def worker(rank: int, world_size: int) -> None:
torch.distributed.init_process_group(backend="nccl",
init_method='tcp://127.0.0.1:8999',
world_size=world_size,
rank=rank)
model = TestModelWithChangedTrain(freezing_stages=1)
model.cuda()
model.to(rank)
nncf_config = NNCFConfig()
nncf_config.update({
"input_info": {
"sample_size": [1, 1, 30, 30]
},
"compression": {
"algorithm": "quantization",
"initializer": {
"range": {
"num_init_samples": 10
},
"batchnorm_adaptation": {
"num_bn_adaptation_samples": 10
}
}
}
})
dataloader = create_random_mock_dataloader(nncf_config, num_samples=10)
register_default_init_args(nncf_config, dataloader)
_, compressed_model = create_compressed_model(model, nncf_config)
# At this part the additional processes may be freezing
_ = torch.nn.parallel.DistributedDataParallel(compressed_model,
device_ids=[rank])
def test_quantization_configuration_stats(data):
config = get_basic_quantization_config()
config['compression']['ignored_scopes'] = data.ignored_scopes
config['input_info']['sample_size'] = [2, 3, 299, 299]
ctrl, _ = create_compressed_model(
test_models.Inception3(aux_logits=True, transform_input=True), config)
stats = ShareEdgesQuantizedDataPathStatisticsCollector(ctrl.model,
ctrl).collect()
for attr_name, expected_value in data.expected.items():
actual_value = as_dict(getattr(stats, attr_name))
assert expected_value == actual_value
def test_memory_consumption_stats(data):
config = get_basic_quantization_config()
config['compression']['initializer'].update(data.initializers)
config['compression']['weights'] = data.weights
config['compression']['ignored_scopes'] = data.ignored_scopes
config['target_device'] = data.target_device
ctrl, _ = create_compressed_model(test_models.AlexNet(), config)
stats = MemoryConsumptionStatisticsCollector(
ctrl.model, ctrl.weight_quantizers,
ctrl.non_weight_quantizers).collect()
for attr_name, expected_value in data.expected.items():
actual_value = getattr(stats, attr_name)
assert expected_value == pytest.approx(actual_value, rel=1e-2)
def test_quantization_share_and_bitwidth_distribution_stats(data):
config = get_basic_quantization_config()
config['compression']['initializer'].update(data.initializers)
config['compression']['activations'] = data.activations
config['compression']['weights'] = data.weights
config['compression']['ignored_scopes'] = data.ignored_scopes
config['target_device'] = data.target_device
ctrl, _ = create_compressed_model(test_models.AlexNet(), config)
nncf_stats = ctrl.statistics()
quantization_stats = nncf_stats.quantization
for attr_name, expected_value in data.expected.items():
actual_value = as_dict(getattr(quantization_stats, attr_name))
assert expected_value == actual_value
def create_model(config: SampleConfig, resuming_checkpoint: dict = None):
input_info_list = create_input_infos(config.nncf_config)
image_size = input_info_list[0].shape[-1]
ssd_net = build_ssd(config.model, config.ssd_params, image_size,
config.num_classes, config)
weights = config.get('weights')
if weights:
sd = torch.load(weights,
map_location='cpu',
pickle_module=restricted_pickle_module)
sd = sd["state_dict"]
load_state(ssd_net, sd)
ssd_net.to(config.device)
model_state_dict, compression_state = extract_model_and_compression_states(
resuming_checkpoint)
compression_ctrl, compressed_model = create_compressed_model(
ssd_net, config.nncf_config, compression_state)
if model_state_dict is not None:
load_state(compressed_model, model_state_dict, is_resume=True)
compressed_model, _ = prepare_model_for_execution(compressed_model, config)
compressed_model.train()
return compression_ctrl, compressed_model
def wrap_nncf_model(model,
cfg,
checkpoint_dict=None,
datamanager_for_init=None):
# Note that we require to import it here to avoid cyclic imports when import get_no_nncf_trace_context_manager
# from mobilenetv3
from torchreid.data.transforms import build_inference_transform
