def half_incr():
classes = get_classes()
included = []
num_classes = 0
for c in classes:
included += c
num_classes = len(included)
print("{} classes: {}".format(num_classes, c))
root, annFile = getds('train2014')
dataset = COCOLoader(root, annFile, included=c)
root, annFile = getds('val2014')
dataset_test = COCOLoader(root, annFile, included=included)
print('data prepared, train data: {}'.format(len(dataset)))
print('data prepared, test data: {}'.format(len(dataset_test)))
yield num_classes, dataset, dataset_test
def create_coco_dataset(split):
loader = COCOLoader(split)
sz = len(loader.get_filenames())
print("Contains %i files" % sz)
if split in ['minival2014', 'valminusminival2014']:
realsplit = 'val2014'
else:
realsplit = split
shard_size = 5000
num_shards = ceil(sz/shard_size)
print("So we decided to split it in %i shards" % num_shards)
image_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(tf.expand_dims(image_placeholder, 2))
with tf.Session('') as sess:
for shard in range(num_shards):
print("Shard %i/%i is starting" % (shard, num_shards))
output_file = os.path.join(DATASETS_ROOT, 'coco-seg-%s-%.5d-of-%.5d' % (split, shard, num_shards))
writer = tf.python_io.TFRecordWriter(output_file)
for i in range(shard*shard_size, min(sz, (shard+1)*shard_size)):
f = loader.get_filenames()[i]
img = loader.coco.loadImgs(f)[0]
path = '%simages/%s/%s' % (loader.root, realsplit, img['file_name'])
with tf.gfile.FastGFile(path, 'rb') as ff:
image_data = ff.read()
gt_bb, gt_cats, w, h, diff = loader.read_annotations(f)
gt_bb = normalize_bboxes(gt_bb, w, h)
segmentation = np.zeros((h, w), dtype=np.uint8)
coco_anns = loader._get_coco_annotations(f, only_instances=False)
# instances = np.zeros([instance_nums, h, w], dtype=np.uint8)
instances = []
for num_ann, ann in enumerate(coco_anns):
mask = loader._read_segmentation(ann, h, w)
cid = loader.coco_ids_to_internal[ann['category_id']]
assert mask.shape == segmentation.shape
segmentation[mask > 0] = cid
instance = np.zeros((h, w), dtype=np.uint8)
instance[mask > 0] = cid
instance = list(instance.reshape(-1))
instances.extend(instance)
# if num_ann < instance_nums:
# instances[num_ann] = instance
# instances = instances.reshape(-1)
png_string = sess.run(encoded_image,
feed_dict={image_placeholder: segmentation})
example = _convert_to_example(path, image_data, gt_bb, gt_cats, diff, png_string,
instances, h, w)
if i % 100 == 0:
print("%i files are processed" % i)
writer.write(example.SerializeToString())
writer.close()
print("Done")
def offline():
classes = [list(range(0, 41))] + [[*range(41, 81)]
] #offlien half and last half
included = []
num_classes = 0
for c in classes:
included += c
num_classes = len(included)
print("{} classes: {}".format(num_classes, c))
root, annFile = getds('train2014')
dataset = COCOLoader(root, annFile, included=included)
root, annFile = getds('val2014')
dataset_test = COCOLoader(root, annFile, included=included)
print('data prepared, train data: {}'.format(len(dataset)))
print('data prepared, test data: {}'.format(len(dataset_test)))
yield num_classes, dataset, dataset_test
def main(argv=None): # pylint: disable=unused-argument
assert args.ckpt > 0 or args.batch_eval
assert args.detect or args.segment, "Either detect or segment should be True"
if args.trunk == 'resnet50':
net = ResNet
depth = 50
if args.trunk == 'resnet101':
net = ResNet
depth = 101
if args.trunk == 'vgg16':
net = VGG
depth = 16
net = net(config=net_config, depth=depth, training=False)
if args.dataset == 'voc07' or args.dataset == 'voc07+12':
loader = VOCLoader('07', 'test')
if args.dataset == 'voc12':
loader = VOCLoader('12', 'val', segmentation=args.segment)
if args.dataset == 'coco':
loader = COCOLoader(args.split)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)) as sess:
detector = Detector(sess,
