def benchmark_train(args):
cfg = setup(args)
model = build_model(cfg)
logger.info("Model:\n{}".format(model))
if comm.get_world_size() > 1:
model = DistributedDataParallel(model,
device_ids=[comm.get_local_rank()],
broadcast_buffers=False)
optimizer = build_optimizer(cfg, model)
checkpointer = DetectionCheckpointer(model, optimizer=optimizer)
checkpointer.load(cfg.MODEL.WEIGHTS)
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0
data_loader = build_detection_train_loader(cfg)
dummy_data = list(itertools.islice(data_loader, 100))
def f():
data = DatasetFromList(dummy_data, copy=False)
while True:
yield from data
max_iter = 400
trainer = SimpleTrainer(model, f(), optimizer)
trainer.register_hooks([
hooks.IterationTimer(),
hooks.PeriodicWriter([CommonMetricPrinter(max_iter)])
])
trainer.train(1, max_iter)
def __init__(self, size: int):
"""
Args:
size (int): the total number of data of the underlying dataset to sample from
"""
self._size = size
assert size > 0
self._rank = comm.get_rank()
self._world_size = comm.get_world_size()
shard_size = (self._size - 1) // self._world_size + 1
begin = shard_size * self._rank
end = min(shard_size * (self._rank + 1), self._size)
self._local_indices = range(begin, end)
def default_setup(cfg, args):
"""
Perform some basic common setups at the beginning of a job, including:
1. Set up the mydl logger
2. Log basic information about environment, cmdline arguments, and config
3. Backup the config to the output directory
Args:
cfg (CfgNode): the full config to be used
args (argparse.NameSpace): the command line arguments to be logged
"""
output_dir = cfg.OUTPUT_DIR
if comm.is_main_process() and output_dir:
PathManager.mkdirs(output_dir)
rank = comm.get_rank()
setup_logger(output_dir, distributed_rank=rank, name="fvcore")
logger = setup_logger(output_dir, distributed_rank=rank)
logger.info("Rank of current process: {}. World size: {}".format(
rank, comm.get_world_size()))
logger.info("Environment info:\n" + collect_env_info())
logger.info("Command line arguments: " + str(args))
if hasattr(args, "config_file") and args.config_file != "":
logger.info("Contents of args.config_file={}:\n{}".format(
args.config_file,
PathManager.open(args.config_file, "r").read()))
logger.info("Running with full config:\n{}".format(cfg))
if comm.is_main_process() and output_dir:
# Note: some of our scripts may expect the existence of
# config.yaml in output directory
path = os.path.join(output_dir, "config.yaml")
with PathManager.open(path, "w") as f:
f.write(cfg.dump())
logger.info("Full config saved to {}".format(os.path.abspath(path)))
if cfg.VERSION == 2:
# make sure each worker has a different, yet deterministic seed if specified
seed_all_rng(None if cfg.SEED < 0 else cfg.SEED + rank)
# cudnn benchmark has large overhead. It shouldn't be used considering the small size of
# typical validation set.
if not (hasattr(args, "eval_only") and args.eval_only):
torch.backends.cudnn.benchmark = cfg.CUDNN_BENCHMARK
def main(args):
cfg = setup(args)
model = build_model(cfg)
logger.info("Model:\n{}".format(model))
if args.eval_only:
DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
cfg.MODEL.WEIGHTS, resume=args.resume
)
return do_test(cfg, model)
distributed = comm.get_world_size() > 1
if distributed:
model = DistributedDataParallel(
model, device_ids=[comm.get_local_rank()], broadcast_buffers=False
)
do_train(cfg, model)
return do_test(cfg, model)
def __init__(self,
size: int,
shuffle: bool = True,
seed: Optional[int] = None):
"""
Args:
size (int): the total number of data of the underlying dataset to sample from
shuffle (bool): whether to shuffle the indices or not
seed (int): the initial seed of the shuffle. Must be the same
across all workers. If None, will use a random seed shared
among workers (require synchronization among all workers).
"""
self._size = size
assert size > 0
self._shuffle = shuffle
if seed is None:
seed = comm.shared_random_seed()
self._seed = int(seed)
self._rank = comm.get_rank()
self._world_size = comm.get_world_size()
def reduce_loss_dict(loss_dict):
"""
Reduce the loss dictionary from all processes so that process with rank
0 has the averaged results. Returns a dict with the same fields as
loss_dict, after reduction.
