class lwpr_dyn_model(DynamicsLearnerInterface):
def __init__(self,
history_length,
prediction_horizon,
difference_learning,
averaging,
streaming,
settings=None):
super().__init__(history_length,
prediction_horizon,
difference_learning,
averaging=averaging,
streaming=streaming)
self.model_ = LWPR(self._get_input_dim(), self.observation_dimension)
# Default values.
init_D = 25
init_alpha = 175
self.time_threshold = np.inf
if settings:
init_D = settings['init_D']
init_alpha = settings['init_alpha']
self.time_threshold = settings.get('time_threshold', np.inf)
self.model_.init_D = init_D * np.eye(self._get_input_dim())
self.model_.init_alpha = init_alpha * np.eye(self._get_input_dim())
def _learn(self, training_inputs, training_targets):
def gen(inputs, targets):
for i in range(inputs.shape[0]):
yield targets[i], inputs[i]
self._learn_from_stream(gen(training_inputs, training_targets),
training_inputs.shape[0])
def _learn_from_stream(self, training_generator, generator_size):
deck = deque(maxlen=100)
for count in range(generator_size):
training_target, training_input = next(training_generator)
assert training_input.shape[0] == self._get_input_dim()
assert training_target.shape[0] == self.observation_dimension
time_before_update = time.perf_counter()
self.model_.update(training_input, training_target)
elapsed_time = time.perf_counter() - time_before_update
deck.append(elapsed_time)
if count and count % 1000 == 0:
median_time = sorted(deck)[deck.maxlen // 2]
print('Update time for iter {} is {}'.format(
count, median_time))
if median_time > self.time_threshold:
break
def _predict(self, inputs):
assert self.model_, "a trained model must be available"
prediction = np.zeros((inputs.shape[0], self.observation_dimension))
for idx in range(inputs.shape[0]):
prediction[idx, :] = self.model_.predict(inputs[idx])
return prediction
def name(self):
return "LWPR"
# initialize the LWPR model
model = LWPR(1, 1)
model.init_D = 20 * eye(1)
model.update_D = True
model.init_alpha = 40 * eye(1)
model.meta = False
print Xtr
print Ytr
# train the model
for k in range(20):
ind = random.permutation(Ntr)
mse = 0
for i in range(Ntr):
yp = model.update(Xtr[ind[i]], Ytr[ind[i]])
mse = mse + (Ytr[ind[i], :] - yp)**2
nMSE = mse / Ntr / var(Ytr)
print "#Data: %5i #RFs: %3i nMSE=%5.3f" % (model.n_data, model.num_rfs,
nMSE)
# test the model with unseen data
Ntest = 500
Xtest = linspace(0, 10, Ntest)
Ytest = zeros((Ntest, 1))
Conf = zeros((Ntest, 1))
for k in range(500):
Ytest[k, :], Conf[k, :] = model.predict_conf(array([Xtest[k]]))
print model
# train the model
# for k in range(20):
# ind = np.random.permutation(Ntr)
# mse = 0
# for i in range(Ntr):
# yp = model.update(Xtr[ind[i]], Ytr[ind[i]])
# mse = mse + (Ytr[ind[i], :] - yp)**2
# nMSE = mse / Ntr / np.var(Ytr)
# print "#Data: %5i #RFs: %3i nMSE=%5.3f" % (model.n_data, model.num_rfs, nMSE)
for i in range(Ntr):
model.update(Xtr[i], Ytr[i])
print model.num_rfs
# test the model with unseen data
Ntest = 5000
Ttest = np.linspace(0, time, Ntest)
Xtest = np.exp(-ax * Ttest / time)
Ytest = np.zeros((Ntest, 1))
Conf = np.zeros((Ntest, 1))
for k in range(Ntest):
Ytest[k, :], Conf[k, :] = model.predict_conf(
np.array([Xtest[k] * (Ytr[-1] - Ytr[0])]))
plt.plot(t, Ytr, 'r.')
