def setup_logging(output_dir=None):
"""
Sets up the logging for multiple processes. Only enable the logging for the
master process, and suppress logging for the non-master processes.
"""
# Set up logging format.
_FORMAT = "[%(levelname)s: %(filename)s: %(lineno)4d]: %(message)s"
if du.is_master_proc():
# Enable logging for the master process.
logging.root.handlers = []
else:
# Suppress logging for non-master processes.
_suppress_print()
logger = logging.getLogger()
logger.setLevel(logging.DEBUG)
logger.propagate = False
plain_formatter = logging.Formatter(
"[%(asctime)s][%(levelname)s] %(filename)s: %(lineno)3d: %(message)s",
datefmt="%m/%d %H:%M:%S",
)
if du.is_master_proc():
ch = logging.StreamHandler(stream=sys.stdout)
ch.setLevel(logging.DEBUG)
ch.setFormatter(plain_formatter)
logger.addHandler(ch)
if output_dir is not None and du.is_master_proc(du.get_world_size()):
filename = os.path.join(output_dir, "stdout.log")
fh = logging.StreamHandler(_cached_log_stream(filename))
fh.setLevel(logging.DEBUG)
fh.setFormatter(plain_formatter)
logger.addHandler(fh)
def test(cfg):
"""
Perform multi-view testing on the pretrained video model.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
assert (len(test_loader.dataset) %
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS) == 0)
# Create meters for multi-view testing.
test_meter = TestMeter(
len(test_loader.dataset) //
(cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# Set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# # Perform multi-view test on the entire dataset.
test_meter = perform_test(test_loader, model, test_meter, cfg, writer)
if writer is not None:
writer.close()
def save_checkpoint(path_to_job, model, optimizer, epoch, cfg):
"""
Save a checkpoint.
Args:
model (model): model to save the weight to the checkpoint.
optimizer (optim): optimizer to save the historical state.
epoch (int): current number of epoch of the model.
cfg (CfgNode): configs to save.
"""
# Save checkpoints only from the master process.
if not du.is_master_proc(cfg.NUM_GPUS * cfg.NUM_SHARDS):
return
# Ensure that the checkpoint dir exists.
g_pathmgr.mkdirs(get_checkpoint_dir(path_to_job))
# Omit the DDP wrapper in the multi-gpu setting.
sd = model.module.state_dict() if cfg.NUM_GPUS > 1 else model.state_dict()
normalized_sd = sub_to_normal_bn(sd)
# Record the state.
checkpoint = {
"epoch": epoch,
"model_state": normalized_sd,
"optimizer_state": optimizer.state_dict(),
"cfg": cfg.dump(),
}
# Write the checkpoint.
path_to_checkpoint = get_path_to_checkpoint(path_to_job, epoch + 1)
with g_pathmgr.open(path_to_checkpoint, "wb") as f:
torch.save(checkpoint, f)
return path_to_checkpoint
def build_trainer(cfg):
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
optimizer = optim.construct_optimizer(model, cfg)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = loader.construct_loader(cfg,
"train",
is_precise_bn=True)
# Create meters.
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
return (
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
)
def make_checkpoint_dir(path_to_job):
"""
Creates the checkpoint directory (if not present already).
Args:
path_to_job (string): the path to the folder of the current job.
"""
checkpoint_dir = os.path.join(path_to_job, "checkpoints")
# Create the checkpoint dir from the master process
if du.is_master_proc() and not g_pathmgr.exists(checkpoint_dir):
try:
g_pathmgr.mkdirs(checkpoint_dir)
except Exception:
pass
return checkpoint_dir
def build_trainer(cfg):
"""
Build training model and its associated tools, including optimizer,
dataloaders and meters.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
Returns:
model (nn.Module): training model.
optimizer (Optimizer): optimizer.
train_loader (DataLoader): training data loader.
val_loader (DataLoader): validatoin data loader.
precise_bn_loader (DataLoader): training data loader for computing
precise BN.
train_meter (TrainMeter): tool for measuring training stats.
val_meter (ValMeter): tool for measuring validation stats.