from nncf import NNCFConfig
from nncf.torch import create_compressed_model, load_state
from nncf.torch.initialization import register_default_init_args
from nncf.torch.dynamic_graph.io_handling import nncf_model_input
from nncf.torch.dynamic_graph.trace_tensor import TracedTensor
from nncf.torch.initialization import PTInitializingDataLoader
if checkpoint_dict is None:
checkpoint_path = cfg.model.load_weights
resuming_checkpoint = safe_load_checkpoint(
checkpoint_path, map_location=torch.device('cpu'))
else:
checkpoint_path = 'pretrained_dict'
resuming_checkpoint = checkpoint_dict
if datamanager_for_init is None and not is_nncf_state(resuming_checkpoint):
raise RuntimeError('Either datamanager_for_init or NNCF pre-trained '
'model checkpoint should be set')
nncf_metainfo = None
if is_nncf_state(resuming_checkpoint):
nncf_metainfo = _get_nncf_metainfo_from_state(resuming_checkpoint)
nncf_config_data = nncf_metainfo['nncf_config']
datamanager_for_init = None
logger.info(f'Read NNCF metainfo with NNCF config from the checkpoint:'
f'nncf_metainfo=\n{pformat(nncf_metainfo)}')
else:
resuming_checkpoint = None
nncf_config_data = cfg.get('nncf_config')
if nncf_config_data is None:
logger.info('Cannot read nncf_config from config file')
else:
logger.info(f' nncf_config=\n{pformat(nncf_config_data)}')
h, w = cfg.data.height, cfg.data.width
if not nncf_config_data:
logger.info('Using the default NNCF int8 quantization config')
nncf_config_data = get_default_nncf_compression_config(h, w)
# do it even if nncf_config_data is loaded from a checkpoint -- for the rare case when
# the width and height of the model's input was changed in the config
# and then finetuning of NNCF model is run
nncf_config_data.setdefault('input_info', {})
nncf_config_data['input_info']['sample_size'] = [1, 3, h, w]
nncf_config = NNCFConfig(nncf_config_data)
logger.info(f'nncf_config =\n{pformat(nncf_config)}')
if not nncf_metainfo:
nncf_metainfo = create_nncf_metainfo(enable_quantization=True,
enable_pruning=False,
nncf_config=nncf_config_data)
else:
# update it just to be on the safe side
nncf_metainfo['nncf_config'] = nncf_config_data
class ReidInitializeDataLoader(PTInitializingDataLoader):
def get_inputs(self, dataloader_output):
# define own InitializingDataLoader class using approach like
# parse_data_for_train and parse_data_for_eval in the class Engine
# dataloader_output[0] should be image here
args = (dataloader_output[0], )
return args, {}
@torch.no_grad()
def model_eval_fn(model):
"""
Runs evaluation of the model on the validation set and
returns the target metric value.
Used to evaluate the original model before compression
if NNCF-based accuracy-aware training is used.
"""
from torchreid.metrics.classification import evaluate_classification
if test_loader is None:
raise RuntimeError(
'Cannot perform a model evaluation on the validation '
'dataset since the validation data loader was not passed '
'to wrap_nncf_model')
model_type = get_model_attr(model, 'type')
targets = list(test_loader.keys())
use_gpu = cur_device.type == 'cuda'
for dataset_name in targets:
domain = 'source' if dataset_name in datamanager_for_init.sources else 'target'
print(f'##### Evaluating {dataset_name} ({domain}) #####')
if model_type == 'classification':
cmc, _, _ = evaluate_classification(
test_loader[dataset_name]['query'], model, use_gpu=use_gpu)
accuracy = cmc[0]
elif model_type == 'multilabel':
mAP, _, _, _, _, _, _ = evaluate_multilabel_classification(
test_loader[dataset_name]['query'], model, use_gpu=use_gpu)
accuracy = mAP
else:
raise ValueError(
f'Cannot perform a model evaluation on the validation dataset'
f'since the model has unsupported model_type {model_type or "None"}'
)
return accuracy
cur_device = next(model.parameters()).device
logger.info(f'NNCF: cur_device = {cur_device}')
if resuming_checkpoint is None:
logger.info(
'No NNCF checkpoint is provided -- register initialize data loader'
