def build_hooks(self):
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
cfg.TEST.EVAL_PERIOD,
self.model,
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
if comm.is_main_process():
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
return ret
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
# Put `PeriodicDump` after writers so that can dump all the files,
# including the files generated by writers
if cfg.OUTPUT_DUMP.ENABLED:
period = cfg.OUTPUT_DUMP.PERIOD
assert period >= cfg.SOLVER.CHECKPOINT_PERIOD, (
"Dump period ({}) must large than checkpoint period ({}).".
format(period, cfg.SOLVER.CHECKPOINT_PERIOD))
platforms = cfg.OUTPUT_DUMP.PLATFORMS
buckets = cfg.OUTPUT_DUMP.BUCKETS
ret.append(
hooks.PeriodicDump(cfg.OUTPUT_DIR, period, platforms,
buckets))
return ret
def __init__(self, period, model, data_loader, num_iter):
"""
Args:
period (int): the period this hook is run, or 0 to not run during training.
The hook will always run in the end of training.
model (nn.Module): a module whose all BN layers in training mode will be
updated by precise BN.
Note that user is responsible for ensuring the BN layers to be
updated are in training mode when this hook is triggered.
data_loader (iterable): it will produce data to be run by `model(data)`.
num_iter (int): number of iterations used to compute the precise
statistics.
"""
self._logger = logging.getLogger(__name__)
if len(get_bn_modules(model)) == 0:
self._logger.info(
"PreciseBN is disabled because model does not contain BN layers in training mode."
)
self._disabled = True
return
self._model = model
self._data_loader = data_loader
self._num_iter = num_iter
self._period = period
self._disabled = False
self._data_iter = None
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
## insert custom hook
ret.insert(
-1,
LossEvalHook(
self.cfg.TEST.EVAL_PERIOD, self.model,
build_detection_test_loader(self.cfg,
self.cfg.DATASETS.TEST[0],
DatasetMapper(self.cfg, True))))
## change dataset index (custom)
cfg.Test_index += 1
if cfg.Test_index == cfg.Test_index_MAX:
cfg.Test_index = 0
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
if cfg.SOLVER.CHECKPOINT_BY_EPOCH:
ckpt_period = cfg.SOLVER.CHECKPOINT_PERIOD * self.iters_per_epoch
else:
ckpt_period = cfg.SOLVER.CHECKPOINT_PERIOD
ret.append(
MyPeriodicCheckpointer(self.checkpointer,
ckpt_period,
max_to_keep=cfg.SOLVER.get(
"NUM_CKPT_KEEP", 5),
iters_per_epoch=self.iters_per_epoch))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(
hooks.PeriodicWriter(self.build_writers(),
period=cfg.TRAIN.get("PRINT_FREQ", 100)))
return ret
def build_hooks(self):
"""
Build a list of default hooks.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers()))
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(
hooks.PeriodicWriter([TensorboardXWriter(self.cfg.OUTPUT_DIR)
]))
return ret
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
hooks.GradcamHook(
# which layer need to visualize
cfg.VISUALIZED.LAYER_NAMES,
# visualization of period
cfg.VIS_PERIOD,
) if cfg.VIS_PERIOD > 0 else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# Here the default print/log frequency of each writer is used.
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def build_hooks(self):
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
)
if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(
self.checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD
)
)
def test_and_save_results_student():
self._last_eval_results_student = self.test(self.cfg, self.model)
_last_eval_results_student = {
k + "_student": self._last_eval_results_student[k]
for k in self._last_eval_results_student.keys()
}
return _last_eval_results_student
def test_and_save_results_teacher():
self._last_eval_results_teacher = self.test(
self.cfg, self.model_teacher)
return self._last_eval_results_teacher
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD,
test_and_save_results_student))
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD,
test_and_save_results_teacher))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone() # cfg is a copy, will be revised
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(), # 用于计算step, global time 的
hooks.LRScheduler(), # 执行学习率 step()
hooks.PreciseBN( # 用于更新精确的BN 的信息
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD, # 7330
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER, # 200
)
# TEST.PRECISE_BN.ENABLED: defualt:False
if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model) else
None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(self.cfg, self.model)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
# eval_period: 7330
# to be read...2021.2.26
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def __init__(self, period, model, data_loader, num_iter):
self._logger = logging.getLogger(__name__)
if len(get_bn_modules(model)) == 0:
self._logger.info(
"PreciseBN is disabled because model does not contain BN layers in training mode."
