def main_worker(current_gpu, config):
configure_device(current_gpu, config)
config.mlflow = SafeMLFLow(config)
if is_main_process():
configure_logging(logger, config)
print_args(config)
logger.info(config)
dataset = get_dataset(config.dataset)
color_encoding = dataset.color_encoding
num_classes = len(color_encoding)
if config.metrics_dump is not None:
write_metrics(0, config.metrics_dump)
train_loader = val_loader = criterion = None
resuming_checkpoint_path = config.resuming_checkpoint_path
nncf_config = config.nncf_config
pretrained = is_pretrained_model_requested(config)
def criterion_fn(model_outputs, target, criterion_):
labels, loss_outputs, _ = \
loss_funcs.do_model_specific_postprocessing(config.model, target, model_outputs)
return criterion_(loss_outputs, labels)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
loaders, w_class = load_dataset(dataset, config)
train_loader, val_loader, init_loader = loaders
criterion = get_criterion(w_class, config)
def autoq_test_fn(model, eval_loader):
return test(model, eval_loader, criterion, color_encoding, config)
nncf_config = register_default_init_args(nncf_config, init_loader,
criterion, criterion_fn,
autoq_test_fn, val_loader,
config.device)
model = load_model(config.model,
pretrained=pretrained,
num_classes=num_classes,
model_params=config.get('model_params', {}),
weights_path=config.get('weights'))
model.to(config.device)
resuming_model_sd = None
resuming_checkpoint = None
if resuming_checkpoint_path is not None:
resuming_model_sd, resuming_checkpoint = load_resuming_model_state_dict_and_checkpoint_from_path(
resuming_checkpoint_path)
compression_ctrl, model = create_compressed_model(
model, nncf_config, resuming_state_dict=resuming_model_sd)
model, model_without_dp = prepare_model_for_execution(model, config)
if config.distributed:
compression_ctrl.distributed()
log_common_mlflow_params(config)
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if is_main_process():
print_statistics(compression_ctrl.statistics())
if config.mode.lower() == 'test':
logger.info(model)
model_parameters = filter(lambda p: p.requires_grad,
model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
logger.info("Trainable argument count:{params}".format(params=params))
model = model.to(config.device)
test(model, val_loader, criterion, color_encoding, config)
elif config.mode.lower() == 'train':
train(model, model_without_dp, compression_ctrl, train_loader,
val_loader, criterion, color_encoding, config,
resuming_checkpoint)
else:
# Should never happen...but just in case it does
raise RuntimeError(
"\"{0}\" is not a valid choice for execution mode.".format(
config.mode))
def main_worker(current_gpu, config: SampleConfig):
config.current_gpu = current_gpu
config.distributed = config.execution_mode in (
ExecutionMode.DISTRIBUTED, ExecutionMode.MULTIPROCESSING_DISTRIBUTED)
if config.distributed:
configure_distributed(config)
config.device = get_device(config)
if is_main_process():
configure_logging(logger, config)
print_args(config)
if config.seed is not None:
manual_seed(config.seed)
cudnn.deterministic = True
cudnn.benchmark = False
# define loss function (criterion)
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(config.device)
train_loader = train_sampler = val_loader = None
resuming_checkpoint_path = config.resuming_checkpoint_path
nncf_config = config.nncf_config
pretrained = is_pretrained_model_requested(config)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
# Data loading code
train_dataset, val_dataset = create_datasets(config)
train_loader, train_sampler, val_loader = create_data_loaders(
config, train_dataset, val_dataset)
nncf_config = register_default_init_args(nncf_config, criterion,
train_loader)
# create model
model_name = config['model']
model = load_model(model_name,
pretrained=pretrained,
num_classes=config.get('num_classes', 1000),
model_params=config.get('model_params'),
weights_path=config.get('weights'))
model.to(config.device)
resuming_model_sd = None
resuming_checkpoint = None
if resuming_checkpoint_path is not None:
resuming_checkpoint = load_resuming_checkpoint(
resuming_checkpoint_path)
resuming_model_sd = resuming_checkpoint['state_dict']
compression_ctrl, model = create_compressed_model(
model, nncf_config, resuming_state_dict=resuming_model_sd)
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
model, _ = prepare_model_for_execution(model, config)
if config.distributed:
compression_ctrl.distributed()
# define optimizer
params_to_optimize = get_parameter_groups(model, config)
optimizer, lr_scheduler = make_optimizer(params_to_optimize, config)
best_acc1 = 0
# optionally resume from a checkpoint
if resuming_checkpoint_path is not None:
if config.mode.lower() == 'train' and config.to_onnx is None:
config.start_epoch = resuming_checkpoint['epoch']
best_acc1 = resuming_checkpoint['best_acc1']
compression_ctrl.scheduler.load_state_dict(
resuming_checkpoint['scheduler'])
optimizer.load_state_dict(resuming_checkpoint['optimizer'])
logger.info(
"=> loaded checkpoint '{}' (epoch: {}, best_acc1: {:.3f})".
