def advance(self):
#if it's taking up the whole camera it's definately hit or just about to
#not sure what the whole camera is though.
if(self.lastKnownBuoyLoc.width < 500):
forward(.3)
else:
forward(0.0)
def main():
args = docopt(__doc__)
use_duration = True if args["-D"] else False
window = int(args["-w"])
models = model.extract_models(model.ZIPPED_MODELS)
if args["-B"]:
models = classification.binarize(models)
for line in sys.stdin:
splitted = line.strip().split("\t")
if len(splitted) == 5:
mid_terminal, uid, bid, aid, histories = splitted
term = classification.terminal_name(mid_terminal)
paths = stream.decode(histories)
assert len(paths) == 1
decision = classification.query(term, paths[0], models, use_duration=use_duration, window=window)
if decision:
forward(
mid_terminal,
uid,
bid,
aid,
decision.decision,
decision.prob,
decision.path,
decision.probs,
decision.principal,
)
def execute(self, userdata):
forward(0.5)
for i in range(1000):
if self.preempt_requested():
self.service_preempt()
return 'preempted'
rospy.sleep(1)
return 'preempted'
def dump(key, model):
split = lambda l, n: (l[x: x + n] for x in xrange(0, len(l), n))
splitted = split(encode(model), MAX_SIZE)
for i, code in enumerate(splitted):
keys = key.split('-')
if len(keys) > 1:
forward('-'.join(keys[0:-1]), keys[-1], i, code)
else:
forward(key, 'NULL', i, code)
def execute(self, userdata):
move('Depth', 'Command', 2.0)
for i in range(4):
if self.preempt_requested():
self.service_preempt()
return 'preempted'
rospy.sleep(1)
forward(.75)
for i in range(self.GO_FORWARD_TIMEOUT):
if self.preempt_requested():
self.service_preempt()
return 'preempted'
rospy.sleep(1)
return 'succeeded'
def train_save(self):
train_data, train_target, test_data, test_target = self.get_train_test_data(
)
print(len(train_data), len(train_target))
print(len(test_data), len(test_target))
# train_data = np.array(train_data, dtype = 'float')
# test_data = np.array(test_data, dtype = 'float')
self.cl.fit(train_data, train_target)
predict = self.cl.predict(test_data)
utils.forward(self)
Loader.save_model(self.cl, self.model_path, "svm")
print(metrics.accuracy_score(test_target, predict))
def main():
X, T = get_facialexpression(balance_ones=True)
# X, T = np.shuffle(X,T)
label_map = [
'Anger', 'Disgust', 'Fear', 'Happy', 'Sad', 'Surprise', 'Neutral'
]
# klass =3 error_rate=0.0
# klass =4 error_rate=0.0
# klass =5 error_rate=0.0
# klass =0
klass = 4
N, D = X.shape
X = np.concatenate(
(np.ones((N, 1)), X),
axis=1,
)
T = T.astype(np.int32)
X = X.astype(np.float32)
#Fix for forecasting on one image
T = class1detect(T, detect=klass)
D += 1
# params
lr = 5e-7
max_iteration = 150
W = np.random.randn(D) / np.sqrt(D)
cost = []
error = []
for i in xrange(max_iteration):
Y = forward(W, X)
cost.append(cross_entropy(T, Y))
error.append(error_rate(T, Y))
W += lr * X.T.dot(T - Y)
if i % 5 == 0:
print "i=%d\tcost=%.3f\terror=%.3f" % (i, cost[-1], error[-1])
if i % 5 == 0:
print "i=%d\tcost=%.3f\terror=%.3f" % (i, cost[-1], error[-1])
print "Final weight:", W
print T
print np.round(Y)
plt.title('logistic regression ' + label_map[klass])
plt.xlabel('iterations')
plt.ylabel('cross entropy')
plt.plot(cost)
plt.show()
plt.title('logistic regression ' + label_map[klass])
plt.xlabel('iterations')
plt.ylabel('error rate')
plt.plot(error)
plt.show()
def from_pretrained(cls, path, output_hidden_states=False):
if os.path.exists(path):
parameters = utils.read_yaml(path)
else:
raise ValueError("{} doesn't exists.".format(path))
model = cls(**parameters)
