def load_data(args, fixed_subset=False):
return apputils.load_data(args.dataset, os.path.expanduser(args.data),
args.batch_size, args.workers,
args.validation_split, args.deterministic,
args.effective_train_size,
args.effective_valid_size,
args.effective_test_size, fixed_subset)
def automated_deep_compression(model, criterion, optimizer, loggers, args):
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size,
args.effective_test_size)
args.display_confusion = True
validate_fn = partial(test,
test_loader=test_loader,
criterion=criterion,
loggers=loggers,
args=args,
activations_collectors=None)
train_fn = partial(train,
train_loader=train_loader,
criterion=criterion,
loggers=loggers,
args=args)
save_checkpoint_fn = partial(apputils.save_checkpoint,
arch=args.arch,
dir=msglogger.logdir)
optimizer_data = {
'lr': args.lr,
'momentum': args.momentum,
'weight_decay': args.weight_decay
}
adc.do_adc(model, args, optimizer_data, validate_fn, save_checkpoint_fn,
train_fn)
def greedy(model, criterion, optimizer, loggers, args):
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size, args.effective_test_size)
test_fn = partial(test, test_loader=test_loader, criterion=criterion,
loggers=loggers, args=args, activations_collectors=None)
train_fn = partial(train, train_loader=train_loader, criterion=criterion, args=args)
assert args.greedy_target_density is not None
distiller.pruning.greedy_filter_pruning.greedy_pruner(model, args,
args.greedy_target_density,
args.greedy_pruning_step,
test_fn, train_fn)
def load_data(args, fixed_subset=False, sequential=False, load_train=True, load_val=True, load_test=True):
train_loader, val_loader, test_loader, _ = apputils.load_data(args.dataset,
os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size, args.effective_test_size,
fixed_subset, sequential)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler))
loaders = (train_loader, val_loader, test_loader)
flags = (load_train, load_val, load_test)
loaders = [loaders[i] for i, flag in enumerate(flags) if flag]
if len(loaders) == 1:
# Unpack the list for convinience
loaders = loaders[0]
return loaders
def main():
script_dir = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(script_dir, '..', '..'))
global msglogger
# Parse arguments
args = parser.get_parser().parse_args()
if args.epochs is None:
args.epochs = 90
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
msglogger = apputils.config_pylogger(os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir)
# Log various details about the execution environment. It is sometimes useful
# to refer to past experiment executions and this information may be useful.
apputils.log_execution_env_state(args.compress, msglogger.logdir, gitroot=module_path)
msglogger.debug("Distiller: %s", distiller.__version__)
start_epoch = 0
ending_epoch = args.epochs
perf_scores_history = []
if args.evaluate:
args.deterministic = True
if args.deterministic:
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
distiller.set_deterministic() # Use a well-known seed, for repeatability of experiments
else:
# Turn on CUDNN benchmark mode for best performance. This is usually "safe" for image
# classification models, as the input sizes don't change during the run
# See here: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/3
cudnn.benchmark = True
if args.cpu or not torch.cuda.is_available():
# Set GPU index to -1 if using CPU
args.device = 'cpu'
args.gpus = -1
else:
args.device = 'cuda'
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
raise ValueError('ERROR: Argument --gpus must be a comma-separated list of integers only')
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
raise ValueError('ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
# Infer the dataset from the model name
args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
args.num_classes = 10 if args.dataset == 'cifar10' else 1000
if args.earlyexit_thresholds:
args.num_exits = len(args.earlyexit_thresholds) + 1
args.loss_exits = [0] * args.num_exits
args.losses_exits = []
args.exiterrors = []
# Create the model
model = create_model(args.pretrained, args.dataset, args.arch,
parallel=not args.load_serialized, device_ids=args.gpus)
compression_scheduler = None
# Create a couple of logging backends. TensorBoardLogger writes log files in a format
# that can be read by Google's Tensor Board. PythonLogger writes to the Python logger.
tflogger = TensorBoardLogger(msglogger.logdir)
pylogger = PythonLogger(msglogger)
# capture thresholds for early-exit training
if args.earlyexit_thresholds:
msglogger.info('=> using early-exit threshold values of %s', args.earlyexit_thresholds)
# TODO(barrh): args.deprecated_resume is deprecated since v0.3.1
if args.deprecated_resume:
msglogger.warning('The "--resume" flag is deprecated. Please use "--resume-from=YOUR_PATH" instead.')