format(resuming_checkpoint_path, resuming_checkpoint['epoch'],
best_acc1))
else:
logger.info(
"=> loaded checkpoint '{}'".format(resuming_checkpoint_path))
if config.execution_mode != ExecutionMode.CPU_ONLY:
cudnn.benchmark = True
if config.mode.lower() == 'test':
print_statistics(compression_ctrl.statistics())
validate(val_loader, model, criterion, config)
if config.mode.lower() == 'train':
is_inception = 'inception' in model_name
train(config, compression_ctrl, model, criterion, is_inception,
lr_scheduler, model_name, optimizer, train_loader, train_sampler,
val_loader, best_acc1)
def main_worker(current_gpu, config):
#################################
# Setup experiment environment
#################################
config.current_gpu = current_gpu
config.distributed = config.execution_mode in (
ExecutionMode.DISTRIBUTED, ExecutionMode.MULTIPROCESSING_DISTRIBUTED)
if config.distributed:
configure_distributed(config)
if is_on_first_rank(config):
configure_logging(logger, config)
print_args(config)
config.device = get_device(config)
config.start_iter = 0
##########################
# Prepare metrics log file
##########################
if config.metrics_dump is not None:
write_metrics(0, config.metrics_dump)
###########################
# Criterion
###########################
criterion = MultiBoxLoss(config,
config['num_classes'],
overlap_thresh=0.5,
prior_for_matching=True,
bkg_label=0,
neg_mining=True,
neg_pos=3,
neg_overlap=0.5,
encode_target=False,
device=config.device)
train_data_loader = test_data_loader = None
resuming_checkpoint_path = config.resuming_checkpoint_path
###########################
# Prepare data
###########################
pretrained = is_pretrained_model_requested(config)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
test_data_loader, train_data_loader = create_dataloaders(config)
config.nncf_config = register_default_init_args(
config.nncf_config, criterion, train_data_loader)
##################
# Prepare model
##################
resuming_checkpoint_path = config.resuming_checkpoint_path
resuming_checkpoint = None
resuming_model_state_dict = None
if resuming_checkpoint_path:
logger.info(
'Resuming from checkpoint {}...'.format(resuming_checkpoint_path))
resuming_checkpoint = torch.load(resuming_checkpoint_path,
map_location='cpu')
# use checkpoint itself in case only the state dict was saved,
# i.e. the checkpoint was created with `torch.save(module.state_dict())`
resuming_model_state_dict = resuming_checkpoint.get(
'state_dict', resuming_checkpoint)
compression_ctrl, net = create_model(config, resuming_model_state_dict)
if config.distributed:
config.batch_size //= config.ngpus_per_node
config.workers //= config.ngpus_per_node
compression_ctrl.distributed()
###########################
# Optimizer
###########################
params_to_optimize = get_parameter_groups(net, config)
optimizer, lr_scheduler = make_optimizer(params_to_optimize, config)
#################################
# Load additional checkpoint data
#################################
if resuming_checkpoint is not None and config.mode.lower(
) == 'train' and config.to_onnx is None:
compression_ctrl.scheduler.load_state_dict(
resuming_checkpoint['scheduler'])
optimizer.load_state_dict(
resuming_checkpoint.get('optimizer', optimizer.state_dict()))
config.start_iter = resuming_checkpoint.get('iter', 0) + 1
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if config.mode.lower() == 'test':
with torch.no_grad():
print_statistics(compression_ctrl.statistics())
net.eval()
mAp = test_net(net,
config.device,
test_data_loader,
distributed=config.distributed)
if config.metrics_dump is not None:
write_metrics(mAp, config.metrics_dump)
return
train(net, compression_ctrl, train_data_loader, test_data_loader,
criterion, optimizer, config, lr_scheduler)
def staged_quantization_main_worker(current_gpu, config):
config.current_gpu = current_gpu
config.distributed = config.execution_mode in (
ExecutionMode.DISTRIBUTED, ExecutionMode.MULTIPROCESSING_DISTRIBUTED)
if config.distributed:
configure_distributed(config)
config.device = get_device(config)
if is_main_process():
configure_logging(logger, config)
print_args(config)
if config.seed is not None:
manual_seed(config.seed)
cudnn.deterministic = True
cudnn.benchmark = False
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(config.device)
train_loader = train_sampler = val_loader = None
resuming_checkpoint_path = config.resuming_checkpoint_path
nncf_config = config.nncf_config
pretrained = is_pretrained_model_requested(config)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
# Data loading code
train_dataset, val_dataset = create_datasets(config)
train_loader, train_sampler, val_loader = create_data_loaders(
config, train_dataset, val_dataset)
nncf_config = register_default_init_args(nncf_config, criterion,
train_loader)
# create model
model_name = config['model']
model = load_model(model_name,
pretrained=pretrained,
num_classes=config.get('num_classes', 1000),
model_params=config.get('model_params'),
weights_path=config.get('weights'))
original_model = copy.deepcopy(model)
model.to(config.device)
resuming_model_sd = None
resuming_checkpoint = None
if resuming_checkpoint_path is not None:
resuming_checkpoint = load_resuming_checkpoint(
resuming_checkpoint_path)
resuming_model_sd = resuming_checkpoint['state_dict']
compression_ctrl, model = create_compressed_model(model, nncf_config,
resuming_model_sd)
if not isinstance(compression_ctrl,
(BinarizationController, QuantizationController)):
raise RuntimeError(
"The stage quantization sample worker may only be run with the binarization and quantization algorithms!"
