def _load_file(self, f):
# catalog lookup
if f.startswith("catalog://"):
paths_catalog = import_file(
"detectron.config.paths_catalog", self.cfg.PATHS_CATALOG, True
)
catalog_f = paths_catalog.ModelCatalog.get(f[len("catalog://") :])
self.logger.info("{} points to {}".format(f, catalog_f))
f = catalog_f
# download url files
if f.startswith("http"):
# if the file is a url path, download it and cache it
cached_f = cache_url(f)
self.logger.info("url {} cached in {}".format(f, cached_f))
f = cached_f
# convert Caffe2 checkpoint from pkl
if f.endswith(".pkl"):
return load_c2_format(self.cfg, f)
# load native detectron.pytorch checkpoint
if self.is_train and True in self.cfg.MODEL.VOVNET.STAGE_WITH_DCN:
return load_deformable_vovnet(self.cfg, f)
else:
loaded = super(DetectronCheckpointer, self)._load_file(f)
if "model" not in loaded:
loaded = dict(model=loaded)
return loaded
def setup_custom_environment(custom_module_path):
"""Load custom environment setup from a Python source file and run the setup
function.
"""
module = import_file("detectron.utils.env.custom_module",
custom_module_path)
assert hasattr(module, "setup_environment") and callable(
module.setup_environment), (
"Custom environment module defined in {} does not have the "
"required callable attribute 'setup_environment'."
).format(custom_module_path)
module.setup_environment()
def make_data_loader(cfg, is_train=True, start_iter=0, is_for_period=False):
if is_train:
images_per_batch = cfg.SOLVER.IMS_PER_BATCH
shuffle = True
else:
images_per_batch = cfg.TEST.IMS_PER_BATCH
shuffle = False
start_iter = 0
# group images which have similar aspect ratio. In this case, we only
# group in two cases: those with width / height > 1, and the other way around,
# but the code supports more general grouping strategy
paths_catalog = import_file("detectron.config.paths_catalog",
cfg.PATHS_CATALOG, True)
DatasetCatalog = paths_catalog.DatasetCatalog
dataset_list = cfg.DATASETS.TRAIN if is_train else cfg.DATASETS.TEST
# If bbox aug is enabled in testing, simply set transforms to None and we will apply transforms later
transforms = None if not is_train and cfg.TEST.BBOX_AUG.ENABLED else build_transforms(
cfg, is_train)
datasets = build_dataset(dataset_list,
transforms,
DatasetCatalog,
is_train or is_for_period,
with_masks=cfg.MODEL.MASK_ON)
if is_train:
# save category_id to label name mapping
save_labels(datasets, cfg.OUTPUT_DIR)
data_loaders = []
for dataset in datasets:
collator = BBoxAugCollator() if not is_train and cfg.TEST.BBOX_AUG.ENABLED else \
BatchCollator(cfg.DATALOADER.SIZE_DIVISIBILITY,is_train=is_train)
num_workers = cfg.DATALOADER.NUM_WORKERS
data_loader = dataset
data_loader.collate_batch = collator
data_loader.shuffle = shuffle
data_loader.num_workers = num_workers
data_loader.batch_size = images_per_batch
data_loader.is_train = is_train
data_loaders.append(data_loader)
if is_train or is_for_period:
# during training, a single (possibly concatenated) data_loader is returned
assert len(data_loaders) == 1
return data_loaders[0]
return data_loaders
def make_data_loader(cfg, is_train=True, is_distributed=False, start_iter=0):
num_gpus = get_world_size()
if is_train:
images_per_batch = cfg.SOLVER.IMS_PER_BATCH
assert (
images_per_batch % num_gpus == 0
), "SOLVER.IMS_PER_BATCH ({}) must be divisible by the number "
"of GPUs ({}) used.".format(images_per_batch, num_gpus)
images_per_gpu = images_per_batch // num_gpus
shuffle = True
num_iters = cfg.SOLVER.MAX_ITER
else:
images_per_batch = cfg.TEST.IMS_PER_BATCH
assert (
images_per_batch % num_gpus == 0
), "TEST.IMS_PER_BATCH ({}) must be divisible by the number "
"of GPUs ({}) used.".format(images_per_batch, num_gpus)
images_per_gpu = images_per_batch // num_gpus
shuffle = False if not is_distributed else True
num_iters = None
start_iter = 0
if images_per_gpu > 1:
logger = logging.getLogger(__name__)
logger.warning(
"When using more than one image per GPU you may encounter "
"an out-of-memory (OOM) error if your GPU does not have "
"sufficient memory. If this happens, you can reduce "
"SOLVER.IMS_PER_BATCH (for training) or "
"TEST.IMS_PER_BATCH (for inference). For training, you must "
"also adjust the learning rate and schedule length according "
"to the linear scaling rule. See for example: "
"https://github.com/facebookresearch/Detectron/blob/master/configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml#L14"
)
# group images which have similar aspect ratio. In this case, we only
# group in two cases: those with width / height > 1, and the other way around,
# but the code supports more general grouping strategy
aspect_grouping = [1] if cfg.DATALOADER.ASPECT_RATIO_GROUPING else []
paths_catalog = import_file(
"detectron.config.paths_catalog", cfg.PATHS_CATALOG, True
)
DatasetCatalog = paths_catalog.DatasetCatalog
dataset_list = cfg.DATASETS.TRAIN if is_train else cfg.DATASETS.TEST
transforms = build_transforms(cfg, is_train)
datasets = build_dataset(dataset_list, transforms, DatasetCatalog, is_train)
data_loaders = []
for dataset in datasets:
sampler = make_data_sampler(dataset, shuffle, is_distributed)
batch_sampler = make_batch_data_sampler(
dataset, sampler, aspect_grouping, images_per_gpu, num_iters, start_iter
)
collator = BatchCollator(cfg.DATALOADER.SIZE_DIVISIBILITY)
num_workers = cfg.DATALOADER.NUM_WORKERS
data_loader = torch.utils.data.DataLoader(
dataset,
num_workers=num_workers,
batch_sampler=batch_sampler,
collate_fn=collator,
)
data_loaders.append(data_loader)
if is_train:
# during training, a single (possibly concatenated) data_loader is returned
assert len(data_loaders) == 1
return data_loaders[0]
return data_loaders
