def main():
args = parse_args()
cfg = mmcv.Config.fromfile(args.config)
cfg.data.videos_per_gpu = 1
net = build_recognizer(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
net.eval()
load_checkpoint(net, args.checkpoint, force_matching=True)
input_time_size = cfg.input_clip_length
input_image_size = (tuple(cfg.input_img_size) if isinstance(
cfg.input_img_size, (list, tuple)) else
(cfg.input_img_size, cfg.input_img_size))
input_size = (3, input_time_size) + input_image_size
output_path = args.output_name
if not output_path.endswith('.onnx'):
output_path = '{}.onnx'.format(output_path)
base_output_dir = dirname(output_path)
if not exists(base_output_dir):
makedirs(base_output_dir)
if hasattr(net, 'forward_inference'):
net.forward = net.forward_inference
convert_to_onnx(net, input_size, args.output_name, check=args.check)
def main():
parser = ArgumentParser()
parser.add_argument('config', help='Config file path')
parser.add_argument('--load_from',
help='the checkpoint file to init weights from')
parser.add_argument('--load2d_from',
help='the checkpoint file to init 2D weights from')
parser.add_argument('--update_config',
nargs='+',
action=ExtendedDictAction,
help='arguments in dict')
args = parser.parse_args()
cfg = mmcv.Config.fromfile(args.config)
if args.update_config is not None:
cfg.merge_from_dict(args.update_config)
cfg = update_config(cfg, args)
net = build_recognizer(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
net.eval()
if cfg.load_from:
logger = get_root_logger(log_level=cfg.log_level)
load_checkpoint(net,
cfg.load_from,
strict=False,
logger=logger,
show_converted=True,
force_matching=True)
conv_layers = collect_conv_layers(net)
show_stat(conv_layers)
def main(args):
cfg = mmcv.Config.fromfile(args.config)
if args.update_config is not None:
cfg.merge_from_dict(args.update_config)
cfg.data.videos_per_gpu = 1
if cfg.get('seed'):
print(f'Set random seed to {cfg.seed}')
set_random_seed(cfg.seed)
class_maps = None
if cfg.get('classes'):
class_maps = {0: {k: v for k, v in enumerate(sorted(cfg.classes))}}
model = build_recognizer(cfg.model,
train_cfg=None,
test_cfg=cfg.test_cfg,
class_maps=class_maps)
model.eval()
load_checkpoint(model, args.checkpoint, force_matching=True)
if hasattr(model, 'forward_inference'):
model.forward = model.forward_inference
input_time_size = cfg.input_clip_length
input_image_size = (tuple(cfg.input_img_size) if isinstance(
cfg.input_img_size, (list, tuple)) else
(cfg.input_img_size, cfg.input_img_size))
input_size = (3, input_time_size) + input_image_size
onnx_model_path = join(args.output_dir,
splitext(basename(args.config))[0] + '.onnx')
base_output_dir = dirname(onnx_model_path)
if not exists(base_output_dir):
makedirs(base_output_dir)
convert_to_onnx(model,
input_size,
onnx_model_path,
opset=args.opset,
check=True)
if args.target == 'openvino':
input_shape = (1, ) + input_size
export_to_openvino(cfg, onnx_model_path, args.output_dir, input_shape,
args.input_format)
meta = {'model_classes': model.CLASSES[0]}
with open(args.meta_info, 'w') as output_meta_stream:
json.dump(meta, output_meta_stream)
def main():
args = parse_args()
cfg = mmcv.Config.fromfile(args.config)
cfg.data.videos_per_gpu = 1
cfg.model.type += '_Inference'
cfg.model.backbone.type += '_Inference'
cfg.model.backbone.inference = True
cfg.model.cls_head.type += '_Inference'
if args.num_classes is not None and args.num_classes > 0:
cfg.num_test_classes = args.num_classes
cfg.model.cls_head.num_classes = args.num_classes
net = build_recognizer(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
net.eval()
load_checkpoint(net, args.checkpoint, force_matching=True)
time_length = cfg.data.test.out_length if hasattr(cfg.data.test, 'out_length') else cfg.data.test.new_length
input_size = (3, time_length) + cfg.data.test.input_size
convert_to_onnx(net, input_size, args.output_name, check=args.check)
def main():
parser = ArgumentParser()
parser.add_argument('--config',
type=str,
required=True,
help='Test config file path')
parser.add_argument('--checkpoint',
type=str,
required=True,
help='Checkpoint file')
parser.add_argument('--data_dir',
type=str,
required=True,
help='The dir with dataset')
parser.add_argument('--out_dir',
type=str,
required=True,
help='Output directory')
parser.add_argument('--dataset',
type=str,
required=True,
help='Dataset name')
parser.add_argument('--gpus',
default=1,
type=int,
help='GPU number used for annotating')
parser.add_argument('--proc_per_gpu',
default=2,
type=int,
