def fit(self,
train,
test=None,
validate_every=100,
show_norms=False,
show_output=False,
error_logging=True):
""" Fit model
Pass in test to compute test error and report during
training.
train : ndarray (T x n_in)
validation_frequency : int
in terms of number of epochs
"""
if test is not None:
self.interactive = True
test_set = self.shared_dataset(test)
else:
self.interactive = False
train_set = self.shared_dataset(train)
# compute number of minibatches for training
# note that cases are the second dimension, not the first
n_train = train_set.get_value(borrow=True).shape[0]
n_train_batches = int(np.ceil(1.0 * n_train / self.batch_size))
if self.interactive:
n_test = test_set.get_value(borrow=True).shape[0]
n_test_batches = int(np.ceil(1.0 * n_test / self.batch_size))
#validate_every is specified in terms of epochs
validation_frequency = validate_every * n_train_batches
######################
# BUILD ACTUAL MODEL #
######################
logger.info('... building the model')
index = T.lscalar('index') # index to a [mini]batch
n_ex = T.lscalar('n_ex') # total number of examples
# learning rate (may change)
l_r = T.scalar('l_r', dtype=theano.config.floatX)
mom = T.scalar('mom', dtype=theano.config.floatX) # momentum
# Proper implementation of variable-batch size evaluation
# Note that classifier.errors() returns the mean error
# But the last batch may be a smaller size
# So we keep around the effective_batch_size (whose last element may
# be smaller than the rest)
# And weight the reported error by the batch_size when we average
# Also, by keeping batch_start and batch_stop as symbolic variables,
# we make the theano function easier to read
batch_start = index * self.batch_size
batch_stop = T.minimum(n_ex, (index + 1) * self.batch_size)
effective_batch_size = batch_stop - batch_start
get_batch_size = theano.function(inputs=[index, n_ex],
outputs=effective_batch_size)
compute_train_error = []
compute_each_train_error = []
compute_test_error = []
train_model = []
for da in self.estimator.dA_layers:
loss = da.loss
updates = da.get_updates(l_r, mom)
f_ctrain = theano.function(
inputs=[index, n_ex],
outputs=loss,
givens={self.x: train_set[batch_start:batch_stop]},
mode=mode)
compute_train_error.append(f_ctrain)
f_cetrain = theano.function(
inputs=[index, n_ex],
outputs=da.each_loss,
givens={self.x: train_set[batch_start:batch_stop]},
mode=mode)
compute_train_error.append(f_ctrain)
compute_each_train_error.append(f_cetrain)
if self.interactive:
f_ctest = theano.function(
inputs=[index, n_ex],
outputs=loss,
givens={self.x: test_set[batch_start:batch_stop]},
mode=mode)
compute_test_error.append(f_ctest)
# compiling a Theano function `train_model` that returns the
# cost, but in the same time updates the parameter of the
# model based on the rules defined in `updates`
f_train = theano.function(
inputs=[index, n_ex, l_r, mom],
outputs=loss,
updates=updates,
givens={self.x: train_set[batch_start:batch_stop]},
mode=mode)
train_model.append(f_train)
###############
# TRAIN MODEL #
###############
keyMonitoringThread = Threads.KeyMonitoringThread()
keyMonitoringThread.start()
initial_learning_rate = self.learning_rate
t0 = time.time()
for n in xrange(self.estimator.n_layers):
logger.info('... training dA layer[%d]' % n)
epoch = 0
this_train_loss = np.inf
stopFlg = False
t0_l = time.time()
self.learning_rate = initial_learning_rate
self.errorlog.append([])
while (epoch < self.n_epochs) and (
this_train_loss > self.t_error) and (stopFlg is False):
epoch = epoch + 1
effective_momentum = self.final_momentum \
if epoch > self.momentum_switchover \
else self.initial_momentum
for minibatch_idx in xrange(n_train_batches):
minibatch_avg_cost = train_model[n](minibatch_idx, n_train,
self.learning_rate,
effective_momentum)