)
self._disabled = True
return
self._model = model
self._data_loader = data_loader
self._num_iter = num_iter
self._period = period
self._disabled = False
self._data_iter = None
def train(cfg):
# Build model
model = build_model(cfg)
optimizer = optim.construct_optimizer(model, cfg)
# load checkpoint
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Build data loader
train_loader = dataloader.construct_loader(cfg, "train")
val_loader = dataloader.construct_loader(cfg, "val")
precise_bn_loader = dataloader.construct_loader(cfg, "train")
best_accuracy = 0
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
# Train for one epoch.
train_epoch(train_loader, model, optimizer, cur_epoch, cfg)
is_eval_epoch = cur_epoch > 0
# Compute precise BN stats.
if (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,
)
_ = aggregate_sub_bn_stats(
model) # for SubBatchNorm3d call before eval
# Evaluate the model on validation set.
if is_eval_epoch:
results = eval_epoch(val_loader, model, cur_epoch, cfg)
accuracy = results['top1']
if accuracy > best_accuracy:
print("*** Saving best ****")
best_accuracy = accuracy
torch.save(
{
'epoch': cur_epoch + 1,
'model_state': model.state_dict(),
'optimizer_state': optimizer.state_dict()
}, os.path.join(cfg.OUTPUT_DIR, 'best_ckpt.pth'))
def build_hooks(self):
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
)
if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
if comm.is_main_process():
ret.append(hooks.PeriodicCheckpointer(self.checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
res = self._last_eval_results = self.test(self.cfg, self.model)
eval_dir = os.path.join(self.cfg.OUTPUT_DIR, 'evals')
os.makedirs(eval_dir, exist_ok=True)
pd.DataFrame(res).to_csv(os.path.join(eval_dir, f'{self.round}.csv'))
return self._last_eval_results
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
def build_hooks(self):
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
# def test_and_save_results():
# self._last_eval_results = self.test(self.cfg, self.model)
# return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
# ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
return ret
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.
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)
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
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):
"""
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 random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = model_builder.build_model(cfg)
if du.is_master_proc():
misc.log_model_info(model)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model,
cfg.NUM_GPUS > 1, optimizer)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
# Create the video train and val loaders.
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
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))
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.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg)
# Compute precise BN stats.
if cfg.BN.USE_PRECISE_STATS and len(get_bn_modules(model)) > 0:
calculate_and_update_precise_bn(train_loader, model,
cfg.BN.NUM_BATCHES_PRECISE)