"""
world_size = get_world_size()
if world_size < 2:
return loss_dict
with torch.no_grad():
loss_names = []
all_losses = []
for k in sorted(loss_dict.keys()):
loss_names.append(k)
all_losses.append(loss_dict[k])
all_losses = torch.stack(all_losses, dim=0)
dist.reduce(all_losses, dst=0)
if dist.get_rank() == 0:
# only main process gets accumulated, so only divide by
# world_size in this case
all_losses /= world_size
reduced_losses = {k: v for k, v in zip(loss_names, all_losses)}
return reduced_losses
def __init__(self, dataset_dicts, repeat_thresh, shuffle=True, seed=None):
"""
Args:
dataset_dicts (list[dict]): annotations in mydl dataset format.
repeat_thresh (float): frequency threshold below which data is repeated.
shuffle (bool): whether to shuffle the indices or not
seed (int): the initial seed of the shuffle. Must be the same
across all workers. If None, will use a random seed shared
among workers (require synchronization among all workers).
"""
self._shuffle = shuffle
if seed is None:
seed = comm.shared_random_seed()
self._seed = int(seed)
self._rank = comm.get_rank()
self._world_size = comm.get_world_size()
# Get fractional repeat factors and split into whole number (_int_part)
# and fractional (_frac_part) parts.
rep_factors = self._get_repeat_factors(dataset_dicts, repeat_thresh)
self._int_part = torch.trunc(rep_factors)
self._frac_part = rep_factors - self._int_part
def forward(self, input):
if comm.get_world_size() == 1 or not self.training:
return super().forward(input)
assert input.shape[0] > 0, "SyncBatchNorm does not support empty inputs"
C = input.shape[1]
mean = torch.mean(input, dim=[0, 2, 3])
meansqr = torch.mean(input * input, dim=[0, 2, 3])
vec = torch.cat([mean, meansqr], dim=0)
vec = AllReduce.apply(vec) * (1.0 / dist.get_world_size())
mean, meansqr = torch.split(vec, C)
var = meansqr - mean * mean
self.running_mean += self.momentum * (mean.detach() -
self.running_mean)
self.running_var += self.momentum * (var.detach() - self.running_var)
invstd = torch.rsqrt(var + self.eps)
scale = self.weight * invstd
bias = self.bias - mean * scale
scale = scale.reshape(1, -1, 1, 1)
bias = bias.reshape(1, -1, 1, 1)
return input * scale + bias
def __init__(self, cfg):
"""
Args:
cfg (CfgNode):
"""
logger = logging.getLogger("mydl")
if not logger.isEnabledFor(
logging.INFO): # setup_logger is not called for d2
setup_logger()
# Assume these objects must be constructed in this order.
model = self.build_model(cfg)
optimizer = self.build_optimizer(cfg, model)
data_loader = self.build_train_loader(cfg)
# For training, wrap with DDP. But don't need this for inference.
if comm.get_world_size() > 1:
model = DistributedDataParallel(model,
device_ids=[comm.get_local_rank()],
broadcast_buffers=False)
super().__init__(model, data_loader, optimizer)
self.scheduler = self.build_lr_scheduler(cfg, optimizer)
