def __get_images_labels_bboxes(self, data_sources, num_samples,
is_training_data):
self.dataset = pascalvoc_datasets.get_dataset_info(
data_sources, num_samples)
self.is_training_data = is_training_data
if self.is_training_data:
shuffle = True
#make sure most samples can be fetched in one epoch
self.num_readers = 2
else:
#make sure data is fetchd in sequence
shuffle = False
self.num_readers = 1
provider = slim.dataset_data_provider.DatasetDataProvider(
self.dataset,
shuffle=shuffle,
num_readers=self.num_readers,
common_queue_capacity=30 * self.batch_size,
common_queue_min=10 * self.batch_size)
# Get for SSD network: image, labels, bboxes.
[image, shape, format, self.filename, glabels, gbboxes,
gdifficults] = provider.get([
'image', 'shape', 'format', 'filename', 'object/label',
'object/bbox', 'object/difficult'
])
# Pre-processing image, labels and bboxes.
self.image, self.glabels, self.gbboxes = self.__preprocess_data(
image, glabels, gbboxes)
# anchors_1 = g_ssd_model.get_allanchors(minmaxformat=False)
anchors = g_ssd_model.get_allanchors(minmaxformat=True)
print(anchors[-1][-4:])
#flattent the anchors
temp_anchors = []
for i in range(len(anchors)):
temp_anchors.append(tf.reshape(anchors[i], [-1, 4]))
anchors = tf.concat(temp_anchors, axis=0)
self.jaccard = g_ssd_model.compute_jaccard(self.gbboxes, anchors)
# Assign groundtruth information for all default/anchor boxes
# gclasses, glocalisations, gscores = g_ssd_model.tf_ssd_bboxes_encode(glabels, gbboxes)
return
def __get_images_labels_bboxes(self, data_sources, num_samples,
is_training_data):
'''
:param data_sources: .tfrecord files
:param num_samples: number of samples in all the files
:param is_training_data:
:return:
'''
self.dataset = pascalvoc_datasets.get_dataset_info(
data_sources, num_samples)
self.is_training_data = is_training_data
if self.is_training_data:
shuffle = True
#make sure most samples can be fetched in one epoch
self.num_readers = 2
else:
#make sure data is fetchd in sequence
shuffle = False
self.num_readers = 1
provider = slim.dataset_data_provider.DatasetDataProvider(
self.dataset,
shuffle=shuffle,
num_readers=self.num_readers,
common_queue_capacity=20 * self.batch_size,
common_queue_min=10 * self.batch_size)
# Get for SSD network: image, labels, bboxes.
[image, shape, format, filename, glabels, gbboxes,
gdifficults] = provider.get([
'image', 'shape', 'format', 'filename', 'object/label',
'object/bbox', 'object/difficult'
])
glabels -= self.labels_offset
# Pre-processing image, labels and bboxes.
image, glabels, gbboxes = self.__preprocess_data(
image, glabels, gbboxes, shape)
gclasses_list, glocalisations_list, gscores_list = g_ssd_model.match_achors(
glabels, gbboxes, matching_threshold=self.matched_thresholds)
return self.__batching_data(image, glabels, format, filename, gbboxes,
gdifficults, gclasses_list,
glocalisations_list, gscores_list)