# iteration number (how many weight updates have we made?)
# epoch is 1-based, index is 0 based
iter = (epoch - 1) * n_train_batches + minibatch_idx + 1
if iter % validation_frequency == 0:
# compute loss on training set
train_losses = [
compute_train_error[n](i, n_train)
for i in xrange(n_train_batches)
]
train_batch_sizes = [
get_batch_size(i, n_train)
for i in xrange(n_train_batches)
]
this_train_loss = np.average(train_losses,
weights=train_batch_sizes)
# compute each output unit loss on training set
if error_logging is True:
train_each_losses = np.array([
compute_each_train_error[n](i, n_train)
for i in xrange(n_train_batches)
])
train_batch_sizes_for_each = []
for i in xrange(train_each_losses.shape[1]):
train_batch_sizes_for_each.append(
train_batch_sizes)
this_train_each_loss = np.average(
train_each_losses.T,
weights=train_batch_sizes_for_each,
axis=1)
el = np.r_[np.array([epoch]), this_train_each_loss]
self.errorlog[n] = np.vstack((self.errorlog[n], el)) \
if self.errorlog[n] != [] \
else np.array([el])
# エラーの推移をpngに保存
self.save_errorlog_png()
# self.save_errorlog_png(fname='dALayer%d' % n)
if self.interactive:
test_losses = [
compute_test_error[n](i, n_test)
for i in xrange(n_test_batches)
]
test_batch_sizes = [
get_batch_size(i, n_test)
for i in xrange(n_test_batches)
]
this_test_loss = np.average(
test_losses, weights=test_batch_sizes)
logger.info('*** dA layer[%d] *** epoch %i, mb %i/%i, tr loss %f '
'te loss %f lr: %f mom: %f'% \
(n, epoch, minibatch_idx + 1, n_train_batches,
this_train_loss, this_test_loss,
self.learning_rate, effective_momentum))
else:
logger.info(
'*** dA layer[%d] *** epoch %i, mb %i/%i, train loss %f'
' lr: %f mom: %f' %
(n, epoch, minibatch_idx + 1, n_train_batches,
this_train_loss, self.learning_rate,
effective_momentum))
self.optional_output(train_set, show_norms,
show_output)
self.learning_rate *= self.learning_rate_decay
# 学習途中のパラメータをスナップショットとして保存
if self.snapshot_every is not None:
if (epoch + 1) % self.snapshot_every == 0:
date_obj = datetime.datetime.now()
date_str = date_obj.strftime('%Y-%m-%d-%H:%M:%S')
class_name = self.__class__.__name__
fname = '%s.%s-snapshot-%d' % (class_name, date_str,
epoch + 1)
self.save(fpath=self.snapshot_path, fname=fname)
# 学習中のコマンド入力を別スレッドで受け取る
var = keyMonitoringThread.GetInput()
# 'q' を受け取った場合、学習を途中で切り上げる
if var == 'q':
stopFlg = True
h, m = divmod(time.time() - t0_l, 3600)
m, s = divmod(m, 60)
print "*** dA layer[%d] *** Elapsed time: %d hour %d min %f sec" % (
n, int(h), int(m), s)
h, m = divmod(time.time() - t0, 3600)
m, s = divmod(m, 60)
print "Elapsed time: %d hour %d min %f sec" % (int(h), int(m), s)
# コマンド入力スレッドの停止
# ('q'入力による学習の途中終了ではなく、終了条件を満たして
# 学習が正常に終了した場合、スレッドを明示的に終了する必要がある)
keyMonitoringThread.Stop()
print 'Press any key...'
def fit(self, X_train, Y_train, X_test=None, Y_test=None,
validate_every=100, optimizer='sgd', compute_zero_one=False,
show_norms=True, show_output=True, error_logging=True):
""" Fit model
Pass in X_test, Y_test to compute test error and report during
training.
X_train : ndarray (T x n_in)
Y_train : ndarray (T x n_out)
validation_frequency : int
in terms of number of epochs
optimizer : string
Optimizer type.
Possible values:
'sgd' : batch stochastic gradient descent
'cg' : nonlinear conjugate gradient algorithm
(scipy.optimize.fmin_cg)
'bfgs' : quasi-Newton method of Broyden, Fletcher, Goldfarb,
and Shanno (scipy.optimize.fmin_bfgs)
'l_bfgs_b' : Limited-memory BFGS (scipy.optimize.fmin_l_bfgs_b)
compute_zero_one : bool
in the case of binary output, compute zero-one error in addition to
cross-entropy error
show_norms : bool
Show L2 norms of individual parameter groups while training.
show_output : bool
Show the model output on first training case while training.
