def _genDef(self, train_imgs, train_masks, attnlist, class_weight):
"""
图片取块生成器模板
:param train_imgs: 原始图
:param train_masks: 原始图groundtruth
:param attnlist: 目标区域list
:return: 取出的训练样本
"""
while 1:
Nimgs = train_imgs.shape[0]
for t in range(
int(self.config.subsample * self.config.total_train /
self.config.batch_size)):
if self.multi_proportion:
X = np.zeros([
self.config.batch_size, self.config.patch_height,
self.config.patch_width, 1
])
Y = np.zeros([
self.config.batch_size,
self.config.patch_height * self.config.patch_width,
self.config.seg_num + 1
])
else:
X = np.zeros([
self.config.batch_size, self.config.patch_height,
self.config.patch_width, 3
])
Y = np.zeros([
self.config.batch_size,
self.config.patch_height * self.config.patch_width,
self.config.seg_num + 1
])
for j in range(self.config.batch_size):
[i_center, x_center,
y_center] = self._CenterSampler(attnlist, class_weight,
Nimgs)
patch = train_imgs[
i_center,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2), :]
patch_mask = train_masks[
i_center, :,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2)]
X[j, :, :, :] = patch
Y[j, :, :] = genMasks(
np.reshape(patch_mask, [
1, self.config.seg_num, self.config.patch_height,
self.config.patch_width
]), self.config.seg_num)
yield (X, Y)
def _genDef(self, train_imgs, train_masks, attnlist, class_weight):
"""
Image block generator template
:param train_imgs: The original image
:param train_masks: Original map groundtruth
:param attnlist: Target area list
:return: Take out training samples
"""
while 1:
Nimgs = train_imgs.shape[0]
for t in range(
int(self.config.subsample * self.config.total_train /
self.config.batch_size)):
X = np.zeros([
self.config.batch_size, self.config.patch_height,
self.config.patch_width, 1
])
Y = np.zeros([
self.config.batch_size,
self.config.patch_height * self.config.patch_width,
self.config.seg_num + 1
])
for j in range(self.config.batch_size):
[i_center, x_center,
y_center] = self._CenterSampler(attnlist, class_weight,
Nimgs)
patch = train_imgs[
i_center,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2), :]
patch_mask = train_masks[
i_center, :,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2)]
X[j, :, :, :] = patch
Y[j, :, :] = genMasks(
np.reshape(patch_mask, [
1, self.config.seg_num, self.config.patch_height,
self.config.patch_width
]), self.config.seg_num)
yield (X, Y)
def _genDef(self, train_imgs, train_masks, attnlist, class_weight):
while 1:
Nimgs = train_imgs.shape[0]
for t in range(
int(self.config.subsample * self.config.total_train /
self.config.batch_size)):
X = np.zeros([
self.config.batch_size, self.config.patch_height,
self.config.patch_width, 1
])
Y = np.zeros([
self.config.batch_size,
self.config.patch_height * self.config.patch_width,
self.config.seg_num + 1
])
for j in range(self.config.batch_size):
[i_center, x_center,
y_center] = self._CenterSampler(attnlist, class_weight,
Nimgs)
patch = train_imgs[
i_center,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2), :]
patch_mask = train_masks[
i_center, :,
int(y_center - self.config.patch_height /
2):int(y_center + self.config.patch_height / 2),
int(x_center - self.config.patch_width /
2):int(x_center + self.config.patch_width / 2)]
X[j, :, :, :] = patch
Y[j, :, :] = genMasks(
np.reshape(patch_mask, [
1, self.config.seg_num, self.config.patch_height,
self.config.patch_width
]), self.config.seg_num)
yield (X, Y)