# Save a checkpoint.
if cu.is_checkpoint_epoch(cur_epoch, cfg.TRAIN.CHECKPOINT_PERIOD):
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if misc.is_eval_epoch(cfg, cur_epoch):
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg)
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 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)
# Create a GradScaler for mixed precision training
scaler = torch.cuda.amp.GradScaler(enabled=cfg.TRAIN.MIXED_PRECISION)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR, task=cfg.TASK)
if last_checkpoint is not None:
checkpoint_epoch = cu.load_checkpoint(
last_checkpoint,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
)
start_epoch = checkpoint_epoch + 1
elif "ssl_eval" in cfg.TASK:
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR,
task="ssl")
checkpoint_epoch = cu.load_checkpoint(
last_checkpoint,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
epoch_reset=True,
clear_name_pattern=cfg.TRAIN.CHECKPOINT_CLEAR_NAME_PATTERN,
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "":
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
epoch_reset=cfg.TRAIN.CHECKPOINT_EPOCH_RESET,
clear_name_pattern=cfg.TRAIN.CHECKPOINT_CLEAR_NAME_PATTERN,
)
start_epoch = checkpoint_epoch + 1
else:
start_epoch = 0
# 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)
# if (
# cfg.TASK == "ssl"
# and cfg.MODEL.MODEL_NAME == "ContrastiveModel"
# and cfg.CONTRASTIVE.KNN_ON
# ):
# if hasattr(model, "module"):
# model.module.init_knn_labels(train_loader)
# else:
# model.init_knn_labels(train_loader)
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
train_meter = TrainMeter(1e6, cfg)
val_meter = ValMeter(1e6, 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))
epoch_timer = EpochTimer()
for cur_epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCH):
if cur_epoch > 0 and cfg.DATA.LOADER_CHUNK_SIZE > 0:
num_chunks = math.ceil(cfg.DATA.LOADER_CHUNK_OVERALL_SIZE /
cfg.DATA.LOADER_CHUNK_SIZE)
skip_rows = (cur_epoch) % num_chunks * cfg.DATA.LOADER_CHUNK_SIZE
logger.info(
f"=================+++ num_chunks {num_chunks} skip_rows {skip_rows}"
)
cfg.DATA.SKIP_ROWS = skip_rows
logger.info(f"|===========| skip_rows {skip_rows}")
train_loader = loader.construct_loader(cfg, "train")
loader.shuffle_dataset(train_loader, cur_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,
task=cfg.TASK)
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)
if hasattr(train_loader.dataset, "_set_epoch_num"):
train_loader.dataset._set_epoch_num(cur_epoch)
# Train for one epoch.
epoch_timer.epoch_tic()
train_epoch(
train_loader,
model,
optimizer,
scaler,
train_meter,
cur_epoch,
cfg,
writer,
)
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 if multigrid is None else multigrid.schedule,
) or cur_epoch == cfg.SOLVER.MAX_EPOCH - 1)
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,
scaler if cfg.TRAIN.MIXED_PRECISION else None,
)
# Evaluate the model on validation set.
if is_eval_epoch:
eval_epoch(
val_loader,
model,
val_meter,
cur_epoch,
cfg,
train_loader,
writer,
)
if writer is not None:
writer.close()
result_string = "Top1 Acc: {:.2f} Top5 Acc: {:.2f} MEM: {:.2f}" "".format(
100 - val_meter.min_top1_err,
100 - val_meter.min_top5_err,
misc.gpu_mem_usage(),
)
logger.info("training done: {}".format(result_string))
return result_string
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = [
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg),
cfg.TEST.PRECISE_BN.NUM_ITER,
) if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
self._last_eval_results = self.test(
self.cfg,
self.model,
evaluators=[
LOFAREvaluator(
t,
cfg.OUTPUT_DIR,
sigmabox=cfg.SIGMABOX,
segmentation_dir=
f'/data1/mostertrij/data/cache/segmentation_maps_{cfg.TEST.REMOVE_THRESHOLD}',
remove_unresolved=cfg.TEST.REMOVE_UNRESOLVED)
for t in self.cfg.DATASETS.TEST
])
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(
hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results,
cfg.TEST.EXTRA_EVAL))
if comm.is_main_process():
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers()))
return ret
def train(cfg):
"""
Train an audio 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)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the audio 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)
if cfg.BN.FREEZE:
model.module.freeze_fn(
'bn_parameters') if cfg.NUM_GPUS > 1 else model.freeze_fn(
'bn_parameters')
# Construct the optimizer.
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 audio train and val loaders.
if cfg.TRAIN.DATASET != 'epickitchens' or not cfg.EPICKITCHENS.TRAIN_PLUS_VAL:
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(cfg, "train")
if cfg.BN.USE_PRECISE_STATS else None)
else:
train_loader = loader.construct_loader(cfg, "train+val")
val_loader = loader.construct_loader(cfg, "val")
precise_bn_loader = (loader.construct_loader(cfg, "train+val")
if cfg.BN.USE_PRECISE_STATS else None)
# Create meters.
if cfg.TRAIN.DATASET == 'epickitchens':
train_meter = EPICTrainMeter(len(train_loader), cfg)
val_meter = EPICValMeter(len(val_loader), cfg)
else:
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
if cfg.WANDB.ENABLE and du.is_master_proc(cfg.NUM_GPUS * cfg.NUM_SHARDS):
wandb_log = True
if cfg.TRAIN.AUTO_RESUME and cfg.WANDB.RUN_ID != "":
wandb.init(project='slowfast',
config=cfg,
sync_tensorboard=True,
resume=cfg.WANDB.RUN_ID)
else:
wandb.init(project='slowfast', config=cfg, sync_tensorboard=True)
wandb.watch(model)
else:
wandb_log = False
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
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.
