def __call__(self, X, y=None):
if y is not None:
alpha = self.config.cnf['balance_ratio'] ** self.count
class_weights = self.config.cnf['balance_weights'] * alpha \
+ self.config.cnf['final_balance_weights'] * (1 - alpha)
self.count += 1
indices = data.balance_per_class_indices(y, weights=class_weights)
X = X[indices]
y = y[indices]
return super(ResampleIterator, self).__call__(X, y, weight=None)
def __call__(self, X, y=None, transform=None, color_vec=None):
if y is not None:
alpha = self.config.cnf['balance_ratio'] ** self.count
class_weights = self.config.cnf['balance_weights'] * alpha \
+ self.config.cnf['final_balance_weights'] * (1 - alpha)
self.count += 1
indices = data.balance_per_class_indices(y, weights=class_weights)
X = X[indices]
y = y[indices]
return super(ResampleIterator, self).__call__(X, y, transform=transform,
color_vec=color_vec)
def __iter__(self):
n_samples = self.X.shape[0]
bs = self.batch_size
indices = data.balance_per_class_indices(self.y.ravel())
for i in range((n_samples + bs - 1) // bs):
r = np.random.rand()
if r < self.resample_prob:
sl = indices[np.random.randint(0, n_samples, size=bs)]
elif r < self.shuffle_prob:
sl = np.random.randint(0, n_samples, size=bs)
else:
sl = slice(i * bs, (i + 1) * bs)
Xb = self.X[sl]
if self.y is not None:
yb = self.y[sl]
else:
yb = None
yield self.transform(Xb, yb)
def __iter__(self):
n_samples = self.X.shape[0]
bs = self.batch_size
indices = data.balance_per_class_indices(self.y.ravel())
for i in range((n_samples + bs - 1) // bs):
r = np.random.rand()
if r < self.resample_prob:
sl = indices[np.random.randint(0, n_samples, size=bs)]
elif r < self.shuffle_prob:
sl = np.random.randint(0, n_samples, size=bs)
else:
sl = slice(i * bs, (i + 1) * bs)
Xb = self.X[sl]
if self.y is not None:
yb = self.y[sl]
else:
yb = None
yield self.transform(Xb, yb)