# Assume no other objects need to be checkpointed.
# We can later make it checkpoint the stateful hooks
self.checkpointer = DetectionCheckpointer(
# Assume you want to save checkpoints together with logs/statistics
model,
cfg.OUTPUT_DIR,
optimizer=optimizer,
scheduler=self.scheduler,
)
self.start_iter = 0
self.max_iter = cfg.SOLVER.MAX_ITER
self.cfg = cfg
self.register_hooks(self.build_hooks())
def do_train(
cfg,
model,
data_loader,
data_loader_val,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
test_period,
arguments,
):
logger = logging.getLogger("mydl.trainer")
logger.info("Start training")
meters = MetricLogger(delimiter=" ")
max_iter = len(data_loader)
start_iter = arguments["iteration"]
model.train()
start_training_time = time.time()
end = time.time()
iou_types = ("bbox", )
if cfg.MODEL.MASK_ON:
iou_types = iou_types + ("segm", )
if cfg.MODEL.KEYPOINT_ON:
iou_types = iou_types + ("keypoints", )
dataset_names = cfg.DATASETS.TEST
for iteration, (images, targets, _) in enumerate(data_loader, start_iter):
if any(len(target) < 1 for target in targets):
logger.error(
f"Iteration={iteration + 1} || Image Ids used for training {_} || targets Length={[len(target) for target in targets]}"
)
continue
data_time = time.time() - end
iteration = iteration + 1
arguments["iteration"] = iteration
images = images.to(device)
targets = [target.to(device) for target in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
# reduce losses over all GPUs for logging purposes
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
meters.update(loss=losses_reduced, **loss_dict_reduced)
optimizer.zero_grad()
# Note: If mixed precision is not used, this ends up doing nothing
# Otherwise apply loss scaling for mixed-precision recipe
with amp.scale_loss(losses, optimizer) as scaled_losses:
scaled_losses.backward()
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
meters.update(time=batch_time, data=data_time)
eta_seconds = meters.time.global_avg * (max_iter - iteration)
eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
if iteration % 20 == 0 or iteration == max_iter:
logger.info(
meters.delimiter.join([
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if iteration % checkpoint_period == 0:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if data_loader_val is not None and test_period > 0 and iteration % test_period == 0:
meters_val = MetricLogger(delimiter=" ")
synchronize()
_ = inference( # The result can be used for additional logging, e. g. for TensorBoard
model,
# The method changes the segmentation mask format in a data loader,
# so every time a new data loader is created:
make_data_loader(cfg,
is_train=False,
is_distributed=(get_world_size() > 1),
is_for_period=True),
dataset_name="[Validation]",
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=None,
)
synchronize()
model.train()
with torch.no_grad():
# Should be one image for each GPU:
for iteration_val, (images_val, targets_val,
_) in enumerate(tqdm(data_loader_val)):
images_val = images_val.to(device)
targets_val = [target.to(device) for target in targets_val]
loss_dict = model(images_val, targets_val)
losses = sum(loss for loss in loss_dict.values())
loss_dict_reduced = reduce_loss_dict(loss_dict)
losses_reduced = sum(
loss for loss in loss_dict_reduced.values())
meters_val.update(loss=losses_reduced, **loss_dict_reduced)
synchronize()
logger.info(
meters_val.delimiter.join([
"[Validation]: ",
"eta: {eta}",
"iter: {iter}",
"{meters}",
"lr: {lr:.6f}",
"max mem: {memory:.0f}",
]).format(
eta=eta_string,
iter=iteration,
meters=str(meters_val),
lr=optimizer.param_groups[0]["lr"],
memory=torch.cuda.max_memory_allocated() / 1024.0 / 1024.0,
))
if iteration == max_iter:
checkpointer.save("model_final", **arguments)
total_training_time = time.time() - start_training_time
total_time_str = str(datetime.timedelta(seconds=total_training_time))
logger.info("Total training time: {} ({:.4f} s / it)".format(
total_time_str, total_training_time / (max_iter)))
def make_data_loader(cfg,
is_train=True,
is_distributed=False,
start_iter=0,
is_for_period=False):
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("mydl.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)
if is_train:
# save category_id to label name mapping
save_labels(datasets, cfg.OUTPUT_DIR)
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 = BBoxAugCollator() if not is_train and cfg.TEST.BBOX_AUG.ENABLED else \
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 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 build_detection_train_loader(cfg, mapper=None):
"""
A data loader is created by the following steps:
1. Use the dataset names in config to query :class:`DatasetCatalog`, and obtain a list of dicts.
2. Start workers to work on the dicts. Each worker will:
* Map each metadata dict into another format to be consumed by the model.
* Batch them by simply putting dicts into a list.
The batched ``list[mapped_dict]`` is what this dataloader will return.
Args:
cfg (CfgNode): the config
mapper (callable): a callable which takes a sample (dict) from dataset and
returns the format to be consumed by the model.
By default it will be `DatasetMapper(cfg, True)`.