)
train_loader = datamanager_for_init.train_loader
test_loader = datamanager_for_init.test_loader
wrapped_loader = ReidInitializeDataLoader(train_loader)
nncf_config = register_default_init_args(nncf_config,
wrapped_loader,
model_eval_fn=model_eval_fn,
device=cur_device)
model_state_dict = None
compression_state = None
else:
model_state_dict, compression_state = extract_model_and_compression_states(
resuming_checkpoint)
transform = build_inference_transform(
cfg.data.height,
cfg.data.width,
norm_mean=cfg.data.norm_mean,
norm_std=cfg.data.norm_std,
)
def dummy_forward(model):
prev_training_state = model.training
model.eval()
input_img = random_image(cfg.data.height, cfg.data.width)
input_blob = transform(input_img).unsqueeze(0)
assert len(input_blob.size()) == 4
input_blob = input_blob.to(device=cur_device)
input_blob = nncf_model_input(input_blob)
model(input_blob)
model.train(prev_training_state)
def wrap_inputs(args, kwargs):
assert len(args) == 1
if isinstance(args[0], TracedTensor):
logger.info('wrap_inputs: do not wrap input TracedTensor')
return args, {}
return (nncf_model_input(args[0]), ), kwargs
model.dummy_forward_fn = dummy_forward
if 'log_dir' in nncf_config:
os.makedirs(nncf_config['log_dir'], exist_ok=True)
logger.info(f'nncf_config["log_dir"] = {nncf_config["log_dir"]}')
compression_ctrl, model = create_compressed_model(
model,
nncf_config,
dummy_forward_fn=dummy_forward,
wrap_inputs_fn=wrap_inputs,
compression_state=compression_state)
if model_state_dict:
logger.info(f'Loading NNCF model from {checkpoint_path}')
load_state(model, model_state_dict, is_resume=True)
return compression_ctrl, model, nncf_metainfo
def wrap_nncf_model(model,
cfg,
data_loader_for_init=None,
get_fake_input_func=None,
export=False):
"""
The function wraps mmaction model by NNCF
Note that the parameter `get_fake_input_func` should be the function `get_fake_input`
-- cannot import this function here explicitly
"""
check_nncf_is_enabled()
from nncf.config import NNCFConfig
from nncf.torch import (create_compressed_model,
register_default_init_args)
from nncf.torch.dynamic_graph.io_handling import nncf_model_input
from nncf.torch.dynamic_graph.trace_tensor import TracedTensor
from nncf.torch.initialization import DefaultInitializingDataLoader
class MMInitializeDataLoader(DefaultInitializingDataLoader):
def get_inputs(self, dataloader_output):
return (), dataloader_output
pathlib.Path(cfg.work_dir).mkdir(parents=True, exist_ok=True)
nncf_config = NNCFConfig(cfg.nncf_config)
logger = get_root_logger(cfg.log_level)
if data_loader_for_init:
wrapped_loader = MMInitializeDataLoader(data_loader_for_init)
nncf_config = register_default_init_args(
nncf_config,
wrapped_loader,
device=next(model.parameters()).device)
if cfg.get('resume_from'):
checkpoint_path = cfg.get('resume_from')
assert is_checkpoint_nncf(checkpoint_path), (
'It is possible to resume training with NNCF compression from NNCF checkpoints only. '
'Use "load_from" with non-compressed model for further compression by NNCF.'
)
elif cfg.get('load_from'):
checkpoint_path = cfg.get('load_from')
if not is_checkpoint_nncf(checkpoint_path):
checkpoint_path = None
logger.info('Received non-NNCF checkpoint to start training '
'-- initialization of NNCF fields will be done')
else:
checkpoint_path = None
if not data_loader_for_init and not checkpoint_path:
raise RuntimeError('Either data_loader_for_init or NNCF pre-trained '
'model checkpoint should be set')
if checkpoint_path:
logger.info(f'Loading NNCF checkpoint from {checkpoint_path}')
logger.info(
'Please, note that this first loading is made before addition of '
'NNCF FakeQuantize nodes to the model, so there may be some '
'warnings on unexpected keys')
resuming_state_dict = load_checkpoint(model, checkpoint_path)
logger.info(f'Loaded NNCF checkpoint from {checkpoint_path}')
else:
resuming_state_dict = None
if "nncf_compress_postprocessing" in cfg:
