def next_minibatch(self, batch_size):
'''
Return the next mini-batch, we do data augmentation during constructing each mini-batch.
'''
data_size = len(self.data)
batch_end = min(self.batch_start + batch_size, data_size)
current_batch_size = batch_end - self.batch_start
if current_batch_size < 0:
raise Exception('Reach the end of the training data.')
inputs = np.empty(shape=(current_batch_size, 1, self.width,
self.height),
dtype=np.float32)
targets = np.empty(shape=(current_batch_size, self.emotion_count),
dtype=np.float32)
for idx in range(self.batch_start, batch_end):
index = self.indices[idx]
distorted_image = imgu.distort_img(self.data[index][1],
self.data[index][3], self.width,
self.height, self.max_shift,
self.max_scale, self.max_angle,
self.max_skew, self.do_flip)
final_image = imgu.preproc_img(distorted_image)
inputs[idx - self.batch_start] = final_image
targets[idx - self.batch_start, :] = self._process_target(
self.data[index][2])
self.batch_start += current_batch_size
return inputs, targets, current_batch_size
def next_minibatch(self,
batch_size,
num_data_partitions=1,
partition_index=0):
'''
Return the next mini-batch, we do data augmentation during constructing each mini-batch.
'''
data_size = len(self.data)
total_batch_size = batch_size * num_data_partitions
batch_end = min(self.batch_start + total_batch_size, data_size)
current_batch_size = batch_end - self.batch_start
if current_batch_size < 0:
raise Exception('Reach the end of the training data.')
#determine individual worker batch size, usually equals batch_size except possibly for last partiton of data,
#and corresponding worker_batch_start and worker_batch_end indices
worker_batch_size = current_batch_size // num_data_partitions
worker_batch_start = self.batch_start + worker_batch_size * partition_index
if ((partition_index + 1) == num_data_partitions):
worker_batch_end = batch_end
worker_batch_size = worker_batch_end - worker_batch_start
else:
worker_batch_end = worker_batch_start + worker_batch_size
inputs = np.empty(shape=(worker_batch_size, 1, self.width,
self.height),
dtype=np.float32)
targets = np.empty(shape=(worker_batch_size, self.emotion_count),
dtype=np.float32)
for idx in range(worker_batch_start, worker_batch_end):
index = self.indices[idx]
distorted_image = imgu.distort_img(self.data[index][1],
self.data[index][3], self.width,
self.height, self.max_shift,
self.max_scale, self.max_angle,
self.max_skew, self.do_flip)
final_image = imgu.preproc_img(distorted_image)
inputs[idx - worker_batch_start] = final_image
targets[idx - worker_batch_start, :] = self._process_target(
self.data[index][2])
self.batch_start += current_batch_size
return inputs, targets, worker_batch_size