model.load_state_dict(torch.load(parameters['weights_path']))
if output_hidden_states:
model.rnn.forward = lambda input, hidden: utils.forward(model.rnn, input, hidden, model.param)
return model
def execute(self, userdata):
self.pub = rospy.Publisher('/Module_Enable', ModuleEnableMsg)
self.sub = rospy.Subscriber('/Task_Completion', String, self.task_complete)
self.objSub = rospy.Subscriber('img_rec/paths', ImgRecObject, self.objCallback)
msg = ModuleEnableMsg()
msg.Module = 'NewPathTask'
msg.State = True
self.pub.publish(msg)
dive(4)
while self.timeout>0:
if self.is_complete:
self.disable_task()
return 'succeeded'
if self.preempt_requested():
self.disable_task()
self.service_preempt()
return 'preempted'
forward(0.35)
rospy.sleep(1)
self.timeout-=1
self.disable_task()
return 'timeout'
def train_save(self):
train_data, train_target, test_data, test_target = self.get_train_test_data(
)
print(len(train_data), len(train_target))
print(len(test_data), len(test_target))
train_data = np.array(train_data)
test_data = np.array(test_data)
print(len(train_data))
best = -999999
for i in range(10):
self.model.fit(train_data,
train_target,
batch_size=20,
nb_epoch=1,
shuffle=True)
score = self.model.evaluate(test_data, test_target, batch_size=20)
print(score)
if score[1] > best:
print('forward!!')
utils.forward(self)
best = score[1]
Loader.save_model(self.model, self.model_path, "lstm",
self.params_path)
def forward(self, x):
'''
'''
N = x.shape[0]
y = np.zeros((N, self.dim_out))
self.fx = np.zeros((N, self.hout * self.wout, self.dim_k))
for i in range(N):
self.fx[i, ] = utils.flatten(x[i, :].reshape(self.shape_in),
self.shape_in, self.shape_k, self.pad,
self.stride, self.indice).reshape(
self.hout * self.wout, -1)
y[i, ] = utils.forward(self.fx[i, ], self.w, self.b).T.ravel()
return y
def execute(self, userdata):
self.reset()
sub = rospy.Subscriber('img_rec/paths', ImgRecObject, self.found_path_cb)
dive(self._target_depth)
for i in range(self._dive_time):
if self.preempt_requested():
self.service_preempt()
sub.unregister()
return 'preempted'
rospy.sleep(1)
forward(.3)
for i in range(self._timeout):
if self.found_path:
sub.unregister()
#Start slowing down
forward(-.1)
rospy.Timer(rospy.Duration(self._slow_down_duration), self.slow_down_stop, True)
#done slowing down.
sub.unregister()
return 'found_path'
rospy.sleep(1)
sub.unregister()
self.reset()
return 'timed_out'
def main():
args = docopt(__doc__)
for line in sys.stdin:
term, histories = history.build(line)
if term and histories:
assert len(histories) == 1
if args['train']:
mergers = int(args['-m'])
reducers = int(args['-r'])
key = term
n = int(ceil(pow(reducers, 1.0 / mergers)))
for i in xrange(mergers):
key += '-%d' % randint(0, n)
forward(key, stream.encode(histories))
elif args['classify']:
fields = history.extract_fields(line)
forward(fields.mid_terminal, fields.uid, fields.bid, fields.aid,
stream.encode(histories))
def discriminator(self):
disc = [
ConvLayer(num_filters=64,
kernel_size=4,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
activation=leakyrelu),
ConvLayer(num_filters=128,
kernel_size=4,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=slim.batch_norm,
activation=leakyrelu),
ConvLayer(num_filters=256,
kernel_size=4,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=slim.batch_norm,
activation=leakyrelu),
ConvLayer(num_filters=512,
kernel_size=4,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=slim.batch_norm,
activation=leakyrelu),
ConvLayer(num_filters=1,
kernel_size=4,
