def train_des(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(cfg.OUTPUT_DIR)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_clevrer_model(cfg)
# Construct the optimizer.
optimizer = AdamW(model.parameters(), lr=cfg.SOLVER.BASE_LR, eps=1e-8)
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Create the video train and val loaders.
train_loader = build_dataloader(cfg, "train")
val_loader = build_dataloader(cfg, "val")
total_steps = len(train_loader) * cfg.SOLVER.MAX_EPOCH
# Create the learning rate scheduler.
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=0, # Default value in run_glue.py
num_training_steps=total_steps)
# Create meters.
train_meter = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(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):
# Train for one epoch.
train_epoch(train_loader, model, optimizer, scheduler, train_meter,
cur_epoch, cfg)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None,
)
is_eval_epoch = misc.is_eval_epoch(cfg, cur_epoch, None)
# 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 test_train_iter2():
cfg = get_cfg()
cfg.LOG_PERIOD = 1
train_meter = ClevrerTrainMeter(5, cfg)
train_meter.update_stats(top1_err=40.0,
top5_err=20.0,
mc_opt_err=55.0,
mc_q_err=78.0,
loss_des=6.3,
loss_mc=0.2,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
train_meter.update_stats(top1_err=41.0,
top5_err=21.0,
mc_opt_err=56.0,
mc_q_err=86.0,
loss_des=6.7,
loss_mc=1.2,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
stats = train_meter.log_iter_stats(cur_epoch=1, cur_iter=1)
assert stats['top1_err'] == 41.0
assert stats['top5_err'] == 21.0
assert stats['mc_opt_err'] == 56.0
assert stats['mc_q_err'] == 86.0
assert stats['loss_des'] == 6.7
assert stats['loss_mc'] == 1.2
def test_train_iter4():
cfg = get_cfg()
cfg.LOG_PERIOD = 2
train_meter = ClevrerTrainMeter(5, cfg)
train_meter.update_stats(top1_err=40.0,
top5_err=20.0,
mc_opt_err=55.0,
mc_q_err=78.0,
loss_des=6.3,
loss_mc=0.2,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
train_meter.update_stats(top1_err=10.0,
top5_err=10.0,
mc_opt_err=10.0,
mc_q_err=10.0,
loss_des=10.0,
loss_mc=10.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
stats = train_meter.log_iter_stats(cur_epoch=1, cur_iter=2)
assert stats is None
train_meter.update_stats(top1_err=20.0,
top5_err=20.0,
mc_opt_err=20.0,
mc_q_err=20.0,
loss_des=20.0,
loss_mc=20.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
stats = train_meter.log_iter_stats(cur_epoch=0, cur_iter=3)
assert stats['top1_err'] == 15.0
assert stats['top5_err'] == 15.0
assert stats['mc_opt_err'] == 15.0
assert stats['mc_q_err'] == 15.0
assert stats['loss_des'] == 15.0
assert stats['loss_mc'] == 15.0
def test_train_iter5():
cfg = get_cfg()
cfg.LOG_PERIOD = 3
train_meter = ClevrerTrainMeter(5, cfg)
train_meter.update_stats(top1_err=30.0,
top5_err=30.0,
mc_opt_err=30.0,
mc_q_err=30.0,
loss_des=30.0,
loss_mc=30.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
train_meter.update_stats(top1_err=10.0,
top5_err=10.0,
mc_opt_err=10.0,
mc_q_err=10.0,
loss_des=10.0,
loss_mc=10.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
stats = train_meter.log_iter_stats(cur_epoch=1, cur_iter=1)
assert stats is None
train_meter.update_stats(top1_err=20.0,
top5_err=20.0,
mc_opt_err=20.0,
mc_q_err=20.0,
loss_des=20.0,
loss_mc=20.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=10)
stats = train_meter.log_iter_stats(cur_epoch=0, cur_iter=5)
assert stats['top1_err'] == 20.0, print(stats['top1_err'])
assert stats['top5_err'] == 20.0
assert stats['mc_opt_err'] == 20.0
assert stats['mc_q_err'] == 20.0
assert stats['loss_des'] == 20.0
assert stats['loss_mc'] == 20.0
def test_train_epoch_only_des():
cfg = get_cfg()
cfg.LOG_PERIOD = 3
train_meter = ClevrerTrainMeter(5, cfg)
train_meter.update_stats(top1_err=30.0,
top5_err=30.0,
mc_opt_err=30.0,
mc_q_err=30.0,
loss_des=30.0,
loss_mc=30.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=0)
train_meter.update_stats(top1_err=10.0,
top5_err=10.0,
mc_opt_err=10.0,
mc_q_err=10.0,
loss_des=10.0,
loss_mc=10.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=0)
train_meter.update_stats(top1_err=20.0,
top5_err=20.0,
mc_opt_err=20.0,
mc_q_err=20.0,
loss_des=20.0,
loss_mc=20.0,
lr=0.1,
mb_size_des=10,
mb_size_mc=0)
stats = train_meter.log_epoch_stats(cur_epoch=1)
assert stats['top1_err'] == 20.0, print(stats['top1_err'])
assert stats['top5_err'] == 20.0
assert not 'mc_opt_err' in stats
assert not 'mc_q_err' in stats
assert stats['loss_des'] == 20.0
assert not 'loss_mc' in stats
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 (ClevrerTrainMeter): tool for measuring training stats.
val_meter (ClevrerValMeter): tool for measuring validation stats.