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--environment", type=str, default='AntBulletEnv-v0')
parser.add_argument("--no_data", type=int, default=10000)
args = parser.parse_args()
state_action, state, reward, next_state = gather_data(
args.no_data, args.environment)
assert len(state_action) == len(next_state)
assert len(state_action) == len(reward)
'''
data = pickle.load(open('data.p'))
state2, action2, reward2, next_state2 = data[0]
state_action2 = np.concatenate([state2, action2], axis=-1)
state_action = np.concatenate([state_action, state_action2], axis=0)
reward = np.concatenate([reward, reward2], axis=0)
next_state = np.concatenate([next_state, next_state2], axis=0)
'''
no_data = len(state_action)
model_state = LWPR(state_action.shape[-1], next_state.shape[-1])
model_state.init_D = 5. * np.eye(state_action.shape[-1])
model_state.update_D = True
model_state.init_alpha = 1. * np.eye(state_action.shape[-1])
model_state.meta = True
action_shape = state_action.shape[-1] - next_state.shape[-1]
model_state.norm_in = np.array(([10.] * state.shape[-1]) +
[2.] * action_shape)
model_reward = LWPR(state_action.shape[-1], reward.shape[-1])
model_reward.init_D = 1. * np.eye(state_action.shape[-1])
model_reward.update_D = True
model_reward.init_alpha = 20. * np.eye(state_action.shape[-1])
model_reward.meta = True
#for k in range(20):
for k in range(1):
ind = np.random.permutation(no_data)
for i in range(no_data):
print(k, i)
model_state.update(state_action[ind[i]], next_state[ind[i]])
#model_state.update(state_action[ind[i]], next_state[ind[i]] - state[ind[i]])
model_reward.update(state_action[ind[i]], reward[ind[i]])
uid = str(uuid.uuid4())
for k in range(10):
state_action_test, state_test, reward_test, next_state_test = gather_data_epoch(
1, args.environment)
'''
if k % 2 == 0:
state_action_test, state_test, reward_test, next_state_test = gather_data_epoch(1, args.environment)
else:
idx = np.random.randint(1, len(data))
state_test, action_test, reward_test, next_state_test = data[idx]
state_action_test = np.concatenate([state_test, action_test], axis=-1)
'''
Y = []
confs = []
Y_r = []
confs_r = []
for i in range(len(state_action_test)):
y, conf = model_state.predict_conf(state_action_test[i])
#Y.append(y + state_test[i])
Y.append(y)
confs.append(conf)
y_r, conf_r = model_reward.predict_conf(state_action_test[i])
Y_r.append(y_r)
confs_r.append(conf_r)
Y = np.stack(Y, axis=0)
confs = np.stack(confs, axis=0)
Y_r = np.stack(Y_r, axis=0)
confs_r = np.stack(confs_r, axis=0)
for i in range(next_state.shape[-1]):
plt.figure()
print('Here is the length of the trajectory:',
len(next_state_test))
assert len(next_state_test[:, i:i + 1]) == len(Y[:, i:i + 1])
#plt.plot(np.arange(len(next_state_test[:, i:i+1])), next_state_test[:, i:i+1] - state_test[:, i:i+1])
plt.plot(np.arange(len(next_state_test[:, i:i + 1])),
next_state_test[:, i:i + 1])
plt.errorbar(np.arange(len(Y[:, i:i + 1])),
Y[:, i:i + 1],
yerr=confs[:, i:i + 1],
color='r',
ecolor='y')
plt.grid()
#plt.savefig(args.environment+'_'+'k:'+str(k)+'_'+'dim:'+str(i)+'_'+uid+'.pdf')
plt.figure()
plt.plot(np.arange(len(reward_test)), reward_test)
plt.errorbar(np.arange(len(Y_r)),
Y_r,
yerr=confs_r,
color='r',
ecolor='g')
plt.grid()
plt.show()
class LWPRFA(FA):
parametric = False
def __init__(self, indim, outdim):
FA.__init__(self, indim, outdim)
self.filename = None
def reset(self):
FA.reset(self)
# initialize the LWPR function
self.lwpr = LWPR(self.indim, self.outdim)
self.lwpr.init_D = 10.*np.eye(self.indim)
self.lwpr.init_alpha = 0.1*np.ones([self.indim, self.indim])
self.lwpr.meta = True
def predict(self, inp):
""" predict the output for the given input. """