stride=1,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=slim.batch_norm,
activation=leakyrelu)
]
return forward(disc)
def main():
user_action=3
X, T = get_ecommerce(user_action=user_action)
# X, T = np.shuffle(X,T)
N, D = X.shape
X = np.concatenate((np.ones((N,1)), X), axis=1, )
T = T.astype(np.int32)
X = X.astype(np.float32)
D+=1
# params
lr = 5e-4
max_iteration=1000
W = np.random.randn(D) / np.sqrt(D)
cost = []
error = []
for i in xrange(max_iteration):
Y = forward(W, X)
cost.append(cross_entropy(T,Y))
error.append(error_rate(T,Y))
W += lr*X.T.dot(T-Y)
if i % 5 == 0:
print "i=%d\tcost=%.3f\terror=%.3f" % (i,cost[-1],error[-1])
if i % 5 == 0:
print "i=%d\tcost=%.3f\terror=%.3f" % (i,cost[-1],error[-1])
print "Final weight:", W
plt.title('logistic regression user_action=%d' % (user_action))
plt.xlabel('iterations')
plt.ylabel('cross entropy')
plt.plot(cost)
plt.show()
plt.title('logistic regression user_action=%d' % (user_action))
plt.xlabel('iterations')
plt.ylabel('error rate')
plt.plot(error)
plt.show()
def __call__(self, x):
layers = [
PaddingLayer(self.padding_size, self.padding_type),
ConvLayer(num_filters=self.num_filters,
kernel_size=self.kernel_size,
stride=self.stride,
padding="VALID",
weights_init=self.weights_init,
normalizer=self.normalizer,
activation=self.activation),
PaddingLayer(self.padding_size, self.padding_type),
ConvLayer(num_filters=self.num_filters,
kernel_size=self.kernel_size,
stride=self.stride,
padding="VALID",
weights_init=self.weights_init,
normalizer=self.normalizer,
activation=None)
]
res = forward(layers)(x) + x
return res
def forward(self, x):
self.x = x
return utils.forward(x, self.w, self.b)
def main():
dset = dataset(dconf)
# dset.set_batch_size([25, 25])
input_dim = dconf['input_dim']
output_dim = dconf['output_dim']
lr = 1e-1
# a random labeled example
xl = tf.placeholder(tf.float32, input_dim)
# a random neighbor
xn = tf.placeholder(tf.float32, input_dim)
# a random unlabeled example
xu = tf.placeholder(tf.float32, input_dim)
yl = forward(dconf['netName_label'], xl)
yn = forward(dconf['netName_neighbor'], xn)
yu = forward(dconf['netName_unlabel'], xu)
yt = tf.placeholder(tf.float32, output_dim)
loss_ = supervised_loss('abs_quadratic')(yl, yt)
if not dconf['supervise only']:
# add loss based on manifold neighbor
loss_ += graph_loss('LE')(yl, yn, 1.0)
loss_ += graph_loss('LE')(yl, yu, 0.0)
# add loss based on cluster
loss_ += cluster_loss('S3VM')(yn)
opt = tf.train.AdagradOptimizer(lr)
gvs = opt.compute_gradients(loss_)
clipped_gvs = [(tf.clip_by_value(grad, -1.0, 1.0), var)
for grad, var in gvs]
train_step = opt.apply_gradients(clipped_gvs)
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
# Train
# actually this is the number of labeled samples used in training instead of epoch
# because we cannot perform batch computation here
nb_epoch = 100000
for i in tqdm(range(nb_epoch)):
xl_, xn_, xu_, yt_ = dset.pick()
sess.run(train_step, feed_dict={xl: xl_, xn: xn_, xu: xu_, yt: yt_})
# Test trained model
with sess.as_default():
if dconf['multi-label']:
yl_ = sess.run(yl, feed_dict={xl: dset.test['x']})
yl_ = np.argmax(np.array(yl_), axis=1)
yt_ = dset.test['y']
for metric in (int_accuracy, ):
m_name, m_value = metric(yl_, yt_)
print(m_name, '%.6f' % m_value)
for i in range(len(yt_)):
print((yl_[i], yt_[i]), end='')
else:
for keras_metric in (int_accuracy, ):
print(
keras_metric(yt, yl).eval(feed_dict={
xl: dset.test['x'],
yt: dset.test['y']
}))
yp = sess.run(yl,
feed_dict={
xl: dset.test['x'],
