def convert_caffe2_to_pytorch(config_file, weight_dir, output_name):
cfg.merge_from_file(config_file)
cfg.defrost()
cfg.MODEL.WEIGHT = weight_dir
cfg.freeze()
mkdir(cfg.OUTPUT_DIR)
f = cfg.MODEL.WEIGHT
if f.startswith("catalog://"):
paths_catalog = import_file(
"maskrcnn_benchmark.config.paths_catalog", cfg.PATHS_CATALOG, True
)
f = paths_catalog.ModelCatalog.get(f[len("catalog://"):])
# download url files
if f.startswith("http"):
f = cache_url(f)
if output_name != "":
path_output = os.path.join(cfg.OUTPUT_DIR, output_name)
elif cfg.MODEL.WEIGHT.endswith(".pkl"):
path_output = os.path.join(cfg.OUTPUT_DIR, os.path.basename(f))
else:
str_output = cfg.MODEL.WEIGHT.split('/')
str_output = str_output[-1]
path_output = os.path.join(cfg.OUTPUT_DIR, str_output + '')
transfer_pretrained_weights(cfg, path_output)
def convert_pytorch_to_new_format(config_file, dir_input, dir_output, init_classifier=False):
if not os.path.isfile(dir_input):
return
cfg.merge_from_file(config_file)
cfg.defrost()
cfg.MODEL.WEIGHT = dir_input
cfg.freeze()
mkdir(os.path.dirname(dir_output))
f = cfg.MODEL.WEIGHT
if f.startswith("catalog://"):
paths_catalog = import_file(
"maskrcnn_benchmark.config.paths_catalog", cfg.PATHS_CATALOG, True
)
f = paths_catalog.ModelCatalog.get(f[len("catalog://"):])
# download url files
if f.startswith("http"):
f = cache_url(f)
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
checkpointer = ConvertedCheckpointer(cfg, model, save_dir=cfg.OUTPUT_DIR)
checkpointer.convert(cfg.MODEL.WEIGHT,
use_gn_fpn=False,
use_gn_head_box=True,
use_gn_head_mask=True,
keys=[])
model = {}
model["model"] = checkpointer.model.state_dict()
if init_classifier:
model["model"] = remove_keys_in_model(model["model"])
torch.save(model, dir_output)
print("write model to: " + dir_output)
if os.path.isfile(os.path.join(cfg.OUTPUT_DIR, 'last_checkpoint')):
os.remove(os.path.join(cfg.OUTPUT_DIR, 'last_checkpoint'))
def __init__(
self,
config_file,
checkpoint_path=None
):
cfg.merge_from_file(config_file)
cfg.defrost()
if checkpoint_path:
cfg.MODEL.WEIGHT = checkpoint_path
cfg.freeze()
self.cfg = cfg.clone()
self.model = build_detection_model(cfg)
self.device = torch.device(cfg.MODEL.DEVICE)
self.model.to(self.device)
self.cpu_device = torch.device("cpu")
self.checkpointer = DetectronCheckpointer(cfg, self.model)
_ = self.checkpointer.load(cfg.MODEL.WEIGHT)
if hasattr(self.checkpointer, "classes"):
self.classes = self.checkpointer.classes
def main(args):
seed_torch()
info = ulti.load_json()
num_gpus = get_num_gpus()
args.config_file = os.path.join(
info['training_dir'], 'e2e_faster_rcnn_R_50_FPN_Xconv1fc_1x_gn.yaml')
cfg.merge_from_file(args.config_file)
cfg.defrost()
cfg.OUTPUT_DIR = os.path.join(info['training_dir'], args.sub_dataset)
cfg.MODEL.WEIGHT = os.path.join(info['dataset_dir'], info['experiment'],
'Detector',
'Iter{}.pth'.format(info['iter']))
cfg.SOLVER.IMS_PER_BATCH = num_gpus * 4
cfg.TEST.IMS_PER_BATCH = num_gpus * 16
cfg.SOLVER.BASE_LR = 0.002
cfg.freeze()
mkdir(cfg.OUTPUT_DIR)
if args.sub_dataset is None:
args.sub_dataset = ""
if args.vis_title is None:
args.vis_title = os.path.basename(cfg.OUTPUT_DIR)
logger = setup_logger("maskrcnn_benchmark", cfg.OUTPUT_DIR, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(args)
logger.info("Collecting env info (might take some time)")
# logger.info("\n" + collect_env_info())
DatasetCatalog = None
train_dataset = cfg.DATASETS.TRAIN[0]
test_dataset = cfg.DATASETS.TEST[0]
paths_catalog = import_file("maskrcnn_benchmark.config.paths_catalog",
cfg.PATHS_CATALOG, True)
if args.sub_dataset != "":
DatasetCatalog = paths_catalog.DatasetCatalog
DatasetCatalog.DATASETS[train_dataset]['img_dir'] = os.path.join(
info['dataset_dir'], 'Images')
DatasetCatalog.DATASETS[train_dataset]['ann_file'] = os.path.join(
info['dataset_dir'], 'RCNN_data', 'train.json')
DatasetCatalog.DATASETS[test_dataset]['img_dir'] = os.path.join(
info['dataset_dir'], 'Images')
DatasetCatalog.DATASETS[test_dataset]['ann_file'] = os.path.join(
info['dataset_dir'], 'RCNN_data', 'test.json')
data = json.load(
open(DatasetCatalog.DATASETS[train_dataset]['ann_file']))
else:
data = json.load(
open(paths_catalog.DatasetCatalog.DATASETS[train_dataset]
['ann_file']))
iters_per_epoch = len(data['images'])
iters_per_epoch = math.ceil(iters_per_epoch / cfg.SOLVER.IMS_PER_BATCH)
args.iters_per_epoch = iters_per_epoch
cfg.defrost()
