' % pad_num(i), 255.*reconstruct[0].reshape(64, 64))