train_epoch(train_loader, model, optimizer, train_meter, cur_epoch,
cfg, writer, wandb_log)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
)
is_eval_epoch = misc.is_eval_epoch(
cfg,
cur_epoch,
)
# 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:
is_best_epoch, _ = eval_epoch(val_loader, model, val_meter,
cur_epoch, cfg, writer, wandb_log)
if is_best_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR,
model,
optimizer,
cur_epoch,
cfg,
is_best_epoch=is_best_epoch)
if writer is not None:
writer.close()
def build_hooks(self):
"""
Build a list of default hooks, including timing, evaluation,
checkpointing, lr scheduling, precise BN, writing events.
Returns:
list[HookBase]:
"""
cfg = self.cfg.clone()
cfg.defrost()
cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN
ret = \
[
hooks.IterationTimer(),
hooks.LRScheduler(self.optimizer, self.scheduler),
hooks.PreciseBN(
# Run at the same freq as (but before) evaluation.
cfg.TEST.EVAL_PERIOD,
self.model,
# Build a new data loader to not affect training
self.build_train_loader(cfg,self.mapper_object,self.isShuffleData),
cfg.TEST.PRECISE_BN.NUM_ITER,
)
if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
else None,
]
# Do PreciseBN before checkpointer, because it updates the model and need to
# be saved by checkpointer.
# This is not always the best: if checkpointing has a different frequency,
# some checkpoints may have more precise statistics than others.
if comm.is_main_process():
ret.append(
hooks.PeriodicCheckpointer(self.checkpointer,
cfg.SOLVER.CHECKPOINT_PERIOD))
def test_and_save_results():
# self._last_eval_results = self.test(self.cfg, self.model, self.isTrackAccuracy, self.getter, self.dataset_used)
self._last_eval_results = self.test(self.cfg,
self.model,
self.getter,
self.dataset_used,
self.mapper_object,
evaluators=None)
return self._last_eval_results
# Do evaluation after checkpointer, because then if it fails,
# we can use the saved checkpoint to debug.
ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
if comm.is_main_process():
numberOfSamples = 25
step = -1
if (self.max_iter <= numberOfSamples):
# Eg, maxiter = 20, so step = 20/2 = 10, take a sample every 10
step = int(round(float(self.max_iter) / float(2), 2))
else:
# Eg 10000/20 = 500, so will take a sample every 500 iterations
step = float(self.max_iter) / float(numberOfSamples)
step = int(round(step, 0))
if (step < 1): step = 1
# print("!!!!!!!!!!!!!!STEPS: ", step)
# ret.append(hooks.PeriodicWriter(self.build_writers()))
# run writers in the end, so that evaluation metrics are written
ret.append(hooks.PeriodicWriter(self.build_writers(), period=step))
# ret.append(hooks.PeriodicWriter(self.build_writers(),period=(self.max_iter-1)))
return ret
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 random seed from configs.
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Setup logging format.
logging.setup_logging()
if du.is_master_proc():
if cfg.ENABLE_WANDB:
wandb.login()
wandb.init(project='bbox', entity='slowfast')
# 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 cfg.EPICKITCHENS.USE_BBOX and not cu.has_checkpoint(cfg.OUTPUT_DIR):
slow_fast_model = SlowFast(cfg)
if cfg.EPICKITCHENS.USE_BBOX and cfg.EPICKITCHENS.LOAD_SLOWFAST_PRETRAIN:
if cfg.EPICKITCHENS.SLOWFAST_PRETRAIN_CHECKPOINT_FILE_PATH != "":
_ = cu.load_checkpoint(
cfg.EPICKITCHENS.SLOWFAST_PRETRAIN_CHECKPOINT_FILE_PATH,
slow_fast_model,
False,
optimizer=None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=False,
)
# cfg.TRAIN.CHECKPOINT_TYPE == "caffe2"
logger.info("Load from slowfast.")