cfg.SOLVER.MAX_ITER = round(args.epochs * args.scale * iters_per_epoch)
cfg.SOLVER.STEPS = (round(8 * args.scale * iters_per_epoch),
round(11 * args.scale * iters_per_epoch),
round(16 * args.scale * iters_per_epoch))
cfg.freeze()
# logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
# logger.info(config_str)
# logger.info("Running with config:\n{}".format(cfg))
# logger.info(DatasetCatalog)
output_config_path = os.path.join(cfg.OUTPUT_DIR, 'config.yml')
logger.info("Saving config into: {}".format(output_config_path))
# save overloaded model config in the output directory
save_config(cfg, output_config_path)
if args.train:
args.skip_train = False
logger.info(args)
model = network.train(cfg, args, DatasetCatalog)
if args.test:
network.test(cfg, args, model=None, DatasetCatalog=DatasetCatalog)
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
if cfg.TEST.MULTI_SCALE:
data_loaders_val = []
for min_size_test, max_size_test in cfg.TEST.MULTI_SIZES:
cfg.defrost()
cfg.INPUT.MIN_SIZE_TEST = min_size_test
cfg.INPUT.MAX_SIZE_TEST = max_size_test
cfg.freeze()
data_loaders_val.extend(
make_data_loader(cfg,
is_train=False,
is_distributed=distributed))
output_folders = output_folders * len(cfg.TEST.MULTI_SIZES)
dataset_names = dataset_names * len(cfg.TEST.MULTI_SIZES)
else:
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
predictions = []
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
prediction = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
)
synchronize()
predictions.append(prediction)
if cfg.TEST.MULTI_SCALE:
logger.info("Processing multi-scale bbox voting....")
voted_predictions = voting(
predictions,
args.local_rank) # box_voting(predictions, args.local_rank)
torch.save(voted_predictions,
os.path.join(output_folders[0], 'predictions.pth'))
extra_args = dict(
box_only=cfg.MODEL.RPN_ONLY,
iou_types=iou_types,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
)
evaluate(dataset=data_loaders_val[0].dataset,
predictions=voted_predictions,
output_folder=output_folders[0],
**extra_args)
else:
for prediction, output_folder, dataset_name, data_loader_val in zip(
predictions, output_folders, dataset_names, data_loaders_val):
extra_args = dict(
box_only=cfg.MODEL.RPN_ONLY,
iou_types=iou_types,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
)
evaluate(dataset=data_loader_val.dataset,
predictions=prediction,
output_folder=output_folder,
**extra_args)
return 0
def main():
parser = argparse.ArgumentParser(description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default="/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--seq_test", action='store_true')
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(
backend="nccl", init_method="env://"
)
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
save_dir = cfg.OUTPUT_DIR
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank(), filename='test_log.txt')
logger.info("Using {} GPUs".format(num_gpus))
# logger.info(cfg)
logger.info("Collecting env info (might take some time)")
# logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
ori_output_dir = cfg.OUTPUT_DIR
if args.seq_test:
load_dir = cfg.TEST.LOAD_DIR
model_files = glob.glob(load_dir+'/*.pth')
model_files.sort()
min_iter = cfg.TEST.MIN_ITER
max_iter = cfg.TEST.MAX_ITER
# print(model_files)
model_files = [model_file for model_file in model_files if 'final' not in model_file and int(model_file[-11:-4])>=min_iter and int(model_file[-11:-4])<=max_iter]
else:
model_files = [cfg.MODEL.WEIGHT]
for model_file in model_files:
cfg.defrost()
cfg.MODEL.WEIGHT = model_file
logger.info('testing from {} '.format(model_file))
cfg.OUTPUT_DIR = os.path.join(ori_output_dir, model_file[-11:-4])
cfg.freeze()
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox",)
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm",)
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints",)
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
data_loaders_val = make_data_loader(cfg, is_train=False, is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(output_folders, dataset_names, data_loaders_val):
inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False, #False if cfg.MODEL.FCOS_ON or cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
stop_iter=cfg.FEW_SHOT.STOP_ITER
)
synchronize()
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Inference")
parser.add_argument(
"--config-file",
default=