net,
loader,
net_config,
no_gt=args.no_seg_gt)
if args.dataset == 'coco':
tester = COCOEval(detector, loader)
else:
tester = Evaluation(detector,
loader,
iou_thresh=args.voc_iou_thresh)
if not args.batch_eval:
detector.restore_from_ckpt(args.ckpt)
tester.evaluate_network(args.ckpt)
else:
log.info('Evaluating %s' % args.run_name)
ckpts_folder = CKPT_ROOT + args.run_name + '/'
out_file = ckpts_folder + evaluation_logfile
max_checked = get_last_eval(out_file)
log.debug("Maximum checked ckpt is %i" % max_checked)
with open(out_file, 'a') as f:
start = max(args.min_ckpt, max_checked + 1)
ckpt_files = glob(ckpts_folder + '*.data*')
folder_has_nums = np.array(list((map(filename2num,
ckpt_files))),
dtype='int')
nums_available = sorted(
folder_has_nums[folder_has_nums >= start])
nums_to_eval = [nums_available[-1]]
for n in reversed(nums_available):
if nums_to_eval[-1] - n >= args.step:
nums_to_eval.append(n)
nums_to_eval.reverse()
for ckpt in nums_to_eval:
log.info("Evaluation of ckpt %i" % ckpt)
tester.reset()
detector.restore_from_ckpt(ckpt)
res = tester.evaluate_network(ckpt)
f.write(res)
f.flush()
def main(argv=None): # pylint: disable=unused-argument
assert args.ckpt > 0 or args.batch_eval
assert args.detect or args.segment, "Either detect or segment should be True"
if args.trunk == 'resnet50':
net = ResNet
depth = 50
if args.trunk == 'resnet101':
net = ResNet
depth = 101
if args.trunk == 'vgg16':
net = VGG
depth = 16
net = net(config=net_config, depth=depth, training=False)
if args.dataset == 'voc07' or args.dataset == 'voc07+12':
loader = VOCLoader('07', 'test')
if args.dataset == 'voc12':
loader = VOCLoader('12', 'val', segmentation=args.segment)
if args.dataset == 'coco':
loader = COCOLoader(args.split)
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False,
gpu_options=tf.GPUOptions(allow_growth=True,
per_process_gpu_memory_fraction=0.2))) as sess:
if args.use_profile:
profiler = model_analyzer.Profiler(graph=sess.graph)
detector = Detector(sess, net, loader, net_config,
no_gt=args.no_seg_gt, profiler=profiler)
else:
detector = Detector(sess, net, loader, net_config,
no_gt=args.no_seg_gt)
if args.dataset == 'coco':
tester = COCOEval(detector, loader)
else:
tester = Evaluation(detector, loader, iou_thresh=args.voc_iou_thresh)
if not args.batch_eval:
detector.restore_from_ckpt(args.ckpt)
tester.evaluate_network(args.ckpt)
else:
log.info('Evaluating %s' % args.run_name)
ckpts_folder = CKPT_ROOT + args.run_name + '/'
out_file = ckpts_folder + evaluation_logfile
max_checked = get_last_eval(out_file)
log.debug("Maximum checked ckpt is %i" % max_checked)
with open(out_file, 'a') as f:
start = max(args.min_ckpt, max_checked+1)
ckpt_files = glob(ckpts_folder + '*.data*')
folder_has_nums = np.array(list((map(filename2num, ckpt_files))), dtype='int')
nums_available = sorted(folder_has_nums[folder_has_nums >= start])
nums_to_eval = [nums_available[-1]]
for n in reversed(nums_available):
if nums_to_eval[-1] - n >= args.step:
nums_to_eval.append(n)
nums_to_eval.reverse()
for ckpt in nums_to_eval:
log.info("Evaluation of ckpt %i" % ckpt)
tester.reset()
detector.restore_from_ckpt(ckpt)
res = tester.evaluate_network(ckpt)
f.write(res)
f.flush()
if args.use_profile:
profile_scope_builder = option_builder.ProfileOptionBuilder(
# option_builder.ProfileOptionBuilder.trainable_variables_parameter()
)
profile_scope_builder.with_max_depth(4)
profile_scope_builder.with_min_memory(int(2e6))
profile_scope_builder.with_step(2)
profile_scope_builder.select(['bytes'])
# profile_scope_builder.with_node_names(show_name_regexes=['.*resnet.*', '.*ssd.*'])
# profile_scope_builder.with_node_names(hide_name_regexes=['.*resnet.*', '.*ssd.*'])
# profile_scope_builder.order_by('output_bytes')
detector.profiler.profile_name_scope(profile_scope_builder.build())