def __init__(self, class_vector, p_vec, batch_size, verbose=False):
"""Stratified Sampling
Args:
class_vector (np.array): a vector of class labels
p_vec (list[float]): list of probabilities for each class
batch_size (int): batch_size
verbose
"""
self.n_splits = int(class_vector.shape[0] / batch_size)
self.class_vector = class_vector
self.p_vec = p_vec
self.batch_size = batch_size
self.batch_sizes = get_batch_sizes(self.p_vec,
self.batch_size,
verbose=verbose)
# check that the individual batch size will always be > 0
for i, batch_size in enumerate(self.batch_sizes):
if batch_size == 0:
warnings.warn("Batch size for class {} is 0.".format(i))
self.classes = np.arange(len(p_vec))
assert np.all(
np.sort(pd.Series(self.class_vector).unique()) == self.classes)
idx_all = np.arange(len(self.class_vector))
self.class_idx_iterators = [
iterable_cycle(
np.random.permutation(idx_all[self.class_vector == cls]))
for cls in self.classes
]
def batch_train_iter(self, cycle=True, **kwargs):
"""Returns samples directly useful for training the model:
(x["inputs"],x["targets"])
Args:
cycle: when True, the returned iterator will run indefinitely go through the dataset
Use True with `fit_generator` in Keras.
**kwargs: Arguments passed to self.batch_iter(**kwargs)
"""
if cycle:
return ((to_numpy(x["inputs"]), to_numpy(x["targets"]))
for x in iterable_cycle(self._batch_iterable(**kwargs)))
else:
return ((x["inputs"], x["targets"]) for x in self.batch_iter(**kwargs))