yt: dset.test['y']
})
for i in range(20):
print(yp[i], dset.test['y'][i])
def update(doc, parent, user, version, diff):
if not diff:
return
models.lock_acquire(doc)
now = time.time()
patch = models.get(doc)
if not patch:
patch = []
pre = []
version_count = models.get('version:%s:%s' % (user, doc))
if not version_count:
version_count = 0
version_max = models.get('versionmax:%s:%s' % (user, doc))
if not version_max:
version_max = 0
version_time = models.get('versiontime:%s:%s' % (user, doc))
if not version_time:
version_time = 0
same = []
if parent != version_time:
models.set('version:%s:%s' % (user, doc), 1, now + 60)
models.set('versionmax:%s:%s' % (user, doc), version, now + 60)
models.set('versiontime:%s:%s' % (user, doc), parent, now + 60)
if version == 1:
same = [(version, diff)]
else:
models.set('versions:%s:%s' % (user, doc), [(version, diff)],
now + 60)
else:
same = models.get('versions:%s:%s' % (user, doc))
if not same:
same = []
version_count += 1
models.set('version:%s:%s' % (user, doc), version_count, now + 60)
if version > version_max:
version_max = version
models.set('versionmax:%s:%s' % (user, doc), version_max, now + 60)
if version_count == version_max:
same.append((version, diff))
models.delete('versions:%s:%s' % (user, doc))
else:
models.set('versions:%s:%s' % (user, doc),
same + [(version, diff)], now + 60)
same = []
if not same:
models.lock_release(doc)
return
same = sorted(same)
version = same[0][0]
for i in reversed(patch):
if i['timestamp'] == parent or (i['user'] == user
and i['version'] + 1 == version):
break
pre = i['diff'] + pre
diff = []
for i in same:
diff += utils.forward(pre, i[1])
version = same[-1][0]
ret = {
'parent': parent,
'timestamp': now,
'user': user,
'version': version,
'diff': diff
}
models.set(doc,
filter(lambda x: x['timestamp'] >= now - 60, patch) + [ret])
models.set('last:%s' % doc, now)
text = models.get('doc:%s' % doc)
if text:
text = text[1]
else:
text = ''
text = utils.text_patch(text, diff)
models.set('doc:%s' % doc, (now, text))
models.lock_release(doc)
def objCallback(self, msg):
self.foundObj=True
self.extendTimeout()
forward(0)
self.centerOnPath(msg)
import pygame
import pickle
import numpy as np
from game import init, iterate
from ann import NeuralNetwork
import utils
# Architecture (Specify archetecture here.)
network = NeuralNetwork(layers=[7, 14, 14, 7, 1],
activations=['sigmoid', 'sigmoid', 'sigmoid', 'tanh'])
lr = 0.1
losses = []
screen, font = init()
# Game Loop / Train Loop
frame_count, score, _, _, x = iterate.iterate(screen, font, 0, 0)
game = True
run = True
prediction = 0
while run:
for event in pygame.event.get():
if event.type == pygame.QUIT:
run = False
prediction = utils.forward(x, network)
frame_count, score, game, run, x = iterate.iterate(screen, font,
frame_count, score,
game, run, prediction)
loss = utils.backward(prediction, x, lr, network)
losses.append(loss)
pygame.quit()
def update(doc, parent, user, version, diff):
if not diff:
return
models.lock_acquire(doc)
now = time.time()
patch = models.get(doc)
if not patch:
patch = []
pre = []
version_count = models.get('version:%s:%s' % (user, doc))
if not version_count:
version_count = 0
version_max = models.get('versionmax:%s:%s' % (user, doc))
if not version_max:
version_max = 0
version_time = models.get('versiontime:%s:%s' % (user, doc))
if not version_time:
version_time = 0
same = []
if parent != version_time:
models.set('version:%s:%s' % (user, doc), 1, now+60)
models.set('versionmax:%s:%s' % (user, doc), version, now+60)