)
model, _ = prepare_model_for_execution(model, config)
original_model.to(config.device)
if config.distributed:
compression_ctrl.distributed()
is_inception = 'inception' in model_name
params_to_optimize = model.parameters()
compression_config = config['compression']
quantization_config = compression_config if isinstance(
compression_config, dict) else compression_config[0]
optimizer = get_quantization_optimizer(params_to_optimize,
quantization_config)
optimizer_scheduler = PolyLRDropScheduler(optimizer, quantization_config)
kd_loss_calculator = KDLossCalculator(original_model)
best_acc1 = 0
# optionally resume from a checkpoint
if resuming_checkpoint is not None and config.to_onnx is None:
config.start_epoch = resuming_checkpoint['epoch']
best_acc1 = resuming_checkpoint['best_acc1']
kd_loss_calculator.original_model.load_state_dict(
resuming_checkpoint['original_model_state_dict'])
compression_ctrl.scheduler.load_state_dict(
resuming_checkpoint['compression_scheduler'])
optimizer.load_state_dict(resuming_checkpoint['optimizer'])
optimizer_scheduler.load_state_dict(
resuming_checkpoint['optimizer_scheduler'])
if config.mode.lower() == 'train':
logger.info(
"=> loaded checkpoint '{}' (epoch: {}, best_acc1: {:.3f})".
format(resuming_checkpoint_path, resuming_checkpoint['epoch'],
best_acc1))
else:
logger.info(
"=> loaded checkpoint '{}'".format(resuming_checkpoint_path))
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if config.execution_mode != ExecutionMode.CPU_ONLY:
cudnn.benchmark = True
if config.mode.lower() == 'test':
print_statistics(compression_ctrl.statistics())
validate(val_loader, model, criterion, config)
if config.mode.lower() == 'train':
batch_multiplier = (quantization_config.get("params", {})).get(
"batch_multiplier", 1)
train_staged(config, compression_ctrl, model, criterion, is_inception,
optimizer_scheduler, model_name, optimizer, train_loader,
train_sampler, val_loader, kd_loss_calculator,
batch_multiplier, best_acc1)
def main_worker(current_gpu, config):
config.current_gpu = current_gpu
config.distributed = config.execution_mode in (
ExecutionMode.DISTRIBUTED, ExecutionMode.MULTIPROCESSING_DISTRIBUTED)
if config.distributed:
configure_distributed(config)
if is_main_process():
configure_logging(logger, config)
print_args(config)
logger.info(config)
config.device = get_device(config)
dataset = get_dataset(config.dataset)
color_encoding = dataset.color_encoding
num_classes = len(color_encoding)
if config.metrics_dump is not None:
write_metrics(0, config.metrics_dump)
train_loader = val_loader = criterion = None
resuming_checkpoint_path = config.resuming_checkpoint_path
nncf_config = config.nncf_config
pretrained = is_pretrained_model_requested(config)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
loaders, w_class = load_dataset(dataset, config)
train_loader, val_loader = loaders
criterion = get_criterion(w_class, config)
if not resuming_checkpoint_path:
nncf_config = register_default_init_args(nncf_config, criterion,
train_loader)
model = load_model(config.model,
pretrained=pretrained,
num_classes=num_classes,
model_params=config.get('model_params', {}),
weights_path=config.get('weights'))
model.to(config.device)
compression_ctrl, model = create_compressed_model(model, nncf_config)
model, model_without_dp = prepare_model_for_execution(model, config)
if config.distributed:
compression_ctrl.distributed()
if resuming_checkpoint_path:
if not config.pretrained:
# Load the previously saved model state
model, _, _, _, _ = \
load_checkpoint(model, resuming_checkpoint_path, config.device,
compression_scheduler=compression_ctrl.scheduler)
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if config.mode.lower() == 'test':
logger.info(model)
model_parameters = filter(lambda p: p.requires_grad,
model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
logger.info("Trainable argument count:{params}".format(params=params))
model = model.to(config.device)