if cfg.NUM_GPUS > 1:
model.module.load_weight_slowfast(slow_fast_model)
else:
model.load_weight_slowfast(slow_fast_model)
# if du.is_master_proc():
# misc.log_model_info(model, cfg, is_train=True)
if cfg.BN.FREEZE:
model.freeze_fn('bn_parameters')
if cfg.EPICKITCHENS.USE_BBOX and cfg.EPICKITCHENS.FREEZE_BACKBONE:
model.freeze_fn('slowfast_bbox')
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
# Load a checkpoint to resume training if applicable.
if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(cfg.OUTPUT_DIR):
logger.info("Load from last checkpoint.")
last_checkpoint = cu.get_last_checkpoint(cfg.OUTPUT_DIR)
checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model,
cfg.NUM_GPUS > 1, optimizer)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "" and not cfg.TRAIN.FINETUNE:
logger.info("Load from given checkpoint file.")
checkpoint_epoch = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
optimizer,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.CHECKPOINT_FILE_PATH != "" and cfg.TRAIN.FINETUNE:
logger.info("Load from given checkpoint file. Finetuning.")
_ = cu.load_checkpoint(
cfg.TRAIN.CHECKPOINT_FILE_PATH,
model,
cfg.NUM_GPUS > 1,
None,
inflation=cfg.TRAIN.CHECKPOINT_INFLATE,
convert_from_caffe2=cfg.TRAIN.CHECKPOINT_TYPE == "caffe2",
)
start_epoch = 0
else:
start_epoch = 0
# Create the video train and val loaders.
if cfg.TRAIN.DATASET != 'epickitchens' or not cfg.EPICKITCHENS.TRAIN_PLUS_VAL:
train_loader = loader.construct_loader(cfg, "train")
val_loader = loader.construct_loader(cfg, "val")
logger.info("Train loader size: {}".format(len(train_loader)))
logger.info("Val loader size: {}".format(len(val_loader)))
else:
train_loader = loader.construct_loader(cfg, "train+val")
val_loader = loader.construct_loader(cfg, "val")
# Create meters.
if cfg.DETECTION.ENABLE:
train_meter = AVAMeter(len(train_loader), cfg, mode="train")
val_meter = AVAMeter(len(val_loader), cfg, mode="val")
else:
if cfg.TRAIN.DATASET == 'epickitchens':
train_meter = EPICTrainMeterSimple(len(train_loader), cfg)
val_meter = EPICValMeterSimple(len(val_loader), cfg)
else:
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))
cnt = 0
# eval_epoch(val_loader, model, val_meter, 0, cfg, cnt)
# test_from_train(model, cfg, cnt=cnt)
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.
cnt = train_epoch(train_loader, model, optimizer, train_meter,
cur_epoch, cfg, cnt)
# Compute precise BN stats.
if cfg.BN.USE_PRECISE_STATS and len(get_bn_modules(model)) > 0:
calculate_and_update_precise_bn(train_loader, model,
cfg.BN.NUM_BATCHES_PRECISE)
# Save a checkpoint.
if cu.is_checkpoint_epoch(cur_epoch, cfg.TRAIN.CHECKPOINT_PERIOD):
cu.save_checkpoint(cfg.OUTPUT_DIR, model, optimizer, cur_epoch,
cfg)
# Evaluate the model on validation set.
if misc.is_eval_epoch(cfg, cur_epoch):
is_best_epoch = eval_epoch(val_loader, model, val_meter, cur_epoch,
cfg, cnt)
if is_best_epoch:
cu.save_checkpoint(cfg.OUTPUT_DIR,
model,
optimizer,
cur_epoch,
cfg,
is_best_epoch=is_best_epoch)