"""
if X_test is not None:
assert(Y_test is not None)
self.interactive = True
test_set_x, test_set_y = self.shared_dataset((X_test, Y_test))
else:
self.interactive = False
train_set_x, train_set_y = self.shared_dataset((X_train, Y_train))
# compute number of minibatches for training
# note that cases are the second dimension, not the first
n_train = train_set_x.get_value(borrow=True).shape[1]
n_train_batches = int(np.ceil(1.0 * n_train / self.batch_size))
if self.interactive:
n_test = test_set_x.get_value(borrow=True).shape[1]
n_test_batches = int(np.ceil(1.0 * n_test / self.batch_size))
#validate_every is specified in terms of epochs
validation_frequency = validate_every * n_train_batches
######################
# BUILD ACTUAL MODEL #
######################
logger.info('... building the model')
index = T.lscalar('index') # index to a [mini]batch
n_ex = T.lscalar('n_ex') # total number of examples
# learning rate (may change)
l_r = T.scalar('l_r', dtype=theano.config.floatX)
mom = T.scalar('mom', dtype=theano.config.floatX) # momentum
cost = self.estimator.loss(self.y) \
+ self.L1_reg * self.estimator.L1 \
+ self.L2_reg * self.estimator.L2_sqr
# Proper implementation of variable-batch size evaluation
# Note that classifier.errors() returns the mean error
# But the last batch may be a smaller size
# So we keep around the effective_batch_size (whose last element may
# be smaller than the rest)
# And weight the reported error by the batch_size when we average
# Also, by keeping batch_start and batch_stop as symbolic variables,
# we make the theano function easier to read
batch_start = index * self.batch_size
batch_stop = T.minimum(n_ex, (index + 1) * self.batch_size)
effective_batch_size = batch_stop - batch_start
get_batch_size = theano.function(inputs=[index, n_ex],
outputs=effective_batch_size)
compute_train_error = theano.function(inputs=[index, n_ex],
outputs=self.estimator.loss(self.y),
givens={self.x: train_set_x[:, batch_start:batch_stop],
self.y: train_set_y[:, batch_start:batch_stop]},
mode=mode)
compute_train_each_error = theano.function(inputs=[index, n_ex],
outputs=self.estimator.each_loss(self.y),
givens={self.x: train_set_x[:, batch_start:batch_stop],
self.y: train_set_y[:, batch_start:batch_stop]},
mode=mode)
if self.interactive:
compute_test_error = theano.function(inputs=[index, n_ex],
outputs=self.estimator.loss(self.y),
givens={self.x: test_set_x[:, batch_start:batch_stop],
self.y: test_set_y[:, batch_start:batch_stop]},
mode=mode)
self.get_norms = {}
for param in self.estimator.params:
self.get_norms[param] = theano.function(inputs=[],
outputs=self.estimator.l2_norms[param], mode=mode)
# compute the gradient of cost with respect to theta using BPTT
gtheta = T.grad(cost, self.estimator.theta)
if optimizer == 'sgd':
updates = {}
theta = self.estimator.theta
theta_update = self.estimator.theta_update
# careful here, update to the shared variable
# cannot depend on an updated other shared variable
# since updates happen in parallel
# so we need to be explicit
upd = mom * theta_update - l_r * gtheta
updates[theta_update] = upd
updates[theta] = theta + upd
# compiling a Theano function `train_model` that returns the
# cost, but in the same time updates the parameter of the
# model based on the rules defined in `updates`
train_model = theano.function(inputs=[index, n_ex, l_r, mom],
outputs=cost,
updates=updates,
givens={self.x: train_set_x[:, batch_start:batch_stop],
self.y: train_set_y[:, batch_start:batch_stop]},
mode=mode)
###############
# TRAIN MODEL #
###############
logger.info('... training')
epoch = 0
this_train_loss = np.inf
stopFlg = False
keyMonitoringThread = Threads.KeyMonitoringThread()
keyMonitoringThread.start()
t0 = time.time()
while (epoch < self.n_epochs) and (this_train_loss > self.t_error) and (stopFlg is False):
epoch = epoch + 1
effective_momentum = self.final_momentum \
if epoch > self.momentum_switchover \
else self.initial_momentum
for minibatch_idx in xrange(n_train_batches):
minibatch_avg_cost = train_model(minibatch_idx, n_train,
self.learning_rate,
effective_momentum)