help='Number of processes per GPU')
parser.add_argument('--mode',
choices=['train', 'val', 'test'],
default='train')
args = parser.parse_args()
assert exists(args.config)
assert exists(args.checkpoint)
assert exists(args.data_dir)
cfg = Config.fromfile(args.config)
cfg = update_config(cfg, args, trg_name=args.dataset)
cfg = propagate_root_dir(cfg, args.data_dir)
dataset = build_dataset(cfg.data, args.mode, dict(test_mode=True))
data_pipeline = Compose(dataset.pipeline.transforms[1:])
print('{} dataset:\n'.format(args.mode) + str(dataset))
tasks = prepare_tasks(dataset, cfg.input_clip_length)
print('Prepared tasks: {}'.format(sum([len(v) for v in tasks.values()])))
if not exists(args.out_dir):
makedirs(args.out_dir)
model = build_model(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint, strict=False)
batch_size = 4 * cfg.data.videos_per_gpu
if args.gpus == 1:
model = MMDataParallel(model, device_ids=[0])
model.eval()
process_tasks(tasks, dataset, model, args.out_dir, batch_size,
cfg.input_clip_length, data_pipeline)
else:
raise NotImplementedError
def main():
args = parse_args()
cfg = mmcv.Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
if args.num_classes is not None and args.num_classes > 0:
cfg.num_test_classes = args.num_classes
if args.data_dir is not None:
cfg = update_data_paths(cfg, args.data_dir)
assert args.mode in cfg.data
data_cfg = getattr(cfg.data, args.mode)
data_cfg.test_mode = True
dataset = obj_from_dict(data_cfg, datasets, dict(test_mode=True))
if args.gpus == 1:
model = build_recognizer(cfg.model,
train_cfg=None,
test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint, strict=False)
model = MMDataParallel(model, device_ids=[0])
data_loader = build_dataloader(
dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
outputs = single_test(model, data_loader)
else:
model_args = cfg.model.copy()
model_args.update(train_cfg=None, test_cfg=cfg.test_cfg)
model_type = getattr(recognizers, model_args.pop('type'))
outputs = parallel_test(model_type,
model_args,
args.checkpoint,
dataset,
_data_func,
range(args.gpus),
workers_per_gpu=args.proc_per_gpu)
gt_labels = []
for i in range(len(dataset)):
ann = dataset.get_ann_info(i)
gt_labels.append(ann['label'])
results = np.array([res.cpu().numpy().mean(axis=0) for res in outputs],
dtype=np.float32)
if cfg.data.num_test_classes is not None and cfg.data.num_test_classes > 0:
results = results[:, :cfg.data.num_test_classes]
top1_value = mean_top_k_accuracy(results, gt_labels, k=1)
top5_value = mean_top_k_accuracy(results, gt_labels, k=5)
print("\nMean Top-1 Accuracy = {:.03f}%".format(top1_value * 100))
print("Mean Top-5 Accuracy = {:.03f}%".format(top5_value * 100))
map_value = mean_average_precision(results, gt_labels)
print("mAP = {:.03f}%".format(map_value * 100))
invalid_ids = invalid_filtered(results, gt_labels)
print('\nNum invalid classes: {} / {}'.format(len(invalid_ids),
cfg.data.num_test_classes))
num_invalid_samples = sum([len(ids) for ids in invalid_ids.values()])
print('Num invalid samples: {} / {}'.format(num_invalid_samples,
len(gt_labels)))
def main(args):
cfg = mmcv.Config.fromfile(args.config)
if args.update_config is not None:
cfg.merge_from_dict(args.update_config)
cfg.data.videos_per_gpu = 1
if cfg.get('seed'):
print(f'Set random seed to {cfg.seed}')
set_random_seed(cfg.seed)
class_maps = None
if cfg.get('classes'):
class_maps = {0: {k: v for k, v in enumerate(sorted(cfg.classes))}}
model = build_recognizer(
cfg.model,
train_cfg=None,
test_cfg=cfg.test_cfg,
class_maps=class_maps
)
model.eval()
load_checkpoint(model, args.checkpoint, force_matching=True)
if hasattr(model, 'forward_inference'):
model.forward = model.forward_inference
input_time_size = cfg.input_clip_length
input_image_size = (tuple(cfg.input_img_size)
if isinstance(cfg.input_img_size, (list, tuple))
else (cfg.input_img_size, cfg.input_img_size))
input_size = (3, input_time_size) + input_image_size
# BEGIN nncf part
was_model_compressed = is_checkpoint_nncf(args.checkpoint)
cfg_contains_nncf = cfg.get('nncf_config')
if cfg_contains_nncf and not was_model_compressed:
raise RuntimeError('Trying to make export with NNCF compression '
'a model snapshot that was NOT trained with NNCF')
if was_model_compressed and not cfg_contains_nncf:
# reading NNCF config from checkpoint
nncf_part = get_nncf_config_from_meta(args.checkpoint)
for k, v, in nncf_part.items():
cfg[k] = v
if cfg.get('nncf_config'):
if torch.cuda.is_available():
model.cuda()
check_nncf_is_enabled()