models.set('versiontime:%s:%s' % (user, doc), parent, now+60)
if version == 1:
same = [(version, diff)]
else:
models.set('versions:%s:%s' % (user, doc), [(version, diff)], now+60)
else:
same = models.get('versions:%s:%s' % (user, doc))
if not same:
same = []
version_count += 1
models.set('version:%s:%s' % (user, doc), version_count, now+60)
if version > version_max:
version_max = version
models.set('versionmax:%s:%s' % (user, doc), version_max, now+60)
if version_count == version_max:
same.append((version, diff))
models.delete('versions:%s:%s' % (user, doc))
else:
models.set('versions:%s:%s' % (user, doc), same+[(version, diff)], now+60)
same = []
if not same:
models.lock_release(doc)
return
same = sorted(same)
version = same[0][0]
for i in reversed(patch):
if i['timestamp'] == parent or (i['user'] == user and i['version']+1 == version):
break
pre = i['diff']+pre
diff = []
for i in same:
diff += utils.forward(pre, i[1])
version = same[-1][0]
ret = {'parent': parent, 'timestamp': now, 'user': user, 'version': version, 'diff': diff}
models.set(doc, filter(lambda x:x['timestamp']>=now-60, patch)+[ret])
models.set('last:%s' % doc, now)
text = models.get('doc:%s' % doc)
if text:
text = text[1]
else:
text = ''
text = utils.text_patch(text, diff)
models.set('doc:%s' % doc, (now, text))
models.lock_release(doc)
'''
if MODE == "federated":
client_headers["step"] = str(global_train_step)
else:
client_headers["step"] = str(cur_epoch)
st = time.time()
xs = x_train[i:i + BATCH_SIZE]
ys = y_train[i:i + BATCH_SIZE]
# Logits == predictions
client_grads, loss, preds = forward(
inputs=(xs, ys),
model=model,
loss_fn=tf.nn.sparse_softmax_cross_entropy_with_logits,
training=True,
)
'''
If federated, then get synchronized gradient updates from server
If weightavg, then apply normal gradients. Federation occurs at end of epoch.
If cyclic, then apply normal gradients. Federation occurs at end of epoch.
If local, then apply normal gradients and train as usual
'''
if MODE == "federated":
grads = federate_vals(URL, client_grads, client_headers)
def centerY(self, msg):
if abs(msg.center_y) > 30:
forward(-(msg.center_y/400))
return False
forward(0)
return True
else:
client_headers["step"] = str(cur_epoch)
st = time.time()
xs, ys = get_training_pair(
ct_vols_train,
mask_vols_train,
patch_indices_train,
batch_size=BATCH_SIZE,
)
# Logits == predictions
client_grads, loss, preds = forward(
inputs=(xs, ys),
model=model,
loss_fn=soft_dice_loss,
training=True,
)
'''
If federated, then get synchronized gradient updates from server
If weightavg, then apply normal gradients. Federation occurs at end of epoch.
If cyclic, then apply normal gradients. Federation occurs at end of epoch.
If local, then apply normal gradients and train as usual
'''
if MODE == "federated":
grads = federate_vals(URL, client_grads, client_headers)
import utils
import numpy as np
layer = [utils.get_layer(3, 8, activation='ReLU', optimizer='Adamoptimizer', regularization=0.1), utils.get_layer(8, 1, activation='linear',optimizer='Adamoptimizer',regularization=0.1)]
batch_x = [[0.5,0.5,0.5]]
batch_y = [[0.5]]
for i in range(10):
layer, total_loss = utils.train(batch_x, batch_y, layer)
print(total_loss)
#print(layer)
utils.save_model(layer, 'model_test.npy')
utils.load_model('model_test.npy')
utils.forward(batch_x, layer)
print(utils.last_layer(layer)['out_'])
c = 's'
while True:
if isData():
c = sys.stdin.read(1)
print(c)
if c == 's':
stop_state = True
elif c == 'h':
happy_state = not happy_state
c = ''
else: #c == '' or else.