if not args.reset_optimizer:
msglogger.warning('If you wish to also reset the optimizer, call with: --reset-optimizer')
args.reset_optimizer = True
args.resumed_checkpoint_path = args.deprecated_resume
# We can optionally resume from a checkpoint
optimizer = None
if args.resumed_checkpoint_path:
model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint(
model, args.resumed_checkpoint_path, model_device=args.device)
elif args.load_model_path:
model = apputils.load_lean_checkpoint(model, args.load_model_path,
model_device=args.device)
if args.reset_optimizer:
start_epoch = 0
if optimizer is not None:
optimizer = None
msglogger.info('\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0')
# Define loss function (criterion)
criterion = nn.CrossEntropyLoss().to(args.device)
if optimizer is None:
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
if args.AMC:
return automated_deep_compression(model, criterion, optimizer, pylogger, args)
if args.greedy:
return greedy(model, criterion, optimizer, pylogger, args)
# This sample application can be invoked to produce various summary reports.
if args.summary:
return summarize_model(model, args.dataset, which_summary=args.summary)
activations_collectors = create_activation_stats_collectors(model, *args.activation_stats)
if args.qe_calibration:
msglogger.info('Quantization calibration stats collection enabled:')
msglogger.info('\tStats will be collected for {:.1%} of test dataset'.format(args.qe_calibration))
msglogger.info('\tSetting constant seeds and converting model to serialized execution')
distiller.set_deterministic()
model = distiller.make_non_parallel_copy(model)
activations_collectors.update(create_quantization_stats_collector(model))
args.evaluate = True
args.effective_test_size = args.qe_calibration
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size, args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler))
if args.sensitivity is not None:
sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2])
return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities)
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args,
compression_scheduler)
if args.compress:
# The main use-case for this sample application is CNN compression. Compression
# requires a compression schedule configuration file in YAML.
compression_scheduler = distiller.file_config(model, optimizer, args.compress, compression_scheduler,
(start_epoch-1) if args.resumed_checkpoint_path else None)
# Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer)
model.to(args.device)
elif compression_scheduler is None:
compression_scheduler = distiller.CompressionScheduler(model)
if args.thinnify:
#zeros_mask_dict = distiller.create_model_masks_dict(model)
assert args.resumed_checkpoint_path is not None, \
"You must use --resume-from to provide a checkpoint file to thinnify"
distiller.remove_filters(model, compression_scheduler.zeros_mask_dict, args.arch, args.dataset, optimizer=None)
apputils.save_checkpoint(0, args.arch, model, optimizer=None, scheduler=compression_scheduler,
name="{}_thinned".format(args.resumed_checkpoint_path.replace(".pth.tar", "")),
dir=msglogger.logdir)
print("Note: your model may have collapsed to random inference, so you may want to fine-tune")
return
args.kd_policy = None
if args.kd_teacher:
teacher = create_model(args.kd_pretrained, args.dataset, args.kd_teacher, device_ids=args.gpus)
if args.kd_resume:
teacher = apputils.load_lean_checkpoint(teacher, args.kd_resume)
dlw = distiller.DistillationLossWeights(args.kd_distill_wt, args.kd_student_wt, args.kd_teacher_wt)
args.kd_policy = distiller.KnowledgeDistillationPolicy(model, teacher, args.kd_temp, dlw)
compression_scheduler.add_policy(args.kd_policy, starting_epoch=args.kd_start_epoch, ending_epoch=args.epochs,
frequency=1)
msglogger.info('\nStudent-Teacher knowledge distillation enabled:')
msglogger.info('\tTeacher Model: %s', args.kd_teacher)
msglogger.info('\tTemperature: %s', args.kd_temp)
msglogger.info('\tLoss Weights (distillation | student | teacher): %s',
' | '.join(['{:.2f}'.format(val) for val in dlw]))
msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch)
if start_epoch >= ending_epoch:
msglogger.error(
'epoch count is too low, starting epoch is {} but total epochs set to {}'.format(
start_epoch, ending_epoch))
raise ValueError('Epochs parameter is too low. Nothing to do.')
