def train_model():
"""Trains the model."""
# Setup training/testing environment
setup_env()
# Construct the model, loss_fun, and optimizer
model = setup_model()
loss_fun = builders.build_loss_fun().cuda()
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
if cfg.TRAIN.AUTO_RESUME and checkpoint.has_checkpoint():
last_checkpoint = checkpoint.get_last_checkpoint()
checkpoint_epoch = checkpoint.load_checkpoint(last_checkpoint, model,
optimizer)
logger.info("Loaded checkpoint from: {}".format(last_checkpoint))
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.WEIGHTS:
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS, model)
logger.info("Loaded initial weights from: {}".format(
cfg.TRAIN.WEIGHTS))
# Create data loaders and meters
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
train_meter = meters.TrainMeter(len(train_loader))
test_meter = meters.TestMeter(len(test_loader))
# Compute model and loader timings
if start_epoch == 0 and cfg.PREC_TIME.NUM_ITER > 0:
benchmark.compute_time_full(model, loss_fun, train_loader, test_loader)
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
# Train for one epoch
train_epoch(train_loader, model, loss_fun, optimizer, train_meter,
cur_epoch)
# Compute precise BN stats
if cfg.BN.USE_PRECISE_STATS:
net.compute_precise_bn_stats(model, train_loader)
# Save a checkpoint
if (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0:
checkpoint_file = checkpoint.save_checkpoint(
model, optimizer, cur_epoch)
logger.info("Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
next_epoch = cur_epoch + 1
if next_epoch % cfg.TRAIN.EVAL_PERIOD == 0 or next_epoch == cfg.OPTIM.MAX_EPOCH:
test_epoch(test_loader, model, test_meter, cur_epoch)
def test_model():
"""Evaluates the model."""
# Setup logging
logging.setup_logging()
# Show the config
logger.info("Config:\n{}".format(cfg))
# Fix the RNG seeds (see RNG comment in core/config.py for discussion)
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Configure the CUDNN backend
torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK
# Build the model (before the loaders to speed up debugging)
model = builders.build_model()
logger.info("Model:\n{}".format(model))
logger.info(logging.dump_json_stats(net.complexity(model)))
# Compute precise time
if cfg.PREC_TIME.ENABLED:
logger.info("Computing precise time...")
loss_fun = builders.build_loss_fun()
prec_time = net.compute_precise_time(model, loss_fun)
logger.info(logging.dump_json_stats(prec_time))
net.reset_bn_stats(model)
# Load model weights
checkpoint.load_checkpoint(cfg.TEST.WEIGHTS, model)
logger.info("Loaded model weights from: {}".format(cfg.TEST.WEIGHTS))
# Create data loaders
test_loader = loader.construct_test_loader()
# Create meters
test_meter = meters.TestMeter(len(test_loader))
# Evaluate the model
test_epoch(test_loader, model, test_meter, 0)
def train_model():
"""Trains the model."""
# Setup logging
logging.setup_logging()
# Show the config
logger.info("Config:\n{}".format(cfg))
# Fix the RNG seeds (see RNG comment in core/config.py for discussion)
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Configure the CUDNN backend
torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK
# Build the model (before the loaders to speed up debugging)
model = builders.build_model()
logger.info("Model:\n{}".format(model))
logger.info(logging.dump_json_stats(net.complexity(model)))
# Define the loss function
loss_fun = builders.build_loss_fun()
# Construct the optimizer
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
if cfg.TRAIN.AUTO_RESUME and checkpoint.has_checkpoint():
last_checkpoint = checkpoint.get_last_checkpoint()
checkpoint_epoch = checkpoint.load_checkpoint(last_checkpoint, model,
optimizer)
logger.info("Loaded checkpoint from: {}".format(last_checkpoint))
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.WEIGHTS:
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS, model)
logger.info("Loaded initial weights from: {}".format(
cfg.TRAIN.WEIGHTS))
# Compute precise time
if start_epoch == 0 and cfg.PREC_TIME.ENABLED:
logger.info("Computing precise time...")
