def get_next_batch(self, batch_size):
if self._reader is None:
self._start_read()
if self._gpu_batch is None:
self._fill_reserved_data()
height, width = self._gpu_batch.data.shape
gpu_data = self._gpu_batch.data
gpu_labels = self._gpu_batch.labels
if self.index + batch_size >= width:
width = width - self.index
labels = gpu_labels[self.index:self.index + batch_size]
#data = gpu_data[:, self.index:self.index + batch_size]
data = gpuarray.zeros((height, width), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + width)
self.index = 0
self._fill_reserved_data()
else:
labels = gpu_labels[self.index:self.index + batch_size]
#data = gpu_data[:, self.index:self.index + batch_size]
data = gpuarray.zeros((height, batch_size), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + batch_size)
#labels = gpu_labels[self.index:self.index + batch_size]
self.index += batch_size
return BatchData(data, labels, self._gpu_batch.epoch)
def get_next_batch(self, batch_size):
if self._reader is None:
self._start_read()
if self._gpu_batch is None:
self._fill_reserved_data()
height, width = self._gpu_batch.data.shape
gpu_data = self._gpu_batch.data
gpu_labels = self._gpu_batch.labels
if self.index + batch_size >= width:
width = width - self.index
labels = gpu_labels[self.index:self.index + batch_size]
#data = gpu_data[:, self.index:self.index + batch_size]
data = gpuarray.zeros((height, width), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + width)
self.index = 0
self._fill_reserved_data()
else:
labels = gpu_labels[self.index:self.index + batch_size]
#data = gpu_data[:, self.index:self.index + batch_size]
data = gpuarray.zeros((height, batch_size), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + batch_size)
#labels = gpu_labels[self.index:self.index + batch_size]
self.index += batch_size
return BatchData(data, labels, self._gpu_batch.epoch, self._gpu_batch.batch_index)
def get_next_batch(self, batch_size):
if self._reader is None:
self._start_read()
if self._gpu_batch is None:
self._fill_reserved_data()
if not self.multiview:
height, width = self._gpu_batch.data.shape
gpu_data = self._gpu_batch.data
gpu_labels = self._gpu_batch.labels
epoch = self._gpu_batch.epoch
if self.index + batch_size >= width:
width = width - self.index
labels = gpu_labels[self.index:self.index + batch_size]
data = gpuarray.zeros((height, width), dtype=np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index,
self.index + width)
self.index = 0
self._fill_reserved_data()
else:
labels = gpu_labels[self.index:self.index + batch_size]
data = gpuarray.zeros((height, batch_size),
dtype=np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index,
self.index + batch_size)
self.index += batch_size
else:
# multiview provider
# number of views should be 10
# when using multiview, do not pre-move data and labels to gpu
height, width = self._cpu_batch.data.shape
cpu_data = self._cpu_batch.data
cpu_labels = self._cpu_batch.labels
epoch = self._cpu_batch.epoch
width /= self.num_view
if self.index + batch_size >= width:
batch_size = width - self.index
labels = cpu_labels[self.index:self.index + batch_size]
data = np.zeros((height, batch_size * self.num_view),
dtype=np.float32)
for i in range(self.num_view):
data[:, i * batch_size:(i + 1) *
batch_size] = cpu_data[:, self.index +
width * i:self.index +
width * i + batch_size]
data = copy_to_gpu(np.require(data, requirements='C'))
labels = copy_to_gpu(np.require(labels, requirements='C'))
self.index = (self.index + batch_size) / width
#util.log_info('Batch: %s %s %s', data.shape, gpu_labels.shape, labels.shape)
return BatchData(data, labels, epoch)
def logreg_cost_multiview(self, label, output, num_view):
unit = self.batch_size / num_view
if self.cost.shape[0] != unit:
self.cost = gpuarray.zeros((unit, 1), dtype = np.float32)
maxid = gpuarray.zeros((self.batch_size, 1), dtype = np.float32)
find_col_max_id(maxid, output)
self.batchCorrect = same_reduce_multiview(label, maxid, num_view)
tmp = gpuarray.zeros((output.shape[0], unit), dtype = np.float32)
gpu_partial_copy_to(output, tmp, 0, output.shape[0], 0, unit)
logreg_cost_col_reduce(tmp, label, self.cost)
def logreg_cost_multiview(self, label, output, num_view):
unit = self.batch_size / num_view
if self.cost.shape[0] != unit:
self.cost = gpuarray.zeros((unit, 1), dtype = np.float32)
maxid = gpuarray.zeros((self.batch_size, 1), dtype = np.float32)
find_col_max_id(maxid, output)
self.batchCorrect = same_reduce_multiview(label, maxid, num_view)
tmp = gpuarray.zeros((output.shape[0], unit), dtype = np.float32)
gpu_partial_copy_to(output, tmp, 0, output.shape[0], 0, unit)
logreg_cost_col_reduce(tmp, label, self.cost)
def get_next_batch(self, batch_size):
if self._reader is None:
self._start_read()
if self._gpu_batch is None:
self._fill_reserved_data()
if not self.multiview:
height, width = self._gpu_batch.data.shape
gpu_data = self._gpu_batch.data
gpu_labels = self._gpu_batch.labels
epoch = self._gpu_batch.epoch
if self.index + batch_size >= width:
width = width - self.index
labels = gpu_labels[self.index:self.index + batch_size]
data = gpuarray.zeros((height, width), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + width)
self.index = 0
self._fill_reserved_data()
else:
labels = gpu_labels[self.index:self.index + batch_size]
data = gpuarray.zeros((height, batch_size), dtype = np.float32)
gpu_partial_copy_to(gpu_data, data, 0, height, self.index, self.index + batch_size)
self.index += batch_size
else:
# multiview provider
# number of views should be 10
# when using multiview, do not pre-move data and labels to gpu
height, width = self._cpu_batch.data.shape
cpu_data = self._cpu_batch.data
cpu_labels = self._cpu_batch.labels
epoch = self._cpu_batch.epoch
width /= self.num_view
if self.index + batch_size >= width:
batch_size = width - self.index
labels = cpu_labels[self.index:self.index + batch_size]
data = np.zeros((height, batch_size * self.num_view), dtype = np.float32)
for i in range(self.num_view):
data[:, i* batch_size: (i+ 1) * batch_size] = cpu_data[:, self.index + width * i : self.index + width * i + batch_size]
data = copy_to_gpu(np.require(data, requirements = 'C'))
labels = copy_to_gpu(np.require(labels, requirements = 'C'))
self.index = (self.index + batch_size) / width
#util.log_info('Batch: %s %s %s', data.shape, gpu_labels.shape, labels.shape)
return BatchData(data, labels, epoch)