# iteration number (how many weight updates have we made?)
# epoch is 1-based, index is 0 based
iter = (epoch - 1) * n_train_batches + minibatch_idx + 1
if iter % validation_frequency == 0:
# compute loss on training set
train_losses = [compute_train_error(i, n_train)
for i in xrange(n_train_batches)]
train_batch_sizes = [get_batch_size(i, n_train)
for i in xrange(n_train_batches)]
this_train_loss = np.average(train_losses,
weights=train_batch_sizes)
# compute each output unit loss on training set
if error_logging is True:
train_each_losses = np.array([compute_train_each_error(i, n_train)
for i in xrange(n_train_batches)])
train_batch_sizes_for_each = []
for i in xrange(self.n_out):
train_batch_sizes_for_each.append(train_batch_sizes)
this_train_each_loss = np.average(train_each_losses.T,
weights=train_batch_sizes_for_each, axis=1)
el = np.r_[np.array([epoch]), this_train_each_loss]
self.errorlog = np.vstack((self.errorlog, el)) \
if self.errorlog != [] \
else np.array([el])
# エラーの推移をpngに保存
self.save_errorlog_png()
if self.interactive:
test_losses = [compute_test_error(i, n_test)
for i in xrange(n_test_batches)]
test_batch_sizes = [get_batch_size(i, n_test)
for i in xrange(n_test_batches)]
this_test_loss = np.average(test_losses,
weights=test_batch_sizes)
logger.info('epoch %i, mb %i/%i, tr loss %f '
'te loss %f lr: %f mom: %f'% \
(epoch, minibatch_idx + 1, n_train_batches,
this_train_loss, this_test_loss,
self.learning_rate, effective_momentum))
else:
logger.info('epoch %i, mb %i/%i, train loss %f'
' lr: %f mom: %f' % (epoch,
minibatch_idx + 1,
n_train_batches,
this_train_loss,
self.learning_rate,
effective_momentum))
self.optional_output(train_set_x, show_norms,
show_output)
self.learning_rate *= self.learning_rate_decay
# 学習途中のパラメータをスナップショットとして保存
if self.snapshot_every is not None:
if (epoch + 1) % self.snapshot_every == 0:
date_obj = datetime.datetime.now()
date_str = date_obj.strftime('%Y-%m-%d-%H:%M:%S')
class_name = self.__class__.__name__
fname = '%s.%s-snapshot-%d' % (class_name,
date_str, epoch + 1)
self.save(fpath=self.snapshot_path, fname=fname)
# 学習中のコマンド入力を別スレッドで受け取る
var = keyMonitoringThread.GetInput()
# 'q' を受け取った場合、学習を途中で切り上げる
if var == 'q':
stopFlg = True
keyMonitoringThread.Stop()
h, m = divmod(time.time() - t0, 3600)
m, s = divmod(m, 60)
print "Elapsed time: %d hour %d min %f sec" % (int(h), int(m), s)
# コマンド入力スレッドの停止
# ('q'入力による学習の途中終了ではなく、終了条件を満たして
# 学習が正常に終了した場合、スレッドを明示的に終了する必要がある)
keyMonitoringThread.Stop()
print 'Press any key...'
elif optimizer == 'cg' or optimizer == 'bfgs' or optimizer == 'l_bfgs_b':
# compile a theano function that returns the cost of a minibatch
batch_cost = theano.function(inputs=[index, n_ex],
outputs=cost,
givens={self.x: train_set_x[:, batch_start:batch_stop],
self.y: train_set_y[:, batch_start:batch_stop]},
mode=mode, name="batch_cost")
# compile a theano function that returns the gradient of the
# minibatch with respect to theta
batch_grad = theano.function(inputs=[index, n_ex],
outputs=T.grad(cost, self.estimator.theta),
givens={self.x: train_set_x[:, batch_start:batch_stop],
self.y: train_set_y[:, batch_start:batch_stop]},
mode=mode, name="batch_grad")
# creates a function that computes the average cost on the training
# set
def train_fn(theta_value):
theta_value=np.array(theta_value, dtype=theano.config.floatX)
self.estimator.theta.set_value(theta_value, borrow=True)
train_losses = [batch_cost(i, n_train)
for i in xrange(n_train_batches)]
train_batch_sizes = [get_batch_size(i, n_train)
for i in xrange(n_train_batches)]
return np.average(train_losses, weights=train_batch_sizes)
# creates a function that computes the average gradient of cost
# with respect to theta
def train_fn_grad(theta_value):
theta_value=np.array(theta_value, dtype=theano.config.floatX)
self.estimator.theta.set_value(theta_value, borrow=True)
train_grads = [batch_grad(i, n_train)
for i in xrange(n_train_batches)]