prec_time = net.compute_precise_time(model, loss_fun)
logger.info(logging.dump_json_stats(prec_time))
net.reset_bn_stats(model)
# Create data loaders
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
# Create meters
train_meter = meters.TrainMeter(len(train_loader))
test_meter = meters.TestMeter(len(test_loader))
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
# Train for one epoch
train_epoch(train_loader, model, loss_fun, optimizer, train_meter,
cur_epoch)
# Compute precise BN stats
if cfg.BN.USE_PRECISE_STATS:
net.compute_precise_bn_stats(model, train_loader)
# Save a checkpoint
if checkpoint.is_checkpoint_epoch(cur_epoch):
checkpoint_file = checkpoint.save_checkpoint(
model, optimizer, cur_epoch)
logger.info("Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
if is_eval_epoch(cur_epoch):
test_epoch(test_loader, model, test_meter, cur_epoch)
def test_ftta_model(corruptions, levels):
"""Use feed back to fine-tune some part of the model. (with all kind of corruptions)"""
all_results = []
for corruption_level in levels:
lvl_results = []
for corruption_type in corruptions:
cfg.TRAIN.CORRUPTION = corruption_type
cfg.TRAIN.LEVEL = corruption_level
cfg.TEST.CORRUPTION = corruption_type
cfg.TEST.LEVEL = corruption_level
# Setup training/testing environment
setup_env()
# Construct the model, loss_fun, and optimizer
model = setup_model()
loss_fun = builders.build_loss_fun().cuda()
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS,
model,
strict=cfg.TRAIN.LOAD_STRICT)
logger.info("Loaded initial weights from: {}".format(
cfg.TRAIN.WEIGHTS))
# Create data loaders and meters
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
train_meter = meters.TrainMeter(len(train_loader))
test_meter = meters.TestMeter(len(test_loader))
# Compute model and loader timings
if start_epoch == 0 and cfg.PREC_TIME.NUM_ITER > 0:
benchmark.compute_time_full(model, loss_fun, train_loader,
test_loader)
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
if cfg.TRAIN.ADAPTATION != 'test_only':
if cfg.TRAIN.ADAPTATION == 'update_bn':
bn_update(model, train_loader)
elif cfg.TRAIN.ADAPTATION == 'min_entropy':
# Train for one epoch
train_epoch(train_loader, model, loss_fun, optimizer,
train_meter, cur_epoch)
bn_update(model, train_loader)
# Save a checkpoint
if (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0:
checkpoint_file = checkpoint.save_checkpoint(
model, optimizer, cur_epoch)
logger.info(
"Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
next_epoch = cur_epoch + 1
if next_epoch % cfg.TRAIN.EVAL_PERIOD == 0 or next_epoch == cfg.OPTIM.MAX_EPOCH:
top1 = test_epoch(test_loader, model, test_meter,
cur_epoch)
lvl_results.append(top1)
all_results.append(lvl_results)
for lvl_idx in range(len(all_results)):
logger.info("corruption level: {}".format(levels[lvl_idx]))
logger.info("corruption types: {}".format(corruptions))
logger.info(all_results[lvl_idx])
# show_parameters(model)
return all_results
def train_model():
"""Trains the model."""
# Setup training/testing environment
setup_env()
# Construct the model, loss_fun, and optimizer
model = setup_model()
loss_fun = builders.build_loss_fun().cuda()
if "search" in cfg.MODEL.TYPE:
params_w = [v for k, v in model.named_parameters() if "alphas" not in k]
params_a = [v for k, v in model.named_parameters() if "alphas" in k]
optimizer_w = torch.optim.SGD(
params=params_w,
lr=cfg.OPTIM.BASE_LR,
momentum=cfg.OPTIM.MOMENTUM,
weight_decay=cfg.OPTIM.WEIGHT_DECAY,
dampening=cfg.OPTIM.DAMPENING,
nesterov=cfg.OPTIM.NESTEROV
)
if cfg.OPTIM.ARCH_OPTIM == "adam":
optimizer_a = torch.optim.Adam(
params=params_a,
lr=cfg.OPTIM.ARCH_BASE_LR,
betas=(0.5, 0.999),
weight_decay=cfg.OPTIM.ARCH_WEIGHT_DECAY
)
elif cfg.OPTIM.ARCH_OPTIM == "sgd":
optimizer_a = torch.optim.SGD(
params=params_a,
lr=cfg.OPTIM.ARCH_BASE_LR,
momentum=cfg.OPTIM.MOMENTUM,
weight_decay=cfg.OPTIM.ARCH_WEIGHT_DECAY,
dampening=cfg.OPTIM.DAMPENING,
nesterov=cfg.OPTIM.NESTEROV
)
optimizer = [optimizer_w, optimizer_a]
else:
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
if cfg.TRAIN.AUTO_RESUME and checkpoint.has_checkpoint():
last_checkpoint = checkpoint.get_last_checkpoint()
checkpoint_epoch = checkpoint.load_checkpoint(last_checkpoint, model, optimizer)
logger.info("Loaded checkpoint from: {}".format(last_checkpoint))
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.WEIGHTS:
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS, model)
logger.info("Loaded initial weights from: {}".format(cfg.TRAIN.WEIGHTS))