if c == 'w':
forward(pwm, pwm)
stop_state = False
elif c == 'x':
backward(pwm, pwm)
stop_state = False
elif c == 'a':
spin_left(pwm, pwm)
stop_state = False
elif c == 'd':
spin_right(pwm, pwm)
stop_state = False
elif c == 'q':
pwm = pwm + pwm_increment if pwm <= max_speed - pwm_increment else pwm
print("pwm: ", pwm)
elif c == 'e':
def predict(X,theta):
X = np.array(X)
X.reshape((1,7))
return forward(X,theta)
y_true = []
cur_step = 1
iterator = iter(test_dataset)
elapsed_time = 0.0
while True:
try:
batch_start_time = time.time()
data = next(iterator)
except StopIteration:
break
else:
xs, ys = data
loss, logits = forward(
inputs=data,
model=model,
loss_fn=tf.nn.softmax_cross_entropy_with_logits,
training=False,
)
preds = tf.nn.softmax(logits)
test_loss.update_state(loss)
test_accuracy.update_state(
tf.argmax(ys, axis=1),
tf.argmax(preds, axis=1),
)
y_true.extend(tf.argmax(ys, axis=1))
y_pred.extend(tf.argmax(preds, axis=1))
batch_end_time = time.time()
def main():
parser = argparse.ArgumentParser()
# Training
parser.add_argument('--adv', default=False, action='store_true')
parser.add_argument('--easy', default=False, action='store_true')
parser.add_argument('--num_epochs', '-N', type=int, default=100)
parser.add_argument('--batch_size', '-B', type=int, default=128)
parser.add_argument('--gpu_id', '-G', type=int, default=-1)
parser.add_argument('--seed', type=int, default=820)
# Log
parser.add_argument('--results_dir', '-R', type=str, default='results')
# Dataset -- X
parser.add_argument('--X_std', type=float, default=0.04)
parser.add_argument('--X_trn_sz', type=int, default=512 * 4)
parser.add_argument('--X_val_sz', type=int, default=512 * 2)
# Dataset -- Z
parser.add_argument('--Z_std', type=float, default=1.0)
parser.add_argument('--Z_trn_sz', type=int, default=512 * 4)
parser.add_argument('--Z_val_sz', type=int, default=512 * 2)
# Optimizer
parser.add_argument('--n_gen', type=int, default=5)
parser.add_argument('--n_dis', type=int, default=1)
parser.add_argument('--alpha', type=float, default=1e-3)
parser.add_argument('--beta1', type=float, default=0.5)
parser.add_argument('--beta2', type=float, default=0.999)
# Models
parser.add_argument('--in_features', '-in', type=int, default=2)
parser.add_argument('--latent_features', '-lf', type=int, default=2)
parser.add_argument('--noise_features', '-nf', type=int, default=2)
parser.add_argument('--gen_num_layers', '-gnl', type=int, default=2)
parser.add_argument('--gen_hidden_features', '-ghf', type=int, default=256)
parser.add_argument('--gen_out_features', '-gof', default=None)
parser.add_argument('--inf_num_layers', '-inl', type=int, default=2)
parser.add_argument('--inf_hidden_features', '-ihf', type=int, default=256)
parser.add_argument('--inf_out_features', '-iof', default=None)
parser.add_argument('--dis_num_layers', '-dnl', type=int, default=2)
parser.add_argument('--dis_hidden_features', '-dhf', type=int, default=256)
parser.add_argument('--dis_out_features', '-dof', type=int, default=1)
args = parser.parse_args()
if args.gen_out_features is None:
args.gen_out_features = args.in_features
if args.inf_out_features is None:
args.inf_out_features = args.latent_features
if args.adv:
_name = "_ALICE_toydata_unsupervised_adversarial_reconstruction"
else:
_name = "_ALICE_toydata_unsupervised_MSE_reconstruction"
args.results_dir = os.path.join(
args.results_dir,
datetime.datetime.now().strftime('%y%m%d-%H%M%S') + _name)
if not os.path.isdir(args.results_dir):
os.makedirs(args.results_dir)