cfg.load_from = args.checkpoint
cfg.resume_from = None
compression_ctrl, model = wrap_nncf_model(model, cfg, None, get_fake_input, export=True)
compression_ctrl.prepare_for_export()
# END nncf part
onnx_model_path = join(args.output_dir, splitext(basename(args.config))[0] + '.onnx')
base_output_dir = dirname(onnx_model_path)
if not exists(base_output_dir):
makedirs(base_output_dir)
convert_to_onnx(model, input_size, onnx_model_path, opset=args.opset, check=True)
if args.target == 'openvino':
input_shape = (1,) + input_size
export_to_openvino(cfg, onnx_model_path, args.output_dir, input_shape, args.input_format)
meta = {'model_classes': model.CLASSES[0]}
with open(args.meta_info, 'w') as output_meta_stream:
json.dump(meta, output_meta_stream)
def main():
parser = ArgumentParser()
parser.add_argument('--config', '-c', type=str, required=True)
parser.add_argument('--checkpoint', '-w', type=str, required=True)
parser.add_argument('--dataset_name', '-n', type=str, required=True)
parser.add_argument('--data_dir', '-d', type=str, required=True)
parser.add_argument('--predictions', '-p', type=str, required=True)
parser.add_argument('--movements', '-m', type=str, required=True)
parser.add_argument('--keypoints', '-k', type=str, required=True)
parser.add_argument('--out_annotation', '-o', type=str, required=True)
args = parser.parse_args()
assert exists(args.config)
assert exists(args.weights)
assert exists(args.data_dir)
assert exists(args.predictions)
assert exists(args.movements)
assert exists(args.keypoints)
assert args.dataset_name is not None and args.dataset_name != ''
assert args.out_annotation is not None and args.out_annotation != ''
cfg = Config.fromfile(args.config)
cfg = update_config(cfg, args, trg_name=args.dataset_name)
cfg = propagate_root_dir(cfg, args.data_dir)
dataset = build_dataset(cfg.data, 'train', dict(test_mode=True))
data_pipeline = Compose(dataset.pipeline.transforms[1:])
print('{} dataset:\n'.format(args.mode) + str(dataset))
model = build_model(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint, strict=False)
model = MMDataParallel(model, device_ids=[0])
model.eval()
annotation_path = join(args.data_dir, cfg.data.train.sources[0],
cfg.data.train.ann_file)
records = load_annotation(annotation_path)
predictions = load_distributed_data(args.predictions,
parse_predictions_file, 'txt')
movements = load_distributed_data(args.movements, parse_movements_file,
'txt')
hand_kpts = load_distributed_data(args.keypoints, parse_kpts_file, 'json')
print('Loaded records: {}'.format(len(records)))
invalid_stat = dict()
all_candidates = []
ignore_candidates = get_ignore_candidates(records, IGNORE_LABELS)
all_candidates += ignore_candidates
static_candidates, static_invalids = get_regular_candidates(
records,
predictions,
movements,
hand_kpts,
cfg.data.output.length,
False,
STATIC_LABELS,
NEGATIVE_LABEL,
NO_MOTION_LABEL,
min_score=0.9,
min_length=4,
max_distance=1)
all_candidates += static_candidates
invalid_stat = update_stat(invalid_stat, static_invalids)
print('Static candidates: {}'.format(len(static_candidates)))
if len(invalid_stat) > 0:
print('Ignored records after static analysis:')
for ignore_label, ignore_values in invalid_stat.items():
print(' - {}: {}'.format(ignore_label.replace('_', ' '),
len(ignore_values)))
dynamic_candidates, dynamic_invalids = get_regular_candidates(
records,
predictions,
movements,
hand_kpts,
cfg.data.output.length,
True,
DYNAMIC_LABELS,
NEGATIVE_LABEL,
NO_MOTION_LABEL,
min_score=0.9,
min_length=4,
max_distance=1)
all_candidates += dynamic_candidates
invalid_stat = update_stat(invalid_stat, dynamic_invalids)
print('Dynamic candidates: {}'.format(len(dynamic_candidates)))
if len(invalid_stat) > 0:
print('Ignored records after dynamic analysis:')
for ignore_label, ignore_values in invalid_stat.items():
print(' - {}: {}'.format(ignore_label.replace('_', ' '),
len(ignore_values)))
fixed_records, fix_stat = find_best_match(all_candidates, model, dataset,
NEGATIVE_LABEL)
invalid_stat = update_stat(invalid_stat, fix_stat)
print('Final records: {}'.format(len(fixed_records)))
if len(invalid_stat) > 0:
print('Final ignored records:')
for ignore_label, ignore_values in invalid_stat.items():
print(' - {}: {}'.format(ignore_label.replace('_', ' '),
len(ignore_values)))
for ignored_record in ignore_values:
print(' - {}'.format(ignored_record.path))
dump_records(fixed_records, args.out_annotation)
print('Fixed annotation has been stored at: {}'.format(
args.out_annotation))