"""
# Build the video model and print model statistics.
model = build_clevrer_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 = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(len(val_loader), cfg)
return (
model,
optimizer,
train_loader,
val_loader,
precise_bn_loader,
train_meter,
val_meter,
)
def test_implementation_des(cfg):
"""
Simulates a train and val epoch to check if the gradients are being updated,
metrics are being calculated correctly
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(cfg.OUTPUT_DIR)
# Print config.
logger.info("Test implementation")
# Build the video model and print model statistics.
model = build_clevrer_model(cfg)
# Construct the optimizer.
optimizer = optim.construct_optimizer(model, cfg)
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
# Create the video train and val loaders.
if cfg.TRAIN.DATASET != 'Clevrer_des':
print("This train script does not support your dataset: -{}-. Only Clevrer_des".format(cfg.TRAIN.DATASET))
exit()
train_loader = build_dataloader(cfg, "train")
val_loader = build_dataloader(cfg, "val")
# Create meters.
train_meter = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(len(val_loader), cfg)
# Perform the training loop.
logger.info("Start epoch: {}".format(start_epoch + 1))
# Train for one epoch.
model_before = copy.deepcopy(model)
cur_epoch = start_epoch
train_epoch(
train_loader, model, optimizer, train_meter, cur_epoch, cfg, test_imp=True
)
print("Check how much parameters changed")
for (p_b_name, p_b), (p_name, p) in zip(model_before.named_parameters(), model.named_parameters()):
if p.requires_grad:
print("Parameter requires grad:")
print(p_name, p_b_name)
#Calculate ratio of change
change = torch.abs(torch.norm(p) - torch.norm(p_b))
print("Ratio of change = {}".format(torch.true_divide(change, torch.norm(p_b))))
if (p_b != p).any():
print("--Check--")
else:
print("ALERT - WEIGHTS DID NOT CHANGE WITH TRAINING.")
else:
print("Parameter does not require grad:")
print(p_name)
print(p)
print("Val epoch")
eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, test_imp=True)
def train_des(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(cfg.OUTPUT_DIR)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = build_clevrer_model(cfg)
# 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.
if cfg.TRAIN.DATASET != 'Clevrer_des':
print("This train script does not support your dataset: -{}-. Only Clevrer_des".format(cfg.TRAIN.DATASET))
exit()
# Create the video train and val loaders.
train_loader = build_dataloader(cfg, "train")
val_loader = build_dataloader(cfg, "val")
# Create meters.
train_meter = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(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
)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None,
)
is_eval_epoch = misc.is_eval_epoch(
cfg, cur_epoch, None
)
# 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_clevrer_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.
train_meter = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(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_des(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(cfg.OUTPUT_DIR)
# Print config.
logger.info("Train with config:")
logger.info(pprint.pformat(cfg))
# Build the video model and print model statistics.
model = MONET_BERT(cfg)
if cfg.NUM_GPUS:
# 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)
# Construct the optimizer.
# optimizer = AdamW(model.parameters(),
# lr = cfg.SOLVER.BASE_LR,
# eps = 1e-8
# )
optimizer = Lamb(model.parameters(),
lr=cfg.SOLVER.BASE_LR,
betas=(0.9, 0.999),
eps=1e-6,
weight_decay=cfg.SOLVER.WEIGHT_DECAY,
adam=False)
# optimizer = optim.Adam(model.parameters(), lr=cfg.SOLVER.BASE_LR)
start_epoch = cu.load_train_checkpoint(cfg, model, optimizer)
optimizer = Lamb(model.parameters(),
lr=cfg.SOLVER.BASE_LR,
betas=(0.9, 0.999),
eps=1e-6,
weight_decay=cfg.SOLVER.WEIGHT_DECAY,
adam=False)
# Create the video train and val loaders.
train_loader = build_dataloader(cfg, "train")
val_loader = build_dataloader(cfg, "val")
total_steps = len(train_loader) * cfg.SOLVER.MAX_EPOCH
# Create the learning rate scheduler.
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=4000, # Default value in run_glue.py
num_training_steps=total_steps)
# Create meters.
train_meter = ClevrerTrainMeter(len(train_loader), cfg)
val_meter = ClevrerValMeter(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):
# Train for one epoch.
train_epoch(train_loader, model, optimizer, scheduler, train_meter,
cur_epoch, cfg)
is_checkp_epoch = cu.is_checkpoint_epoch(
cfg,
cur_epoch,
None,
)
is_eval_epoch = misc.is_eval_epoch(cfg, cur_epoch, None)
# 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)