"/private/home/fmassa/github/detectron.pytorch_v2/configs/e2e_faster_rcnn_R_50_C4_1x_caffe2.yaml",
metavar="FILE",
help="path to config file",
)
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--data-dir",
default=".",
metavar="DIR",
help="data dir for training",
type=str,
)
parser.add_argument(
"--out-dir",
default=".",
metavar="DIR",
help="output dir for model",
type=str,
)
parser.add_argument(
"--gpu_ids",
default="-1",
help="gpu id",
type=str,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
if args.gpu_ids != '-1':
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_ids
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
distributed = num_gpus > 1
if distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
synchronize()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.MODEL.WEIGHT = args.data_dir + cfg.MODEL.WEIGHT[1:]
cfg.DATA_DIR = args.data_dir + cfg.DATA_DIR[1:]
cfg.OUTPUT_DIR = args.out_dir + cfg.OUTPUT_DIR[1:]
print("cfg.OUTPUT_DIR: ", cfg.OUTPUT_DIR)
print("cfg.MODEL.WEIGHT: ", cfg.MODEL.WEIGHT)
print("cfg.DATA_DIR: ", cfg.DATA_DIR)
print("cfg.MODEL.ATTRIBUTE_ON: ", cfg.MODEL.ATTRIBUTE_ON)
cfg.freeze()
save_dir = ""
logger = setup_logger("maskrcnn_benchmark", save_dir, get_rank())
logger.info("Using {} GPUs".format(num_gpus))
logger.info(cfg)
logger.info("Collecting env info (might take some time)")
logger.info("\n" + collect_env_info())
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
output_folders = [None] * len(cfg.DATASETS.TEST)
dataset_names = cfg.DATASETS.TEST
if cfg.OUTPUT_DIR:
for idx, dataset_name in enumerate(dataset_names):
output_folder = os.path.join(cfg.OUTPUT_DIR, "inference",
dataset_name)
mkdir(output_folder)
output_folders[idx] = output_folder
# evaluate object detection
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_obj = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=False,
)
synchronize()
# evaluate attribute detection
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_attr = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=True,
)
synchronize()
# evaluate RPN
cfg.defrost()
cfg.MODEL.RPN_ONLY = True
cfg.freeze()
logger.info(cfg)
# pdb.set_trace()
model = build_detection_model(cfg)
model.to(cfg.MODEL.DEVICE)
output_dir = cfg.OUTPUT_DIR
checkpointer = DetectronCheckpointer(cfg, model, save_dir=output_dir)
_ = checkpointer.load(cfg.MODEL.WEIGHT)
data_loaders_val = make_data_loader(cfg,
is_train=False,
is_distributed=distributed)
for output_folder, dataset_name, data_loader_val in zip(
output_folders, dataset_names, data_loaders_val):
result_rpn = inference(
model,
data_loader_val,
dataset_name=dataset_name,
iou_types=iou_types,
box_only=False if cfg.MODEL.RETINANET_ON else cfg.MODEL.RPN_ONLY,
device=cfg.MODEL.DEVICE,
expected_results=cfg.TEST.EXPECTED_RESULTS,
expected_results_sigma_tol=cfg.TEST.EXPECTED_RESULTS_SIGMA_TOL,
output_folder=output_folder,
eval_attributes=False,
)
synchronize()
if is_main_process():
results = {**result_rpn, **result_obj, **result_attr}
print(results)
def train(cfg, local_rank, distributed, meters):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
if cfg.MODEL.USE_SYNCBN:
assert is_pytorch_1_1_0_or_later(), \
"SyncBatchNorm is only available in pytorch >= 1.1.0"
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if 'rpn' in cfg.FEW_SHOT.UNTRAINED_KEYWORD:
for param in model.rpn.parameters():
param.requires_grad = False
if 'backbone' in cfg.FEW_SHOT.UNTRAINED_KEYWORD:
for param in model.backbone.parameters():
param.requires_grad = False
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
)
arguments = {}
arguments["iteration"] = 0
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpointer = DetectronCheckpointer(cfg, model, optimizer, scheduler,
output_dir, save_to_disk)
print(cfg.MODEL.WEIGHT)
if cfg.MODEL.FSS_LOAD:
cfg.defrost()
cfg.FEW_SHOT.UNLOAD_KEYWORD = ('rpn', )
cfg.freeze()
extra_checkpoint_data = checkpointer.load(cfg.MODEL.FSS_WEIGHT)
cfg.defrost()
cfg.FEW_SHOT.UNLOAD_KEYWORD = ('backbone', 'roi_head')
cfg.freeze()
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
else:
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
if not cfg.FEW_SHOT.RESUME:
arguments["iteration"] = 0
if isinstance(meters, TensorboardLogger):
meters.iteration = arguments["iteration"]
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# torch.cuda.empty_cache()
# print(model)
do_train(
cfg,
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
meters,
)
return model