# Create data loaders and meters
if cfg.TRAIN.PORTION < 1:
if "search" in cfg.MODEL.TYPE:
train_loader = [loader._construct_loader(
dataset_name=cfg.TRAIN.DATASET,
split=cfg.TRAIN.SPLIT,
batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS),
shuffle=True,
drop_last=True,
portion=cfg.TRAIN.PORTION,
side="l"
),
loader._construct_loader(
dataset_name=cfg.TRAIN.DATASET,
split=cfg.TRAIN.SPLIT,
batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS),
shuffle=True,
drop_last=True,
portion=cfg.TRAIN.PORTION,
side="r"
)]
else:
train_loader = loader._construct_loader(
dataset_name=cfg.TRAIN.DATASET,
split=cfg.TRAIN.SPLIT,
batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS),
shuffle=True,
drop_last=True,
portion=cfg.TRAIN.PORTION,
side="l"
)
test_loader = loader._construct_loader(
dataset_name=cfg.TRAIN.DATASET,
split=cfg.TRAIN.SPLIT,
batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS),
shuffle=False,
drop_last=False,
portion=cfg.TRAIN.PORTION,
side="r"
)
else:
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
train_meter_type = meters.TrainMeterIoU if cfg.TASK == "seg" else meters.TrainMeter
test_meter_type = meters.TestMeterIoU if cfg.TASK == "seg" else meters.TestMeter
l = train_loader[0] if isinstance(train_loader, list) else train_loader
train_meter = train_meter_type(len(l))
test_meter = test_meter_type(len(test_loader))
# Compute model and loader timings
if start_epoch == 0 and cfg.PREC_TIME.NUM_ITER > 0:
l = train_loader[0] if isinstance(train_loader, list) else train_loader
benchmark.compute_time_full(model, loss_fun, l, test_loader)
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
# Train for one epoch
f = search_epoch if "search" in cfg.MODEL.TYPE else train_epoch
f(train_loader, model, loss_fun, optimizer, train_meter, cur_epoch)
# Compute precise BN stats
if cfg.BN.USE_PRECISE_STATS:
net.compute_precise_bn_stats(model, train_loader)
# Save a checkpoint
if (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0:
checkpoint_file = checkpoint.save_checkpoint(model, optimizer, cur_epoch)
logger.info("Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
next_epoch = cur_epoch + 1
if next_epoch % cfg.TRAIN.EVAL_PERIOD == 0 or next_epoch == cfg.OPTIM.MAX_EPOCH:
test_epoch(test_loader, model, test_meter, cur_epoch)
def train_model():
"""Trains the model."""
# Setup training/testing environment
setup_env()
# Construct the model, loss_fun, and optimizer
model = setup_model()
loss_fun = builders.build_loss_fun().cuda()
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
if cfg.TRAIN.AUTO_RESUME and checkpoint.has_checkpoint():
last_checkpoint = checkpoint.get_last_checkpoint()
checkpoint_epoch = checkpoint.load_checkpoint(last_checkpoint, model,
optimizer)
logger.info("Loaded checkpoint from: {}".format(last_checkpoint))
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.WEIGHTS:
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS, model)
logger.info("Loaded initial weights from: {}".format(
cfg.TRAIN.WEIGHTS))
# Create data loaders and meters
if cfg.TEST.DATASET == 'imagenet_dataset' or cfg.TRAIN.DATASET == 'imagenet_dataset':
dataset = loader.construct_train_loader()
train_loader = dataset.train_loader
test_loader = dataset.val_loader
else:
dataset = None
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
train_meter = meters.TrainMeter(len(train_loader))
test_meter = meters.TestMeter(len(test_loader))
# Compute model and loader timings
if start_epoch == 0 and cfg.PREC_TIME.NUM_ITER > 0:
benchmark.compute_time_full(model, loss_fun, train_loader, test_loader)
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
# Train for one epoch
train_epoch(train_loader, model, loss_fun, optimizer, train_meter,
cur_epoch)
# Compute precise BN stats
if cfg.BN.USE_PRECISE_STATS:
net.compute_precise_bn_stats(model, train_loader)
# Save a checkpoint
if (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0:
checkpoint_file = checkpoint.save_checkpoint(
model, optimizer, cur_epoch)
logger.info("Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
next_epoch = cur_epoch + 1
if next_epoch % cfg.TRAIN.EVAL_PERIOD == 0 or next_epoch == cfg.OPTIM.MAX_EPOCH:
logger.info("Start testing")
test_epoch(test_loader, model, test_meter, cur_epoch)
if dataset is not None:
logger.info("Reset the dataset")
train_loader._dali_iterator.reset()
test_loader._dali_iterator.reset()
# clear memory
if torch.cuda.is_available():
torch.cuda.synchronize()
torch.cuda.empty_cache(
) # https://forums.fast.ai/t/clearing-gpu-memory-pytorch/14637
gc.collect()