"""
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = loader.construct_loader(cfg,
"train",
is_precise_bn=True)
# Create meters.
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
return (
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
)
def test(cfg):
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test with config:")
logger.info(cfg)
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=False)
cu.load_test_checkpoint(cfg, model)
# Create video testing loaders.
test_loader = loader.construct_loader(cfg, "test")
logger.info("Testing model for {} iterations".format(len(test_loader)))
assert (
test_loader.dataset.num_videos
% (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS)
== 0
)
# Create meters for multi-view testing.
test_meter = TestMeter(
test_loader.dataset.num_videos
// (cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS),
cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS,
cfg.MODEL.NUM_CLASSES,
len(test_loader),
cfg.DATA.MULTI_LABEL,
cfg.DATA.ENSEMBLE_METHOD,
)
# # Perform multi-view test on the entire dataset.
test_meter = perform_test(test_loader, model, test_meter, cfg)
def train(cfg):
"""
Train a video model for many epochs on train set and evaluate it on val set.
Args:
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Init multigrid.
multigrid = None
if cfg.MULTIGRID.LONG_CYCLE or cfg.MULTIGRID.SHORT_CYCLE:
multigrid = MultigridSchedule()
cfg = multigrid.init_multigrid(cfg)
if cfg.MULTIGRID.LONG_CYCLE:
cfg, _ = multigrid.update_long_cycle(cfg, cur_epoch=0)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if not cfg.TRAIN.FINETUNE:
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
else:
start_epoch = 0
cu.load_checkpoint(cfg.TRAIN.CHECKPOINT_FILE_PATH, model)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(
cfg, "train", is_precise_bn=True)
if cfg.BN.USE_PRECISE_STATS else None)
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# set up writer for logging to Tensorboard format.
if cfg.TENSORBOARD.ENABLE and du.is_master_proc(
cfg.NUM_GPUS * cfg.NUM_SHARDS):
writer = tb.TensorboardWriter(cfg)
else:
writer = None
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
if cfg.MULTIGRID.LONG_CYCLE:
cfg, changed = multigrid.update_long_cycle(cfg, cur_epoch)
if changed:
(
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
) = build_trainer(cfg)
# Load checkpoint.
if cu.has_checkpoint(cfg.OUTPUT_DIR):
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
assert "{:05d}.pyth".format(cur_epoch) in last_checkpoint
else:
last_checkpoint = cfg.TRAIN.CHECKPOINT_FILE_PATH
logger.info("Load from {}".format(last_checkpoint))
cu.load_checkpoint(last_checkpoint, model, cfg.NUM_GPUS > 1,
optimizer)
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg, writer)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None if multigrid is None else multigrid.schedule,
)
is_eval_epoch = misc.is_eval_epoch(
cfg, cur_epoch, None if multigrid is None else multigrid.schedule)
# Compute precise BN stats.
if ((is_checkp_epoch or is_eval_epoch) and cfg.BN.USE_PRECISE_STATS
and len(get_bn_modules(model)) > 0):
calculate_and_update_precise_bn(
precise_bn_loader,
model,
min(cfg.BN.NUM_BATCHES_PRECISE, len(precise_bn_loader)),
cfg.NUM_GPUS > 0,
)
_ = misc.aggregate_sub_bn_stats(model)
# Save a checkpoint.
if is_checkp_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if is_eval_epoch:
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, writer)
if writer is not None:
writer.close()
def train(cfg):
# Set up environment.
du.init_distributed_training(cfg)
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# if True, build apex model and optimizer.
if cfg.TRAIN.ENABLE_APEX:
assert (cfg.NUM_GPUS <= torch.cuda.device_count()
), "Cannot use more GPU devices than available"
# Construct the model
from lib.models import MODEL_REGISTRY
name = cfg.MODEL.MODEL_NAME
model = MODEL_REGISTRY.get(name)(cfg)
import apex
from apex import amp
# using apex synced BN
model = apex.parallel.convert_syncbn_model(model)
# Determine the GPU used by the current process
cur_device = torch.cuda.current_device()
# Transfer the model to the current GPU device
model = model.cuda(device=cur_device)
optimizer = optim.construct_optimizer(model, cfg)
# initialize amp
model, optimizer = amp.initialize(model,
optimizer,
opt_level=cfg.TRAIN.APEX_OPT_LEVEL)
# Use multi-process data parallel model in the multi-gpu setting
if cfg.NUM_GPUS > 1:
# Make model replica operate on the current device
model = apex.parallel.DistributedDataParallel(model,
delay_allreduce=True)
# model = torch.nn.parallel.DistributedDataParallel(
# module=model, device_ids=[cur_device], output_device=cur_device)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
else:
# Build the video model and print model statistics.
model = build_model(cfg)
if du.is_master_proc() and cfg.LOG_MODEL_INFO:
misc.log_model_info(model, cfg, use_train_input=True)
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(
cfg, "train", is_precise_bn=True)
if cfg.BN.USE_PRECISE_STATS else None)
train_meter = TrainMeter(len(train_loader), cfg)
val_meter = ValMeter(len(val_loader), cfg)
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
epoch_timer = EpochTimer()
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Shuffle the dataset.
loader.shuffle_dataset(train_loader, cur_epoch)
# Train for one epoch.
epoch_timer.epoch_tic()
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg)
epoch_timer.epoch_toc()
logger.info(
f"Epoch {cur_epoch} takes {epoch_timer.last_epoch_time():.2f}s. Epochs "
f"from {start_epoch} to {cur_epoch} take "
f"{epoch_timer.avg_epoch_time():.2f}s in average and "
f"{epoch_timer.median_epoch_time():.2f}s in median.")
logger.info(
f"For epoch {cur_epoch}, each iteraction takes "
f"{epoch_timer.last_epoch_time()/len(train_loader):.2f}s in average. "
f"From epoch {start_epoch} to {cur_epoch}, each iteraction takes "
f"{epoch_timer.avg_epoch_time()/len(train_loader):.2f}s in average."
)
is_checkp_epoch = cu.is_checkpoint_epoch(cfg, cur_epoch, None)
is_eval_epoch = misc.is_eval_epoch(cfg, cur_epoch, None)
# Compute precise BN stats.
if ((is_checkp_epoch or is_eval_epoch) and cfg.BN.USE_PRECISE_STATS
and len(get_bn_modules(model)) > 0):
calculate_and_update_precise_bn(
precise_bn_loader,
model,
min(cfg.BN.NUM_BATCHES_PRECISE, len(precise_bn_loader)),
cfg.NUM_GPUS > 0,
)
_ = misc.aggregate_sub_bn_stats(model)
# Save a checkpoint.
if is_checkp_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if is_eval_epoch:
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg)
def run_evaluation(
categories, groundtruth, detections, excluded_keys, verbose=True
):
"""AVA evaluation main logic."""
pascal_evaluator = object_detection_evaluation.PascalDetectionEvaluator(
categories
)
boxes, labels, _ = groundtruth
gt_keys = []
pred_keys = []
for image_key in boxes:
if image_key in excluded_keys:
logging.info(
(
"Found excluded timestamp in ground truth: %s. "
"It will be ignored."
),
image_key,
)
continue
pascal_evaluator.add_single_ground_truth_image_info(
image_key,
{
standard_fields.InputDataFields.groundtruth_boxes: np.array(
boxes[image_key], dtype=float
),
standard_fields.InputDataFields.groundtruth_classes: np.array(
labels[image_key], dtype=int
),
standard_fields.InputDataFields.groundtruth_difficult: np.zeros(
len(boxes[image_key]), dtype=bool
),
},
)
gt_keys.append(image_key)
boxes, labels, scores = detections
for image_key in boxes:
if image_key in excluded_keys:
logging.info(
(
"Found excluded timestamp in detections: %s. "
"It will be ignored."
),
image_key,
)
continue
pascal_evaluator.add_single_detected_image_info(
image_key,
{
standard_fields.DetectionResultFields.detection_boxes: np.array(
boxes[image_key], dtype=float
),
standard_fields.DetectionResultFields.detection_classes: np.array(
labels[image_key], dtype=int
),
standard_fields.DetectionResultFields.detection_scores: np.array(
scores[image_key], dtype=float
),
},
)
pred_keys.append(image_key)
metrics = pascal_evaluator.evaluate()
if du.is_master_proc():
pprint.pprint(metrics, indent=2)
return metrics