# NB: This parameter is used to choose if we should try to make NNCF compression
# for a whole model graph including postprocessing (`nncf_compress_postprocessing=True`),
# or make NNCF compression of the part of the model without postprocessing
# (`nncf_compress_postprocessing=False`).
# Our primary goal is to make NNCF compression of such big part of the model as
# possible, so `nncf_compress_postprocessing=True` is our primary choice, whereas
# `nncf_compress_postprocessing=False` is our fallback decision.
# When we manage to enable NNCF compression for sufficiently many models,
# we should keep one choice only.
nncf_compress_postprocessing = cfg.get('nncf_compress_postprocessing')
logger.debug('set should_compress_postprocessing='
f'{nncf_compress_postprocessing}')
else:
nncf_compress_postprocessing = True
def _get_fake_data_for_forward(cfg, nncf_config, get_fake_input_func):
input_size = nncf_config.get("input_info").get('sample_size')
assert get_fake_input_func is not None
assert len(input_size) == 4 and input_size[0] == 1
H, W, C = input_size[2], input_size[3], input_size[1]
device = next(model.parameters()).device
with no_nncf_trace():
return get_fake_input_func(cfg,
orig_img_shape=tuple([H, W, C]),
device=device)
def dummy_forward(model):
fake_data = _get_fake_data_for_forward(cfg, nncf_config,
get_fake_input_func)
img = fake_data["imgs"]
img = nncf_model_input(img)
if export:
img, _, _ = model.reshape_input(imgs=img)
model(imgs=img)
else:
model(imgs=img, return_loss=False)
def wrap_inputs(args, kwargs):
# during dummy_forward
if not len(kwargs):
if not isinstance(args[0][0], TracedTensor):
args[0][0] = nncf_model_input(args[0][0])
return args, kwargs
# during building original graph
if not kwargs.get('return_loss') and kwargs.get('forward_export'):
return args, kwargs
# during model's forward
assert 'imgs' in kwargs, 'During model forward imgs must be in kwargs'
img = kwargs['imgs']
if isinstance(img, list):
assert len(img) == 1, 'Input list must have a length 1'
assert torch.is_tensor(
img[0]), 'Input for a model must be a tensor'
if not isinstance(img[0], TracedTensor):
img[0] = nncf_model_input(img[0])
else:
assert torch.is_tensor(img), 'Input for a model must be a tensor'
if not isinstance(img, TracedTensor):
img = nncf_model_input(img)
kwargs['imgs'] = img
return args, kwargs
model.dummy_forward_fn = dummy_forward
if 'log_dir' in nncf_config:
os.makedirs(nncf_config['log_dir'], exist_ok=True)
compression_ctrl, model = create_compressed_model(
model,
nncf_config,
dummy_forward_fn=dummy_forward,
wrap_inputs_fn=wrap_inputs,
compression_state=resuming_state_dict)
return compression_ctrl, model
def staged_quantization_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)
set_seed(config)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(config.device)
model_name = config['model']
is_inception = 'inception' in model_name
train_criterion_fn = inception_criterion_fn if is_inception else default_criterion_fn
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)
is_export_only = 'export' in config.mode and (
'train' not in config.mode and 'test' not in config.mode)
if is_export_only:
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, init_loader = create_data_loaders(
config, train_dataset, val_dataset)
def autoq_eval_fn(model, eval_loader):
_, top5, _ = validate(eval_loader, model, criterion, config)
return top5
nncf_config = register_default_init_args(
nncf_config,
init_loader,
criterion=criterion,
criterion_fn=train_criterion_fn,
autoq_eval_fn=autoq_eval_fn,
val_loader=val_loader,
device=config.device)
# 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_checkpoint = None
if resuming_checkpoint_path is not None:
resuming_checkpoint = load_resuming_checkpoint(
resuming_checkpoint_path)
model_state_dict, compression_state = extract_model_and_compression_states(
resuming_checkpoint)
compression_ctrl, model = create_compressed_model(model, nncf_config,
compression_state)
if model_state_dict is not None:
load_state(model, model_state_dict, is_resume=True)
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()
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'])
if 'train' in config.mode:
optimizer.load_state_dict(resuming_checkpoint['optimizer'])
optimizer_scheduler.load_state_dict(
resuming_checkpoint['optimizer_scheduler'])
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))
log_common_mlflow_params(config)
if is_export_only:
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 is_main_process():
statistics = compression_ctrl.statistics()
logger.info(statistics.to_str())
if 'train' in config.mode:
batch_multiplier = (quantization_config.get("params", {})).get(
"batch_multiplier", 1)
train_staged(config, compression_ctrl, model, criterion,
train_criterion_fn, optimizer_scheduler, model_name,
optimizer, train_loader, train_sampler, val_loader,
kd_loss_calculator, batch_multiplier, best_acc1)
if 'test' in config.mode:
validate(val_loader, model, criterion, config)
if 'export' in config.mode:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
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)
set_seed(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)
is_export_only = 'export' in config.mode and (
'train' not in config.mode and 'test' not in config.mode)
if is_export_only:
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)
model_eval_fn = functools.partial(autoq_test_fn,
eval_loader=val_loader)
nncf_config = register_default_init_args(nncf_config,
init_loader,
criterion=criterion,
criterion_fn=criterion_fn,
autoq_eval_fn=autoq_test_fn,
val_loader=val_loader,
model_eval_fn=model_eval_fn,
device=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_checkpoint = None
if resuming_checkpoint_path is not None:
resuming_checkpoint = load_resuming_checkpoint(
resuming_checkpoint_path)
model_state_dict, compression_state = extract_model_and_compression_states(
resuming_checkpoint)
compression_ctrl, model = create_compressed_model(model, nncf_config,
compression_state)
if model_state_dict is not None:
load_state(model, model_state_dict, is_resume=True)
model, model_without_dp = prepare_model_for_execution(model, config)
if config.distributed:
compression_ctrl.distributed()
log_common_mlflow_params(config)
if is_export_only:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))
return
if is_main_process():
statistics = compression_ctrl.statistics()
logger.info(statistics.to_str())
if is_accuracy_aware_training(config) and 'train' in config.mode:
def validate_fn(model, epoch):
return test(model, val_loader, criterion, color_encoding, config)
# training function that trains the model for one epoch (full training dataset pass)
# it is assumed that all the NNCF-related methods are properly called inside of
# this function (like e.g. the step and epoch_step methods of the compression scheduler)
def train_epoch_fn(compression_ctrl, model, optimizer, **kwargs):
ignore_index = None
ignore_unlabeled = config.get("ignore_unlabeled", True)
if ignore_unlabeled and ('unlabeled' in color_encoding):
ignore_index = list(color_encoding).index('unlabeled')
metric = IoU(len(color_encoding), ignore_index=ignore_index)
train_obj = Train(model, train_loader, optimizer, criterion,
compression_ctrl, metric, config.device,
config.model)
train_obj.run_epoch(config.print_step)
# function that initializes optimizers & lr schedulers to start training
def configure_optimizers_fn():
optim_config = config.get('optimizer', {})
optim_params = optim_config.get('optimizer_params', {})
lr = optim_params.get("lr", 1e-4)
params_to_optimize = get_params_to_optimize(
model_without_dp, lr * 10, config)
optimizer, lr_scheduler = make_optimizer(params_to_optimize,
config)
return optimizer, lr_scheduler
acc_aware_training_loop = create_accuracy_aware_training_loop(
config, compression_ctrl)
model = acc_aware_training_loop.run(
model,
train_epoch_fn=train_epoch_fn,
validate_fn=validate_fn,
configure_optimizers_fn=configure_optimizers_fn,
tensorboard_writer=config.tb,
log_dir=config.log_dir)
elif 'train' in config.mode:
train(model, model_without_dp, compression_ctrl, train_loader,
val_loader, criterion, color_encoding, config,
resuming_checkpoint)
if 'test' in config.mode:
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)
if 'export' in config.mode:
compression_ctrl.export_model(config.to_onnx)
logger.info("Saved to {}".format(config.to_onnx))