test(model, val_loader, criterion, color_encoding, config)
print_statistics(compression_ctrl.statistics())
elif config.mode.lower() == 'train':
train(model, model_without_dp, compression_ctrl, train_loader,
val_loader, criterion, color_encoding, config)
else:
# Should never happen...but just in case it does
raise RuntimeError(
"\"{0}\" is not a valid choice for execution mode.".format(
config.mode))
def main_worker(current_gpu, config):
#################################
# Setup experiment environment
#################################
configure_device(current_gpu, config)
config.mlflow = SafeMLFLow(config)
if is_on_first_rank(config):
configure_logging(logger, config)
print_args(config)
config.start_iter = 0
nncf_config = config.nncf_config
##########################
# Prepare metrics log file
##########################
if config.metrics_dump is not None:
write_metrics(0, config.metrics_dump)
###########################
# Criterion
###########################
criterion = MultiBoxLoss(config,
config['num_classes'],
overlap_thresh=0.5,
prior_for_matching=True,
bkg_label=0,
neg_mining=True,
neg_pos=3,
neg_overlap=0.5,
encode_target=False,
device=config.device)
train_data_loader = test_data_loader = None
resuming_checkpoint_path = config.resuming_checkpoint_path
###########################
# Prepare data
###########################
pretrained = is_pretrained_model_requested(config)
if config.to_onnx is not None:
assert pretrained or (resuming_checkpoint_path is not None)
else:
test_data_loader, train_data_loader, init_data_loader = create_dataloaders(
config)
def criterion_fn(model_outputs, target, criterion):
loss_l, loss_c = criterion(model_outputs, target)
return loss_l + loss_c
def autoq_test_fn(model, eval_loader):
# RL is maximization, change the loss polarity
return -1 * test_net(model,
config.device,
eval_loader,
distributed=config.distributed,
loss_inference=True,
criterion=criterion)
nncf_config = register_default_init_args(nncf_config, init_data_loader,
criterion, criterion_fn,
autoq_test_fn,
test_data_loader,
config.device)
##################
# Prepare model
##################
resuming_checkpoint_path = config.resuming_checkpoint_path
resuming_model_sd = None
if resuming_checkpoint_path is not None:
resuming_model_sd, resuming_checkpoint = load_resuming_model_state_dict_and_checkpoint_from_path(
resuming_checkpoint_path)
compression_ctrl, net = create_model(config, resuming_model_sd)
if config.distributed:
config.batch_size //= config.ngpus_per_node
config.workers //= config.ngpus_per_node
compression_ctrl.distributed()
###########################
# Optimizer
###########################
params_to_optimize = get_parameter_groups(net, config)
optimizer, lr_scheduler = make_optimizer(params_to_optimize, config)
#################################
# Load additional checkpoint data
#################################
if resuming_checkpoint_path is not None and config.mode.lower(
) == 'train' and config.to_onnx is None:
compression_ctrl.scheduler.load_state_dict(
resuming_checkpoint['scheduler'])
optimizer.load_state_dict(
resuming_checkpoint.get('optimizer', optimizer.state_dict()))
config.start_iter = resuming_checkpoint.get('iter', 0) + 1
log_common_mlflow_params(config)
if config.to_onnx:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if is_main_process():
print_statistics(compression_ctrl.statistics())
if config.mode.lower() == 'test':
with torch.no_grad():
net.eval()
if config['ssd_params'].get('loss_inference', False):
model_loss = test_net(net,
config.device,
test_data_loader,
distributed=config.distributed,
loss_inference=True,
criterion=criterion)
logger.info("Final model loss: {:.3f}".format(model_loss))
else:
mAp = test_net(net,
config.device,
test_data_loader,
distributed=config.distributed)
if config.metrics_dump is not None:
write_metrics(mAp, config.metrics_dump)
return
train(net, compression_ctrl, train_data_loader, test_data_loader,
criterion, optimizer, config, lr_scheduler)