# the next 3 lines fix a bug when lwpr models are pickled and unpickled again
# without it, a TypeError is thrown "Expected a double precision numpy array."
# even though the numpy array is double precision.
inp = self._asFlatArray(inp)
inp_tmp = np.zeros(inp.shape)
inp_tmp[:] = inp
return self.lwpr.predict(inp_tmp)
def train(self):
for i, t in self.dataset:
i = self._asFlatArray(i)
t = self._asFlatArray(t)
self.lwpr.update(i, t)
def _cleanup(self):
if self.filename and os.path.exists(self.filename):
os.remove(self.filename)
def __getstate__(self):
""" required for pickle. removes the lwpr model from the dictionary
and saves it to file explicitly.
"""
# create unique hash key for filename and write lwpr to file
hashkey = hashlib.sha1(str(self.lwpr) + time.ctime() + str(np.random.random())).hexdigest()[:8]
if not os.path.exists('.lwprmodels'):
os.makedirs('.lwprmodels')
# remove any old files if existing
if self.filename:
os.remove(self.filename)
self.filename = '.lwprmodels/lwpr_%s.binary'%hashkey
self.lwpr.write_binary(self.filename)
# remove lwpr from dictionary and return state
state = self.__dict__.copy()
del state['lwpr']
return state
def __setstate__(self, state):
""" required for pickle. loads the stored lwpr model explicitly.
"""
self.__dict__.update(state)
self.lwpr = LWPR(self.filename)
def train_lwpr(datafile, resultfolder, max_num_train, patience_list, improvement_threshold, init_lwpr_setting, hist_window, start_epoch=0, cmd_scaler=1.0, modelfile='lwpr_model'):
curr_path = os.getcwd()
if resultfolder in os.listdir(curr_path):
print "subfolder exists"
else:
print "Not Exist, so make subfolder"
os.mkdir(resultfolder)
# Load Data
dataset = loadmat(datafile)
train_data_x, train_data_y = dataset['train_data_x'], dataset['train_data_y']
valid_data_x, valid_data_y = dataset['valid_data_x'], dataset['valid_data_y']
num_data, num_valid = train_data_x.shape[0], valid_data_x.shape[0]
speed_hw, cmd_hw = hist_window[0], hist_window[1]
input_dim = 2*(speed_hw+cmd_hw)
# normalize command part
train_data_x[:, 2*speed_hw:] = train_data_x[:, 2*speed_hw:] * cmd_scaler
valid_data_x[:, 2*speed_hw:] = valid_data_x[:, 2*speed_hw:] * cmd_scaler
# Set-up Parameters/Model for Training Procedure
max_num_trials = max_num_train
improvement_threshold = improvement_threshold
error_hist, best_model_error, prev_train_time = [], np.inf, 0
initD, initA, penalty = init_lwpr_setting[0], init_lwpr_setting[1], init_lwpr_setting[2]
w_gen, w_prune = init_lwpr_setting[3], init_lwpr_setting[4]
best_model_epoch = 0
if start_epoch < 1:
# Initialize Two 1-Dimensional Models
LWPR_model_left = LWPR(input_dim, 1)
#LWPR_model_left.init_D = initD * np.eye(input_dim)
tmp_arr = np.ones(input_dim)
tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5]
LWPR_model_left.init_D = initD * np.diag(tmp_arr)
LWPR_model_left.update_D = False # True
#LWPR_model_left.init_alpha = initA * np.eye(input_dim)
tmp_arr = np.ones(input_dim)
tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5]
LWPR_model_left.init_alpha = initA * np.diag(tmp_arr)
LWPR_model_left.penalty = penalty
LWPR_model_left.meta = True
LWPR_model_left.meta_rate = 20
LWPR_model_left.w_gen = w_gen
LWPR_model_left.w_prune = w_prune
LWPR_model_right = LWPR(input_dim, 1)
#LWPR_model_right.init_D = initD * np.eye(input_dim)
tmp_arr = np.ones(input_dim)
tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5]
LWPR_model_right.init_D = initD * np.diag(tmp_arr)
LWPR_model_right.update_D = False # True
#LWPR_model_right.init_alpha = initA * np.eye(input_dim)