with open(os.path.join(args.results_dir, 'args.json'), 'w') as f:
json.dump(args.__dict__, f, indent=4)
if args.gpu_id > -1:
if torch.cuda.is_available():
torch.cuda.set_device(args.gpu_id)
torch.cuda.manual_seed_all(args.seed)
device = torch.device(
'cuda' if args.gpu_id > -1 and torch.cuda.is_available() else 'cpu')
# Prepare dataset X
rt2 = math.sqrt(2)
means = {
'easy': [[0, 0], [5, 5], [-5, 5], [-5, -5], [5, -5]],
'difficult': [[5 * rt2, 0], [5, 5], [0, 5 * rt2], [-5, 5],
[-5 * rt2, 0], [-5, -5], [0, -5 * rt2], [5, -5]]
}
key = 'easy' if args.easy else 'difficult'
means_x = list(
map(lambda x: torch.tensor(x, dtype=torch.float), means[key]))
variances_x = [torch.eye(2) * args.X_std for _ in means_x]
x_trn = gmm_module.GMMData(args.X_trn_sz,
means_x,
variances_x,
seed=args.seed)
x_trn_loader = torch.utils.data.DataLoader(
x_trn,
args.batch_size,
pin_memory=(args.gpu_id > -1 and torch.cuda.is_available()))
# Prepare dataset Z
means_z = list(map(lambda x: torch.tensor(x, dtype=torch.float), [[0, 0]]))
variances_z = [torch.eye(2) * args.Z_std for _ in means_z]
z_trn = gmm_module.GMMData(args.Z_trn_sz,
means_z,
variances_z,
seed=args.seed)
z_trn_loader = torch.utils.data.DataLoader(
z_trn,
args.batch_size,
pin_memory=(args.gpu_id > -1 and torch.cuda.is_available()))
# Prepare models
gen = Generator(args.gen_num_layers, args.in_features, args.noise_features,
args.gen_hidden_features, args.gen_out_features).to(device)
inf = Inference(args.gen_num_layers, args.in_features, args.noise_features,
args.inf_hidden_features, args.inf_out_features).to(device)
dis = Discriminator(args.gen_num_layers, args.in_features,
args.noise_features, args.dis_hidden_features,
args.dis_out_features).to(device)
if args.adv:
dis_x = DiscriminatorXX(args.gen_num_layers, args.in_features,
args.noise_features, args.dis_hidden_features,
args.dis_out_features).to(device)
dis_z = DiscriminatorZZ(args.gen_num_layers, args.in_features,
args.noise_features, args.dis_hidden_features,
args.dis_out_features).to(device)
else:
dis_x, dis_z = None, None
opt_gen_inf = torch.optim.Adam(
list(gen.parameters()) + list(inf.parameters()), args.alpha,
(args.beta1, args.beta2))
_params = list(dis.parameters())
if args.adv:
_params += list(dis_x.parameters()) + list(dis_z.parameters())
opt_dis = torch.optim.Adam(_params, args.alpha, (args.beta1, args.beta2))
# Save figures
x_gmm_samples = x_trn.samples.numpy()
z_gmm_samples = z_trn.samples.numpy()
figure, ax = plt.subplots(1, 1, figsize=(4.5, 4.5))
ax.scatter(x_gmm_samples[:, 0],
x_gmm_samples[:, 1],
label='X',
marker='.',
alpha=0.3,
c=matplotlib.cm.Set1(x_trn.labels.numpy().reshape(
(-1, )) / args.in_features / 2.0))
ax.scatter(z_gmm_samples[:, 0],
z_gmm_samples[:, 1],
label='Z',
marker='.',
alpha=0.1)
ax.set_xlim(-10, 10)
ax.set_ylim(-10, 10)
plt.legend()
plt.savefig(os.path.join(args.results_dir, 'dataset.png'))
plt.close('all')
torch.save(x_trn, os.path.join(args.results_dir, 'x_trn.pkl'))
torch.save(z_trn, os.path.join(args.results_dir, 'z_trn.pkl'))
for epoch in tqdm.tqdm(range(args.num_epochs)):
_mom, epoch_dis_loss, epoch_gen_loss = .9, 0, 0
for i, (x_batch, z_batch) in tqdm.tqdm(
enumerate(zip(x_trn_loader, z_trn_loader))):
x, *_ = x_batch
z, *_ = z_batch
x, z = x.to(device), z.to(device)
iter_dis_loss = .0
for j in range(args.n_dis):
dis_loss_opt, _ = forward(device, args.adv, x, z, gen, inf,
dis, dis_x, dis_z, False)
opt_dis.zero_grad()
dis_loss_opt.backward()
opt_dis.step()
iter_dis_loss += dis_loss_opt.item() / args.n_dis