Returns:
an infinite iterator of training data
"""
num_workers = get_world_size()
images_per_batch = cfg.SOLVER.IMS_PER_BATCH
assert (
images_per_batch % num_workers == 0
), "SOLVER.IMS_PER_BATCH ({}) must be divisible by the number of workers ({}).".format(
images_per_batch, num_workers
)
assert (
images_per_batch >= num_workers
), "SOLVER.IMS_PER_BATCH ({}) must be larger than the number of workers ({}).".format(
images_per_batch, num_workers
)
images_per_worker = images_per_batch // num_workers
# force the training of algorithm without remove image that have not annotations : False
annotation_filter_emtpy = True
if cfg.VERSION == 1:
annotation_filter_emtpy = False
else:
annotation_filter_emtpy = cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS
dataset_dicts = get_detection_dataset_dicts(
cfg.DATASETS.TRAIN,
filter_empty=annotation_filter_emtpy,
min_keypoints=cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE
if cfg.VERSION != 1 and cfg.MODEL.KEYPOINT_ON
else 0,
proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None,
)
dataset = DatasetFromList(dataset_dicts, copy=False)
if mapper is None:
mapper = DatasetMapper(cfg, True)
dataset = MapDataset(dataset, mapper)
sampler_name = cfg.DATALOADER.SAMPLER_TRAIN
logger = logging.getLogger(__name__)
logger.info("Using training sampler {}".format(sampler_name))
if sampler_name == "TrainingSampler":
sampler = samplers.TrainingSampler(len(dataset))
elif sampler_name == "RepeatFactorTrainingSampler":
sampler = samplers.RepeatFactorTrainingSampler(
dataset_dicts, cfg.DATALOADER.REPEAT_THRESHOLD
)
else:
raise ValueError("Unknown training sampler: {}".format(sampler_name))
if cfg.DATALOADER.ASPECT_RATIO_GROUPING:
data_loader = torch.utils.data.DataLoader(
dataset,
sampler=sampler,
num_workers=cfg.DATALOADER.NUM_WORKERS,
batch_sampler=None,
collate_fn=operator.itemgetter(0), # don't batch, but yield individual elements
worker_init_fn=worker_init_reset_seed,
) # yield individual mapped dict
data_loader = AspectRatioGroupedDataset(data_loader, images_per_worker)
else:
batch_sampler = torch.utils.data.sampler.BatchSampler(
sampler, images_per_worker, drop_last=True
)
# drop_last so the batch always have the same size
data_loader = torch.utils.data.DataLoader(
dataset,
num_workers=cfg.DATALOADER.NUM_WORKERS,
batch_sampler=batch_sampler,
collate_fn=trivial_batch_collator,
worker_init_fn=worker_init_reset_seed,
)
return data_loader
def load_cityscapes_instances(image_dir,
gt_dir,
from_json=True,
to_polygons=True):
"""
Args:
image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train".
gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train".
from_json (bool): whether to read annotations from the raw json file or the png files.
to_polygons (bool): whether to represent the segmentation as polygons
(COCO's format) instead of masks (cityscapes's format).
Returns:
list[dict]: a list of dicts in mydl standard format. (See
`Using Custom Datasets </tutorials/datasets.html>`_ )
"""
if from_json:
assert to_polygons, (
"Cityscapes's json annotations are in polygon format. "
"Converting to mask format is not supported now.")
files = []
for image_file in glob.glob(os.path.join(image_dir, "**/*.png")):
suffix = "leftImg8bit.png"
assert image_file.endswith(suffix)
prefix = image_dir
instance_file = gt_dir + image_file[
len(prefix):-len(suffix)] + "gtFine_instanceIds.png"
assert os.path.isfile(instance_file), instance_file
label_file = gt_dir + image_file[
len(prefix):-len(suffix)] + "gtFine_labelIds.png"
assert os.path.isfile(label_file), label_file
json_file = gt_dir + image_file[
len(prefix):-len(suffix)] + "gtFine_polygons.json"
files.append((image_file, instance_file, label_file, json_file))
assert len(files), "No images found in {}".format(image_dir)
logger = logging.getLogger(__name__)
logger.info("Preprocessing cityscapes annotations ...")
# This is still not fast: all workers will execute duplicate works and will
# take up to 10m on a 8GPU server.
pool = mp.Pool(processes=max(mp.cpu_count() // get_world_size() // 2, 4))
ret = pool.map(
functools.partial(cityscapes_files_to_dict,
from_json=from_json,
to_polygons=to_polygons),
files,
)
logger.info("Loaded {} images from {}".format(len(ret), image_dir))
# Map cityscape ids to contiguous ids
from cityscapesscripts.helpers.labels import labels
labels = [l for l in labels if l.hasInstances and not l.ignoreInEval]
dataset_id_to_contiguous_id = {l.id: idx for idx, l in enumerate(labels)}
for dict_per_image in ret:
for anno in dict_per_image["annotations"]:
anno["category_id"] = dataset_id_to_contiguous_id[
anno["category_id"]]
return ret