tmp_arr = np.ones(input_dim)
tmp_arr[input_dim-2*cmd_hw:input_dim] = init_lwpr_setting[5]
LWPR_model_right.init_alpha = initA * np.diag(tmp_arr)
LWPR_model_right.penalty = penalty
LWPR_model_right.meta = True
LWPR_model_right.meta_rate = 20
LWPR_model_right.w_gen = w_gen
LWPR_model_right.w_prune = w_prune
patience = patience_list[0]
else:
modelfile_name = './' + resultfolder + '/' + modelfile + '_left_epoch' + str(start_epoch-1) + '.bin'
LWPR_model_left = LWPR(modelfile_name)
print '\tRead LWPR model for left wheel(%d)' % (LWPR_model_left.num_rfs[0])
modelfile_name = './' + resultfolder + '/' + modelfile + '_right_epoch' + str(start_epoch-1) + '.bin'
LWPR_model_right = LWPR(modelfile_name)
print '\tRead LWPR model for right wheel(%d)' % (LWPR_model_right.num_rfs[0])
result_file_name = './' + resultfolder + '/Result_of_training_epoch' + str(start_epoch-1) + '.mat'
result_file = loadmat(result_file_name)
prev_train_time = result_file['train_time']
patience = result_file['patience']
best_model_error = result_file['best_model_error']
for cnt in range(start_epoch):
error_hist.append([result_file['history_validation_error'][cnt][0], result_file['history_validation_error'][cnt][1], result_file['history_validation_error'][cnt][2]])
# Training Part
model_prediction = np.zeros(valid_data_y.shape)
tmp_x, tmp_y = np.zeros((input_dim, 1)), np.zeros((1,1))
print 'start training'
start_train_time = timeit.default_timer()
for train_cnt in range(start_epoch, max_num_trials):
if patience < train_cnt:
break
rand_ind = np.random.permutation(num_data)
for data_cnt in range(num_data):
tmp_x[:,0] = train_data_x[rand_ind[data_cnt], 0:input_dim]
tmp_y[0,0] = train_data_y[rand_ind[data_cnt], 0]
_ = LWPR_model_left.update(tmp_x, tmp_y)
tmp_y[0,0] = train_data_y[rand_ind[data_cnt], 1]
_ = LWPR_model_right.update(tmp_x, tmp_y)
if data_cnt % 5000 == 0:
print '\ttrain epoch %d, data index %d, #rfs=%d/%d' % (train_cnt, data_cnt, LWPR_model_left.num_rfs, LWPR_model_right.num_rfs)
for data_cnt in range(num_valid):
tmp_x[:,0] = valid_data_x[data_cnt, 0:input_dim]
model_prediction[data_cnt, 0], _ = LWPR_model_left.predict_conf(tmp_x)
model_prediction[data_cnt, 1], _ = LWPR_model_right.predict_conf(tmp_x)
diff = abs(valid_data_y - model_prediction)
new_error = np.asarray([np.sum(diff)/float(num_valid), np.sqrt(np.sum(diff**2)/float(num_valid)), np.max(diff)])
error_hist.append([new_error[0], new_error[1], new_error[2]])
# save result of one training epoch
modelfile_name = './' + resultfolder + '/' + modelfile + '_left_epoch' + str(train_cnt) + '.bin'
LWPR_model_left.write_binary(modelfile_name)
modelfile_name = './' + resultfolder + '/' + modelfile + '_right_epoch' + str(train_cnt) + '.bin'
LWPR_model_right.write_binary(modelfile_name)
if new_error[1] < best_model_error * improvement_threshold:
best_model_epoch = train_cnt
best_model_error = new_error[1]
patience = max(patience, min(train_cnt+10, int(train_cnt * patience_list[1])) )
modelfile_name = './' + resultfolder + '/' + modelfile + '_best_left_epoch' + str(train_cnt) + '.bin'
LWPR_model_left.write_binary(modelfile_name)
modelfile_name = './' + resultfolder + '/' + modelfile + '_best_right_epoch' + str(train_cnt) + '.bin'
LWPR_model_right.write_binary(modelfile_name)
result_file_name = './' + resultfolder + '/Result_of_training_epoch' + str(train_cnt) + '.mat'
result = {}
result['train_time'] = timeit.default_timer() - start_train_time + prev_train_time
result['best_model_error'] = best_model_error
