def get_optimization_func(self, target_dim_list, loss_func, optimizer,
clip):
"""Compile and return optimization function.
Args:
target_dim_list: list of integars. targets' dimension. e.g. target_dim_list=[2]
loss_func: string | function.
optimizer: object.
clip: None | real value.
Return:
optimization function.
"""
# set gt nodes
self.set_gt_nodes(target_dim_list)
# Default loss
if type(loss_func) is str:
assert len(
self.out_nodes_
) == 1, "If the number of out_layers > 1, you need define your own loss_func!"
loss_node = obj.get(loss_func)(self.out_nodes_[0],
self.gt_nodes_[0])
# User defined loss
else:
loss_node = loss_func(self)
# Compute gradient
gparams = K.grad(loss_node + self.reg_value_, self.params_)
# Clip gradient
if clip is not None:
gparams = [K.clip(gparam, -clip, clip) for gparam in gparams]
# Gradient based optimization
param_updates = optimizer.get_updates(self.params_, gparams)
# Get all updates
updates = param_updates + self.inner_updates_
# Compile model
inputs = self.in_nodes_ + self.gt_nodes_ + [K.common_tr_phase_node]
outputs = [loss_node]
f = K.function_no_given(inputs, outputs, updates)
return f
def get_optimization_func(self, target_dim_list, loss_func, optimizer, clip):
"""Compile and return optimization function.
Args:
target_dim_list: list of integars. targets' dimension. e.g. target_dim_list=[2]
loss_func: string | function.
optimizer: object.
clip: None | real value.
Return:
optimization function.
"""
# set gt nodes
self.set_gt_nodes(target_dim_list)
# Default loss
if type(loss_func) is str:
assert len(self.out_nodes_)==1, "If the number of out_layers > 1, you need define your own loss_func!"
loss_node = obj.get(loss_func)(self.out_nodes_[0], self.gt_nodes_[0])
# User defined loss
else:
loss_node = loss_func(self)
# Compute gradient
gparams = K.grad(loss_node + self.reg_value_, self.params_)
# Clip gradient
if clip is not None:
gparams = [K.clip(gparam, -clip, clip) for gparam in gparams]
# Gradient based optimization
param_updates = optimizer.get_updates(self.params_, gparams)
# Get all updates
updates = param_updates + self.inner_updates_
# Compile model
inputs = self.in_nodes_ + self.gt_nodes_ + [K.common_tr_phase_node]
outputs = [loss_node]
f = K.function_no_given(inputs, outputs, updates)
return f
def beta_divergence(y_pred, y_gt, beta):
y_pred = K.clip(y_pred, _EPSILON, np.inf)
y_gt = K.clip(y_gt, _EPSILON, np.inf)
beta_mat = 1. / (beta*(beta-1)) * (K.power(y_gt, beta) + (beta-1) * K.power(y_pred, beta) - beta * y_gt * K.power(y_pred, (beta-1)))
return K.mean(K.sum(beta_mat, axis=-1))
def is_divergence(y_pred, y_gt):
y_pred = K.clip(y_pred, _EPSILON, np.inf)
y_gt = K.clip(y_gt, _EPSILON, np.inf)
is_mat = y_gt / y_pred - K.log(y_gt / y_pred) - 1
return K.mean(K.sum(is_mat, axis=-1))
def kl_divergence(y_pred, y_gt):
y_pred = K.clip(y_pred, _EPSILON, np.inf)
y_gt = K.clip(y_gt, _EPSILON, np.inf)
kl_mat = y_gt * K.log(y_gt / y_pred) - y_gt + y_pred
return K.mean(K.sum(kl_mat, axis=-1))
def binary_crossentropy(p_y_pred, y_gt):
p_y_pred = K.clip(p_y_pred, _EPSILON, 1. - _EPSILON)
return K.mean(K.mean(K.binary_crossentropy(p_y_pred, y_gt), axis=-1))
def sparse_categorical_crossentropy(p_y_pred, y_gt):
p_y_pred = K.clip(p_y_pred, _EPSILON, 1. - _EPSILON)
y_gt = K.to_one_hot(y_gt, )
return K.mean(K.categorical_crossentropy(p_y_pred, y_gt))
def categorical_crossentropy(p_y_pred, y_gt):
p_y_pred = K.clip(p_y_pred, _EPSILON, 1. - _EPSILON)
return K.mean(K.categorical_crossentropy(p_y_pred, y_gt))
def fit(self,
x,
y,
batch_size=100,
n_epochs=10,
loss_func='categorical_crossentropy',
optimizer=SGD(lr=0.01, rho=0.9),
clip=None,
callbacks=[],
shuffle=True,
verbose=1):
x = to_list(x)
y = to_list(y)
# format
x = [K.format_data(e) for e in x]
y = [K.format_data(e) for e in y]
# shuffle data
if shuffle:
x, y = supports.shuffle(x, y)
# check data
self._check_data(y, loss_func)
# init gt_nodes
self._gt_nodes_ = [K.placeholder(e.ndim) for e in y]
# memory usage
print "Train",
self._show_memory_usage(self._layer_list_, batch_size)
# default objective
if type(loss_func) is str:
assert len(self._out_nodes_)==len(self._gt_nodes_), "If you are using default objectives, " \
+ "out_node of out_layers must match ground truth!"
loss_node = sum([
obj.get(loss_func)(pred_node,
gt_node) for pred_node, gt_node in zip(
self._out_nodes_, self._gt_nodes_)
])
# user defined objective
else:
loss_node = loss_func(self._out_nodes_, self._any_nodes_,
self._gt_nodes_)
#loss_node = loss_func( self )
# gradient
gparams = K.grad(loss_node + self._reg_value_, self._params_)
# todo clip gradient
if clip is not None:
gparams = [K.clip(gparam, -clip, clip) for gparam in gparams]
# gradient based opt
param_updates = optimizer.get_updates(self._params_, gparams)
# get all updates
updates = param_updates + self._inner_updates_
# compile for callback
if callbacks is not None:
callbacks = to_list(callbacks)
for callback in callbacks:
callback.compile(self)
# compile model
input_nodes = self._in_nodes_ + self._gt_nodes_
output_nodes = [loss_node]
f = K.function_no_given(input_nodes, self._tr_phase_node_,
output_nodes, updates)
# train
N = len(x[0])
batch_num = int(np.ceil(float(N) / batch_size))
n_abs_epoch = n_epochs + self._epoch_
# callback
print '\n0th epoch:'
for callback in callbacks:
if (self._epoch_ % callback.call_freq == 0):
callback.call()
while self._epoch_ < n_abs_epoch:
self._epoch_ += 1
# train
t1 = time.time()
loss_list = []
for i2 in xrange(batch_num):
batch_x = [
e[i2 * batch_size:min((i2 + 1) * batch_size, N)] for e in x
]
batch_y = [
e[i2 * batch_size:min((i2 + 1) * batch_size, N)] for e in y
]
in_list = batch_x + batch_y + [1.]
loss = f(*in_list)[0] # training phase
loss_list.append(loss)
if verbose == 1:
self._print_progress(self._epoch_, batch_num, i2)
if verbose == 2:
self._print_progress_loss(self._epoch_, batch_num, i2,
loss)
t2 = time.time()
self._tr_time_ += (t2 - t1)
if verbose != 0:
print '\n', ' tr_time: ', "%.2f" % (
t2 - t1), 's' # print an empty line
# callback
for callback in callbacks:
if (self._epoch_ % callback.call_freq == 0):
callback.call()