train_batch_sizes = [get_batch_size(i, n_train)
for i in xrange(n_train_batches)]
return np.average(train_grads, weights=train_batch_sizes,
axis=0)
# validation function, prints useful output after each iteration
def callback(theta_value):
self.epoch += 1
if (self.epoch) % validate_every == 0:
theta_value=np.array(theta_value, dtype=theano.config.floatX)
self.estimator.theta.set_value(theta_value, borrow=True)
# compute loss on training set
train_losses = [compute_train_error(i, n_train)
for i in xrange(n_train_batches)]
train_batch_sizes = [get_batch_size(i, n_train)
for i in xrange(n_train_batches)]
this_train_loss = np.average(train_losses,
weights=train_batch_sizes)
# compute each output unit loss on training set
if error_logging is True:
train_each_losses = np.array([compute_train_each_error(i, n_train)
for i in xrange(n_train_batches)])
train_batch_sizes_for_each = []
for i in xrange(self.n_out):
train_batch_sizes_for_each.append(train_batch_sizes)
this_train_each_loss = np.average(train_each_losses.T,
weights=train_batch_sizes_for_each, axis=1)
el = np.r_[np.array([self.epoch]), this_train_each_loss]
self.errorlog = np.vstack((self.errorlog, el)) \
if self.errorlog is not None \
else np.array([el])
# エラーの推移をpngに保存
self.save_errorlog_png(fname=optimizer)
if self.interactive:
test_losses = [compute_test_error(i, n_test)
for i in xrange(n_test_batches)]
test_batch_sizes = [get_batch_size(i, n_test)
for i in xrange(n_test_batches)]
this_test_loss = np.average(test_losses,
weights=test_batch_sizes)
logger.info('epoch %i, tr loss %f, te loss %f' % \
(self.epoch, this_train_loss,
this_test_loss, self.learning_rate))
else:
logger.info('epoch %i, train loss %f ' % \
(self.epoch, this_train_loss))
self.optional_output(train_set_x, show_norms, show_output)
###############
# TRAIN MODEL #
###############
logger.info('... training')
# using scipy conjugate gradient optimizer
import scipy.optimize
if optimizer == 'cg':
of = scipy.optimize.fmin_cg
elif optimizer == 'bfgs':
of = scipy.optimize.fmin_bfgs
elif optimizer == 'l_bfgs_b':
of = scipy.optimize.fmin_l_bfgs_b
logger.info("Optimizing using %s..." % of.__name__)
start_time = time.clock()
# keep track of epochs externally
# these get updated through callback
self.epoch = 0
# interface to l_bfgs_b is different than that of cg, bfgs
# however, this will be changed in scipy 0.11
# unified under scipy.optimize.minimize
if optimizer == 'cg' or optimizer == 'bfgs':
best_theta = of(
f=train_fn,
x0=self.estimator.theta.get_value(),
#x0=np.zeros(self.estimator.theta.get_value().shape,
# dtype=theano.config.floatX),
fprime=train_fn_grad,
callback=callback,
disp=1,
retall=1,
maxiter=self.n_epochs)
elif optimizer == 'l_bfgs_b':
best_theta, f_best_theta, info = of(
func=train_fn,
x0=self.estimator.theta.get_value(),
fprime=train_fn_grad,
iprint=validate_every,
maxfun=self.n_epochs) # max number of feval
end_time = time.clock()
h, m = divmod(end_time - start_time, 3600)
m, s = divmod(m, 60)
print "Optimization time: %d hour %d min %f sec" % (int(h), int(m), s)
else:
raise NotImplementedError
global IS_GET_OBJECT_SIZE
global CONTROL_TIME
# 学習済みネットワークを呼び出す
model = PrepNetwork('NN_D204060_MSPTS_Short_batch100_f') # 14/10/01 BEST!
# model = PrepNetwork('NN_D204060_MSPTS_batch100f_m500_lr9999') # No Good
# model = PrepNetwork('NN_D204060_MSPT_Short_batch100_f')
# Hostと通信を行うインスタンスを用意
com = HostCommunication()
# Hostとの通信を行うスレッドを開始する
com.start()
keyMonitoringThread = Threads.KeyMonitoringThread()
keyMonitoringThread.start()
rtcFlag = False
print "####################################"
print "# Command #"
print "####################################"
print "# 'i': Set Initial Grasping Pose #"
print "# 's': Start Real Time Control #"
print "# 'p': Pause Real Time Control #"
print "# 'q': Quit Program #"
print "####################################"
while True:
# Hostとの通信が切れるまでループ
if com.CompleteFlag == True:
break