result['history_validation_error'] = error_hist
result['patience'] = patience
result['improvement_threshold'] = improvement_threshold
result['init_D'] = initD
result['init_alpha'] = initA
result['penalty'] = penalty
result['w_generate_criterion'] = w_gen
result['w_prune_criterion'] = w_prune
result['number_speed_in_input'] = 2*speed_hw
result['number_cmd_in_input'] = 2*cmd_hw
savemat(result_file_name, result)
print '\n\tSave Intermediate Result Successfully'
print '\t%d-th learning : #Data=%d/%d, #rfs=%d/%d, error=%f\n' %(train_cnt, LWPR_model_left.n_data, LWPR_model_right.n_data, LWPR_model_left.num_rfs, LWPR_model_right.num_rfs, error_hist[train_cnt][1])
print 'end training'
return best_model_epoch
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--environment", type=str, default='AntBulletEnv-v0')
parser.add_argument("--no_data_start", type=int, default=10000)
parser.add_argument("--train_policy_batch_size", type=int, default=30)
parser.add_argument("--cma_maxiter", type=int, default=1000)
parser.add_argument("--unroll_steps", type=int, default=200)
args = parser.parse_args()
print args
env = gym.make(args.environment)
input_dim = env.observation_space.shape[0]+env.action_space.shape[0]
output_dim = env.observation_space.shape[0] + 1
model = LWPR(input_dim, output_dim)
model.init_D = 1. * np.eye(input_dim)
model.update_D = True
model.init_alpha = 20. * np.eye(input_dim)
model.meta = True
agent = AGENT(env.observation_space.shape[0],
env.action_space.shape[0],
action_space_low=env.action_space.low,
action_space_high=env.action_space.high,
unroll_steps=args.unroll_steps)
init_states = np.stack([env.reset() for _ in range(args.train_policy_batch_size)], axis=0)
#Train the dynamics model the intial data.
data_buffer = gather_data3(env, args.no_data_start)
states, actions, rewards, next_states, _ = zip(*data_buffer)
states = np.stack(states, axis=0)
actions = np.stack(actions, axis=0)
rewards = np.array(rewards)[..., np.newaxis]
next_states = np.stack(next_states, axis=0)
state_actions = np.concatenate([states, actions], axis=-1)
state_diff = next_states - states
targets = np.concatenate([state_diff, rewards], axis=-1)
assert len(state_actions) == len(targets)
ind = np.random.permutation(len(state_actions))
for i in range(len(state_actions)):
model.update(state_actions[ind[i]], targets[ind[i]])
for epoch in range(1000):
agent._fit(model, init_states, args.cma_maxiter)
total_rewards = 0.
state = env.reset()
while True:
action = agent._forward(agent.thetas, state[np.newaxis, ...])[0]
next_state, reward, done, _ = env.step(action)
state_action = np.concatenate([state, action])
state_diff = next_state - state
target = np.append(state_diff, reward)
model.update(state_action, target)
total_rewards += float(reward)
state = next_state.copy()
if done:
print 'epoch:', epoch, 'total_rewards:', total_rewards
break
# initialize lwpr model
model = LWPR(n_in, n_out)
model.init_D = 10*eye(n_in)
model.init_alpha = 0.1* eye(n_in)
# model.kernel = 'BiSquare'
for i in range(10):
for demonstration in demonstrations:
output = np.asarray(demonstration[0])
context = np.asarray(demonstration[1])
# print("added output: " + str(output))
# print("added context: " + str(context))
model.update(context, output)
# generalize
# y = [-1.0, 0.0, 1.0]
y = [-1.0, -0.5, 0.0, 0.5, 1.0]
for y1 in y:
plt.figure()
for y2 in y:
context = np.asarray([y1, y2])
output, conf = model.predict_conf(context)
# print("predicted output: " + str(output))
plt.title("Predicted trajectory")
plt.ylim([y_min, y_max])