for epoch in range(start_epoch, ending_epoch):
# This is the main training loop.
msglogger.info('\n')
if compression_scheduler:
compression_scheduler.on_epoch_begin(epoch,
metrics=(vloss if (epoch != start_epoch) else 10**6))
# Train for one epoch
with collectors_context(activations_collectors["train"]) as collectors:
train(train_loader, model, criterion, optimizer, epoch, compression_scheduler,
loggers=[tflogger, pylogger], args=args)
distiller.log_weights_sparsity(model, epoch, loggers=[tflogger, pylogger])
distiller.log_activation_statsitics(epoch, "train", loggers=[tflogger],
collector=collectors["sparsity"])
if args.masks_sparsity:
msglogger.info(distiller.masks_sparsity_tbl_summary(model, compression_scheduler))
# evaluate on validation set
with collectors_context(activations_collectors["valid"]) as collectors:
top1, top5, vloss = validate(val_loader, model, criterion, [pylogger], args, epoch)
distiller.log_activation_statsitics(epoch, "valid", loggers=[tflogger],
collector=collectors["sparsity"])
save_collectors_data(collectors, msglogger.logdir)
stats = ('Performance/Validation/',
OrderedDict([('Loss', vloss),
('Top1', top1),
('Top5', top5)]))
distiller.log_training_progress(stats, None, epoch, steps_completed=0, total_steps=1, log_freq=1,
loggers=[tflogger])
if compression_scheduler:
compression_scheduler.on_epoch_end(epoch, optimizer)
# Update the list of top scores achieved so far, and save the checkpoint
update_training_scores_history(perf_scores_history, model, top1, top5, epoch, args.num_best_scores)
is_best = epoch == perf_scores_history[0].epoch
checkpoint_extras = {'current_top1': top1,
'best_top1': perf_scores_history[0].top1,
'best_epoch': perf_scores_history[0].epoch}
apputils.save_checkpoint(epoch, args.arch, model, optimizer=optimizer, scheduler=compression_scheduler,
extras=checkpoint_extras, is_best=is_best, name=args.name, dir=msglogger.logdir)
# Finally run results on the test set
test(test_loader, model, criterion, [pylogger], activations_collectors, args=args)
def main():
script_dir = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(script_dir, '..', '..'))
global msglogger
# Parse arguments
args = parser.get_parser().parse_args()
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
msglogger = apputils.config_pylogger(
os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir)
# Log various details about the execution environment. It is sometimes useful
# to refer to past experiment executions and this information may be useful.
apputils.log_execution_env_state(args.compress,
msglogger.logdir,
gitroot=module_path)
msglogger.debug("Distiller: %s", distiller.__version__)
start_epoch = 0
best_epochs = [
distiller.MutableNamedTuple({
'epoch': 0,
'top1': 0,
'sparsity': 0
}) for i in range(args.num_best_scores)
]
if args.deterministic:
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
# In Pytorch, support for deterministic execution is still a bit clunky.
if args.workers > 1:
msglogger.error(
'ERROR: Setting --deterministic requires setting --workers/-j to 0 or 1'
)
exit(1)
# Use a well-known seed, for repeatability of experiments
distiller.set_deterministic()
else:
