def _split_repeated_holdout(self, split_ratio, is_stratified_sampling=True, idx=[]):
assert(self.train_or_test in ['train','sample'])
# split by t/v
if is_stratified_sampling:
labels_num = len(set(idx))
labels_index = [[] for _ in range(labels_num)]
for index, label in enumerate(idx):
labels_index[label].append(index)
train_idx = [[] for _ in range(labels_num)]
validation_idx = [[] for _ in range(labels_num)]
for label in range(labels_num):
label_samples = labels_index[label]
get_rng().shuffle(label_samples)
top_k = int(split_ratio * len(label_samples))
train_idx[label].extend(label_samples[0:top_k])
validation_idx[label].extend((label_samples[top_k:]))
train_idx = reduce(lambda x, y: x+y, train_idx)
validation_idx = reduce(lambda x, y: x+y, validation_idx)
return train_idx, validation_idx
else:
index = range(len(idx))
get_rng().shuffle(index)
top_k = int(split_ratio * len(idx))
train_idx = index[0:top_k]
validation_idx = (index[top_k:])
return train_idx, validation_idx
def __init__(self, inputs):
super(RandomMixData, self).__init__(name=None,
action=None,
inputs=inputs)
self.rng = get_rng(self)
self.input_list_flag = np.ones((len(self._positional_inputs)),
dtype=np.uint8)
def __init__(self, inputs, propabilitys):
super(RandomChooseData, self).__init__(name=None,
action=None,
inputs=inputs)
self.input_list = []
self.input_list.extend([i for i in self._positional_inputs])
self.propability_list = tuple(propabilitys)
self.rng = get_rng(self)
def __init__(self, inputs, capacity=100, samples_per_class=1, classes=[]):
super(Episode, self).__init__(name=None, action=None, inputs=inputs)
self.samples_per_class = samples_per_class
self.classes = classes
self.capacity = capacity
self.consume_num = 0
self.episode_train = _TransparantNode(upper_node=Node.inputs(self),
name='episode-train',
active_request=False)
self.is_episode_ok = False
self.episode_classes_map = {c: i for i, c in enumerate(self.classes)}
self.rng = get_rng(self)
def _split_bootstrap(self, idx):
assert(self.train_or_test in ['train','sample'])
is_ok = False
train_idx = []
validation_idx = []
while not is_ok:
selected_idx = get_rng().randint(low=0, high=len(idx)-1, size=len(idx)).tolist()
train_idx = selected_idx
validation_idx = [i for i in range(len(idx)) if i not in selected_idx]
if len(validation_idx) > 0:
is_ok = True
return train_idx, validation_idx
def __init__(self, inputs=None, horiz=True, vert=False, prob=0.5):
super(Flip, self).__init__(name=None,
action=self.action,
inputs=inputs)
"""
Only one of horiz, vert can be set.
:param horiz: whether or not apply horizontal flip.
:param vert: whether or not apply vertical flip.
:param prob: probability of flip.
"""
if horiz and vert:
raise ValueError("Please use two Flip instead.")
elif horiz:
self.code = 1
elif vert:
self.code = 0
else:
raise ValueError("Are you kidding?")
self.prob = prob
self.rng = get_rng(self)
def rng(self):
if self._is_data_rng:
self.data_rng = get_rng(self)
return self.data_rng
def __init__(self, inputs=None, sigma=10):
super(DisturbNoise, self).__init__(name=None,
action=self.action,
inputs=inputs)
self._sigma = sigma
self.rng = get_rng(self)
def __init__(self, inputs=None, max_lighting=10):
super(DisturbLighting, self).__init__(name=None,
action=self.action,
inputs=inputs)
self._max_lighting = max_lighting
self.rng = get_rng(self)
def __init__(self, inputs=None, max_deg=10):
super(DisturbRotation, self).__init__(name=None,
action=self.action,
inputs=inputs)
self._max_deg = max_deg
self.rng = get_rng(self)
