class Model:
''' simple one layer model for car racing '''
def __init__(self, load_model=True):
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_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 make_env(self, seed=-1, render_mode=False, full_episode=False):
self.render_mode = render_mode
self.env = make_env(self.env_name,
seed=seed,
render_mode=render_mode,
full_episode=full_episode)
def reset(self):
self.state = rnn_init_state(self.rnn)
def encode_obs(self, obs):
# convert raw obs to z, mu, logvar
print(obs.shape)
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 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
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 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, 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
# INIT The Controller After the enviroment Creation.
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)
self.na = self.env.action_space.n # discrete by default.
def init_controller(self):
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.na) # pong. Modify later.
self.bias_output = np.random.randn(self.na)
self.param_count = (self.input_size + 1) * self.hidden_size + (
self.hidden_size + 1) * self.na
else:
self.weight = np.random.randn(self.input_size, self.na)
self.bias = np.random.randn(self.na)
self.param_count = (self.input_size + 1) * self.na
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, epsilon=0.0):
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 np.random.rand() < epsilon:
action = np.random.randint(0, self.na)
else:
if EXP_MODE == MODE_Z_HIDDEN: # one hidden layer
h = np.maximum(
np.dot(h, self.weight_hidden) + self.bias_hidden, 0)
action = np.argmax(
np.dot(h, self.weight_output) + self.bias_output)
else:
action = np.argmax(np.dot(h, self.weight) + self.bias)
oh_action = np.zeros(self.na)
oh_action[action] = 1
# action[1] = (action[1]+1.0) / 2.0
# action[2] = clip(action[2])
# TODO check about this fucntion
self.state = rnn_next_state(self.rnn, z, oh_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[:self.na]
self.weight_output = params_2[self.na:].reshape(
self.hidden_size, self.na)
else:
self.bias = np.array(model_params[:self.na])
self.weight = np.array(model_params[self.na:]).reshape(
self.input_size, self.na)
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 Model:
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 = arglist.obs_size+ arglist.action_space + (arglist.agent_num-1) * arglist.action_space * arglist.timestep,
max_seq_len=1, use_recurrent_dropout=0, is_training=0,
obs_size = arglist.obs_size )
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 = self.rnn.rnn_init_state()
self.rnn_mode = True
print(arglist.inference)
if arglist.inference == True:
self.input_size = self.rnn.rnn_output_size(arglist.exp_mode) +(arglist.agent_num-1) * arglist.action_space
elif arglist.inference == False:
self.input_size = self.rnn.rnn_output_size(arglist.exp_mode) + (arglist.timestep) *( arglist.agent_num-1) * arglist.action_space
# 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 arglist.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
def reset(self):
self.state = self.rnn.rnn_init_state()
# self.oppo_state = lstm_init_state(self.oppo_model)
# 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 get_action(self, obs, act_traj):
h = self.rnn.rnn_output(self.state, obs, act_traj, self.arglist.exp_mode)
if self.arglist.inference:
oppo_intents = []
for i in range(self.arglist.agent_num - 1):
act_traj = self.act_traj[i]
# intent = self.oppo_model .get_inference(act_traj)
intent = [0,0]
oppo_intents.append(intent)
oppo_intents = np.reshape(oppo_intents, ((self.arglist.agent_num-1 )* self.arglist.action_space))
#Oppo intent shape (batch_size, agent_num, action_space)
# reshape oppo_intent agent_num * batch_size * action_space
controller_input = np.concatenate((h, oppo_intents))
else:
controller_input = h
if self.arglist.exp_mode == MODE_Z_HIDDEN: # one hidden layer
x = np.tanh(np.dot(controller_input, self.weight_hidden) + self.bias_hidden)
action = np.tanh(np.dot(x, self.weight_output) + self.bias_output)
else:
action = np.tanh(np.dot(controller_input, self.weight) + self.bias)
for i in range(self.action_space):
action[i] = clip(action[i])
self.state = self.rnn.rnn_next_state(obs, action, self.act_traj, self.state)
# self.oppo_state = oppo_next_state(self.oppo_model, action, self.act_traj, self.oppo_state)
# epsilon exploration
if np.random.uniform(0,1) < 0.2:
action = [np.random.uniform(-3,3)] * len(action)
return action
def set_model_params(self, model_params):
if self.arglist.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.action_space]
self.weight_output = params_2[self.action_space:].reshape(self.hidden_size, self.action_space)
else:
self.bias = np.array(model_params[:self.action_space])
self.weight = np.array(model_params[self.action_space:]).reshape(self.input_size, self.action_space)
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)