# This issue: https://github.com/pytorch/pytorch/issues/3659
# Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that
# results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled.
cudnn.benchmark = True
if args.cpu or not torch.cuda.is_available():
# Set GPU index to -1 if using CPU
args.device = 'cpu'
args.gpus = -1
else:
args.device = 'cuda'
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
msglogger.error(
'ERROR: Argument --gpus must be a comma-separated list of integers only'
)
exit(1)
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
msglogger.error(
'ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
exit(1)
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
# Infer the dataset from the model name
args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
args.num_classes = 10 if args.dataset == 'cifar10' else 1000
if args.earlyexit_thresholds:
args.num_exits = len(args.earlyexit_thresholds) + 1
args.loss_exits = [0] * args.num_exits
args.losses_exits = []
args.exiterrors = []
# Create the model
model = create_model(args.pretrained,
args.dataset,
args.arch,
parallel=not args.load_serialized,
device_ids=args.gpus)
compression_scheduler = None
# Create a couple of logging backends. TensorBoardLogger writes log files in a format
# that can be read by Google's Tensor Board. PythonLogger writes to the Python logger.
tflogger = TensorBoardLogger(msglogger.logdir)
pylogger = PythonLogger(msglogger)
# capture thresholds for early-exit training
if args.earlyexit_thresholds:
msglogger.info('=> using early-exit threshold values of %s',
args.earlyexit_thresholds)
# We can optionally resume from a checkpoint
if args.resume:
model, compression_scheduler, start_epoch = apputils.load_checkpoint(
model, chkpt_file=args.resume)
model.to(args.device)
# Define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().to(args.device)
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
if args.AMC:
return automated_deep_compression(model, criterion, optimizer,
pylogger, args)
if args.greedy:
return greedy(model, criterion, optimizer, pylogger, args)
# This sample application can be invoked to produce various summary reports.
if args.summary:
return summarize_model(model, args.dataset, which_summary=args.summary)
activations_collectors = create_activation_stats_collectors(
model, *args.activation_stats)
if args.qe_calibration:
msglogger.info('Quantization calibration stats collection enabled:')
msglogger.info(
'\tStats will be collected for {:.1%} of test dataset'.format(
args.qe_calibration))
msglogger.info(
'\tSetting constant seeds and converting model to serialized execution'
)
distiller.set_deterministic()
model = distiller.make_non_parallel_copy(model)
activations_collectors.update(
create_quantization_stats_collector(model))
args.evaluate = True
args.effective_test_size = args.qe_calibration
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size,
args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler),
len(test_loader.sampler))
if args.sensitivity is not None:
sensitivities = np.arange(args.sensitivity_range[0],
args.sensitivity_range[1],
args.sensitivity_range[2])
return sensitivity_analysis(model, criterion, test_loader, pylogger,
args, sensitivities)
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger,
activations_collectors, args,
compression_scheduler)
if args.compress:
# The main use-case for this sample application is CNN compression. Compression
# requires a compression schedule configuration file in YAML.
compression_scheduler = distiller.file_config(model, optimizer,
args.compress,
compression_scheduler)
# Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer)
model.to(args.device)
elif compression_scheduler is None:
compression_scheduler = distiller.CompressionScheduler(model)
if args.thinnify:
#zeros_mask_dict = distiller.create_model_masks_dict(model)
assert args.resume is not None, "You must use --resume to provide a checkpoint file to thinnify"
distiller.remove_filters(model,
compression_scheduler.zeros_mask_dict,
args.arch,
args.dataset,
optimizer=None)
apputils.save_checkpoint(0,
args.arch,
model,
optimizer=None,
scheduler=compression_scheduler,
name="{}_thinned".format(
args.resume.replace(".pth.tar", "")),
dir=msglogger.logdir)
print(
"Note: your model may have collapsed to random inference, so you may want to fine-tune"
)
return
args.kd_policy = None
if args.kd_teacher:
teacher = create_model(args.kd_pretrained,
args.dataset,
args.kd_teacher,
device_ids=args.gpus)
if args.kd_resume:
teacher, _, _ = apputils.load_checkpoint(teacher,
chkpt_file=args.kd_resume)
dlw = distiller.DistillationLossWeights(args.kd_distill_wt,
args.kd_student_wt,
args.kd_teacher_wt)
args.kd_policy = distiller.KnowledgeDistillationPolicy(
model, teacher, args.kd_temp, dlw)
compression_scheduler.add_policy(args.kd_policy,
starting_epoch=args.kd_start_epoch,
ending_epoch=args.epochs,
frequency=1)
msglogger.info('\nStudent-Teacher knowledge distillation enabled:')
msglogger.info('\tTeacher Model: %s', args.kd_teacher)
msglogger.info('\tTemperature: %s', args.kd_temp)
msglogger.info('\tLoss Weights (distillation | student | teacher): %s',
' | '.join(['{:.2f}'.format(val) for val in dlw]))
msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch)
for epoch in range(start_epoch, start_epoch + args.epochs):