epoch_dis_loss = epoch_dis_loss * (1 - _mom) + iter_dis_loss * _mom
iter_gen_loss, iter_x, iter_z = .0, .0, .0
for j in range(args.n_gen):
_, gen_loss_opt, cost_x, cost_z = forward(
device, args.adv, x, z, gen, inf, dis, dis_x, dis_z, True)
opt_gen_inf.zero_grad()
gen_loss_opt.backward()
opt_gen_inf.step()
iter_gen_loss += gen_loss_opt.item() / args.n_gen
iter_x += cost_x.item() / args.n_gen
iter_z += cost_z.item() / args.n_gen
epoch_gen_loss = epoch_gen_loss * (1 - _mom) + iter_gen_loss * _mom
if (i + 1) % 8 == 0:
_fmt = "epoch {}/{}, iter {}, dis: {:.05f}, gen: {:.05f}, x: {:.05f}, z: {:.05f}"
tqdm.tqdm.write(
_fmt.format(epoch, args.num_epochs, i + 1, iter_dis_loss,
iter_gen_loss, iter_x, iter_z))
if (epoch + 1) % 10 == 0:
gen.eval()
inf.eval()
x_trn_samples = x_trn.samples.to(device)
z_trn_samples = z_trn.samples.to(device)
with torch.no_grad():
p_x = gen(z_trn_samples)
q_z = inf(x_trn_samples)
x_rec = gen(q_z).cpu().numpy()
z_rec = inf(p_x).cpu().numpy()
tqdm.tqdm.write('Epoch {}/{}, z_rec| mean: {}, var: {}'.format(
epoch + 1, args.num_epochs, z_rec.mean(axis=0),
z_rec.var(axis=0, ddof=1)))
if (epoch + 1) % 10 == 0:
figure, ax = plt.subplots(1, 1, figsize=(4.5, 4.5))
ax.scatter(x_rec[:, 0],
x_rec[:, 1],
label='X_reconstructed',
marker='.',
alpha=0.3)
ax.scatter(z_rec[:, 0],
z_rec[:, 1],
label='Z_reconstructed',
marker='.',
alpha=0.1)
ax.set_xlim(-10, 10)
ax.set_ylim(-10, 10)
plt.legend()
plt.savefig(
os.path.join(args.results_dir,
'reconstructed_{}.png'.format(epoch + 1)))
plt.close('all')
_ckpt = {
'opt_gen_inf': opt_gen_inf.state_dict(),
'opt_dis': opt_dis.state_dict(),
'gen': gen.state_dict(),
'inf': inf.state_dict(),
'dis': dis.state_dict(),
}
if args.adv:
_ckpt.update({
'dis_x': dis_x.state_dict(),
'dis_z': dis_z.state_dict()
})
torch.save(_ckpt, os.path.join(args.results_dir, 'ckpt.pth.tar'))
def test_forward(self):
x = np.array([[1, 2, 3], [4, 5, 6]])
w = np.array([[1, 0, -1], [0, 1, 1], [1, -1, 0]])
b = np.array([[1, 3, 2]])
y = utils.forward(x, w, b)
self.assertTrue(np.allclose(y, [[5, 2, 3], [11, 2, 3]]))
def execute(self, userdata):
turn(0)
dive(0)
forward(0)
rospy.loginfo("motors should be reset now...")
return super(Idle, self).execute(userdata)
def generator(self):
f = 7
p = (f - 1) / 2
gen = [
PaddingLayer(p, "reflect"),
ConvLayer(num_filters=32,
kernel_size=7,
stride=1,
padding="VALID",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=self.normalizer,
activation=tf.nn.relu),
ConvLayer(num_filters=64,
kernel_size=3,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=self.normalizer,
activation=tf.nn.relu),
ConvLayer(num_filters=128,
kernel_size=3,
stride=2,
padding="SAME",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=self.normalizer,
activation=tf.nn.relu)
]
for i in range(self.n_res_blocks):
gen.append(
ResBlock(num_filters=128,
kernel_size=3,
stride=1,
padding_size=1,
padding_type="reflect",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=self.normalizer,
activation=tf.nn.relu))
gen2 = [
ConvTransposeLayer(num_outputs=64,
kernel_size=3,
stride=2,
padding="SAME",
normalizer=self.normalizer,
activation=tf.nn.relu),
ConvTransposeLayer(num_outputs=32,
kernel_size=3,
stride=2,
padding="SAME",
normalizer=self.normalizer,
activation=tf.nn.relu),
PaddingLayer(p, "reflect"),
ConvLayer(num_filters=3,
kernel_size=7,
stride=1,
padding="VALID",
weights_init=tf.truncated_normal_initializer(
stddev=self.weight_stdev),
normalizer=self.normalizer,
activation=tf.tanh)
]
gen = gen + gen2
return forward(gen)