# This is the main training loop.
msglogger.info('\n')
if compression_scheduler:
compression_scheduler.on_epoch_begin(epoch)
# Train for one epoch
with collectors_context(activations_collectors["train"]) as collectors:
train(train_loader,
model,
criterion,
optimizer,
epoch,
compression_scheduler,
loggers=[tflogger, pylogger],
args=args)
distiller.log_weights_sparsity(model,
epoch,
loggers=[tflogger, pylogger])
distiller.log_activation_statsitics(
epoch,
"train",
loggers=[tflogger],
collector=collectors["sparsity"])
if args.masks_sparsity:
msglogger.info(
distiller.masks_sparsity_tbl_summary(
model, compression_scheduler))
# evaluate on validation set
with collectors_context(activations_collectors["valid"]) as collectors:
top1, top5, vloss = validate(val_loader, model, criterion,
[pylogger], args, epoch)
distiller.log_activation_statsitics(
epoch,
"valid",
loggers=[tflogger],
collector=collectors["sparsity"])
save_collectors_data(collectors, msglogger.logdir)
stats = ('Peformance/Validation/',
OrderedDict([('Loss', vloss), ('Top1', top1),
('Top5', top5)]))
distiller.log_training_progress(stats,
None,
epoch,
steps_completed=0,
total_steps=1,
log_freq=1,
loggers=[tflogger])
if compression_scheduler:
compression_scheduler.on_epoch_end(epoch, optimizer)
# Update the list of top scores achieved so far, and save the checkpoint
is_best = top1 > best_epochs[-1].top1
if top1 > best_epochs[0].top1:
best_epochs[0].epoch = epoch
best_epochs[0].top1 = top1
# Keep best_epochs sorted such that best_epochs[0] is the lowest top1 in the best_epochs list
best_epochs = sorted(best_epochs, key=lambda score: score.top1)
for score in reversed(best_epochs):
if score.top1 > 0:
msglogger.info('==> Best Top1: %.3f on Epoch: %d', score.top1,
score.epoch)
apputils.save_checkpoint(epoch, args.arch, model, optimizer,
compression_scheduler, best_epochs[-1].top1,
is_best, args.name, msglogger.logdir)
# Finally run results on the test set
test(test_loader,
model,
criterion, [pylogger],
activations_collectors,
args=args)
def main():
script_dir = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(script_dir, '..', '..'))
global msglogger
# Parse arguments
args = parser.get_parser().parse_args()
if args.epochs is None:
args.epochs = 90
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
msglogger = apputils.config_pylogger(
os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir)
# Log various details about the execution environment. It is sometimes useful
# to refer to past experiment executions and this information may be useful.
apputils.log_execution_env_state(args.compress,
msglogger.logdir,
gitroot=module_path)
msglogger.debug("Distiller: %s", distiller.__version__)
start_epoch = 0
ending_epoch = args.epochs
perf_scores_history = []
if args.evaluate:
args.deterministic = True
if args.deterministic:
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
distiller.set_deterministic(
) # Use a well-known seed, for repeatability of experiments
else:
# Turn on CUDNN benchmark mode for best performance. This is usually "safe" for image
# classification models, as the input sizes don't change during the run
# See here: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/3
cudnn.benchmark = True
if args.cpu or not torch.cuda.is_available():
# Set GPU index to -1 if using CPU
args.device = 'cpu'
args.gpus = -1
else:
args.device = 'cuda'
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
raise ValueError(
'ERROR: Argument --gpus must be a comma-separated list of integers only'
)
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
raise ValueError(
'ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
# Infer the dataset from the model name
args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
args.num_classes = 10 if args.dataset == 'cifar10' else 1000
# Create the model
model = create_model(args.pretrained,
args.dataset,
args.arch,
parallel=not args.load_serialized,
device_ids=args.gpus)
compression_scheduler = None
optimizer = None
if args.resumed_checkpoint_path:
model, compression_scheduler, optimizer, start_epoch = apputils.load_checkpoint(
model,
args.resumed_checkpoint_path,
use_swa_model=args.use_swa_model,
model_device=args.device)
elif args.load_model_path:
model = apputils.load_lean_checkpoint(model,
args.load_model_path,
model_device=args.device)
if args.reset_optimizer:
start_epoch = 0
if optimizer is not None:
optimizer = None
msglogger.info(
'\nreset_optimizer flag set: Overriding resumed optimizer and resetting epoch count to 0'
)
# Create a couple of logging backends. TensorBoardLogger writes log files in a format
# that can be read by Google's Tensor Board. PythonLogger writes to the Python logger.
tflogger = TensorBoardLogger(msglogger.logdir)
pylogger = PythonLogger(msglogger)
# Define loss function (criterion)
criterion = nn.CrossEntropyLoss().to(args.device)
if optimizer is None:
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
# This sample application can be invoked to produce various summary reports.
if args.summary:
return summarize_model(model, args.dataset, which_summary=args.summary)
activations_collectors = create_activation_stats_collectors(
model, *args.activation_stats)
if args.qe_calibration:
msglogger.info('Quantization calibration stats collection enabled:')
msglogger.info(
'\tStats will be collected for {:.1%} of test dataset'.format(
args.qe_calibration))
msglogger.info(
'\tSetting constant seeds and converting model to serialized execution'
)
distiller.set_deterministic()
model = distiller.make_non_parallel_copy(model)
activations_collectors.update(
create_quantization_stats_collector(model))
args.evaluate = True
args.effective_test_size = args.qe_calibration
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size,
args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler),
len(test_loader.sampler))
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger,
activations_collectors, args,
compression_scheduler)
def main():
script_dir = os.path.dirname(__file__)
module_path = os.path.abspath(os.path.join(script_dir, '..', '..'))
global msglogger
# Parse arguments
args = parser.get_parser().parse_args()
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
msglogger = apputils.config_pylogger(
os.path.join(script_dir, 'logging.conf'), args.name, args.output_dir)
# Log various details about the execution environment. It is sometimes useful
# to refer to past experiment executions and this information may be useful.
apputils.log_execution_env_state(args.compress,
msglogger.logdir,
gitroot=module_path)
msglogger.debug("Distiller: %s", distiller.__version__)
start_epoch = 0
best_epochs = list()
if args.deterministic:
if args.loaders is None:
args.loaders = 1
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
# In Pytorch, support for deterministic execution is still a bit clunky.
if args.loaders > 1:
msglogger.error(
'ERROR: Setting --deterministic requires setting --loaders to 0 or 1'
)
exit(1)
# Use a well-known seed, for repeatability of experiments
distiller.set_deterministic()
else:
# This issue: https://github.com/pytorch/pytorch/issues/3659
# Implies that cudnn.benchmark should respect cudnn.deterministic, but empirically we see that
# results are not re-produced when benchmark is set. So enabling only if deterministic mode disabled.
cudnn.benchmark = True
if args.use_cpu or (args.gpus is None
and not torch.cuda.is_available()) or (args.gpus
== ''):
# Set GPU index to -1 if using CPU
args.device = 'cpu'
args.gpus = -1
else:
args.device = 'cuda'
if args.gpus is not None:
try:
args.gpus = [int(s) for s in args.gpus.split(',')]
except ValueError:
msglogger.error(
'ERROR: Argument --gpus must be a comma-separated list of integers only'
)
exit(1)
available_gpus = torch.cuda.device_count()
for dev_id in args.gpus:
if dev_id >= available_gpus:
msglogger.error(
'ERROR: GPU device ID {0} requested, but only {1} devices available'
.format(dev_id, available_gpus))
exit(1)
# Set default device in case the first one on the list != 0
torch.cuda.set_device(args.gpus[0])
if args.loaders is None:
active_gpus = args.gpus if args.gpus is not None else torch.cuda.device_count(
)
args.loaders = max(parser.DEFAULT_LOADERS_COUNT,
parser.DEFAULT_LOADERS_COUNT * active_gpus)
msglogger.debug('Number of data loaders set to: {}'.format(args.loaders))
# Infer the dataset from the model name
args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
args.num_classes = 10 if args.dataset == 'cifar10' else 1000
if args.earlyexit_thresholds:
args.num_exits = len(args.earlyexit_thresholds) + 1
args.loss_exits = [0] * args.num_exits
args.losses_exits = []
args.exiterrors = []
# Create the model
model = create_model(args.pretrained,
args.dataset,
args.arch,
parallel=not args.load_serialized,
device_ids=args.gpus)
compression_scheduler = None
# Create a couple of logging backends. TensorBoardLogger writes log files in a format
# that can be read by Google's Tensor Board. PythonLogger writes to the Python logger.
tflogger = TensorBoardLogger(msglogger.logdir)
pylogger = PythonLogger(msglogger)
# capture thresholds for early-exit training
if args.earlyexit_thresholds:
msglogger.info('=> using early-exit threshold values of %s',
args.earlyexit_thresholds)
# We can optionally resume from a checkpoint
optimizer = None
resumed_training_steps = None
if args.resume or args.load_state_dict:
if args.resume and not args.reset_optimizer:
# initiate SGD with dummy lr
optimizer = torch.optim.SGD(model.parameters(), lr=0.36787944117)
model, compression_scheduler, optimizer, start_epoch, resumed_training_steps = apputils.load_checkpoint(
model, args.resume or args.load_state_dict, optimizer=optimizer)
model.to(args.device)
# Define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().to(args.device)
if optimizer is not None:
# optimizer was resumed from checkpoint
# check if user has tried to set optimizer arguments
# if so, ignore arguments with a warning.
optimizer_group_args = [
'lr', 'learning-rate', 'momentum', 'weight-decay', 'wd'
]
user_optim_args = [
x for x in optimizer_group_args for arg in sys.argv
if arg.startswith('--' + x)
]
if user_optim_args:
msglogger.warning(
'{} optimizer arguments are ignored.'.format(user_optim_args))
msglogger.info(
'setting optimizer arguments when optimizer is resumed '
'from checkpoint is forbidden.')
else:
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
msglogger.info('Optimizer Type: %s', type(optimizer))
msglogger.info('Optimizer Args: %s', optimizer.defaults)
if args.AMC:
return automated_deep_compression(model, criterion, optimizer,
pylogger, args)
if args.greedy:
return greedy(model, criterion, optimizer, pylogger, args)
# This sample application can be invoked to produce various summary reports.
if args.summary:
return summarize_model(model, args.dataset, which_summary=args.summary)
activations_collectors = create_activation_stats_collectors(
model, *args.activation_stats)
if args.qe_calibration:
msglogger.info('Quantization calibration stats collection enabled:')
msglogger.info(
'\tStats will be collected for {:.1%} of test dataset'.format(
args.qe_calibration))
msglogger.info(
'\tSetting constant seeds and converting model to serialized execution'
)
distiller.set_deterministic()
model = distiller.make_non_parallel_copy(model)
activations_collectors.update(
create_quantization_stats_collector(model))
args.evaluate = True
args.effective_test_size = args.qe_calibration
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.loaders, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size,
args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler),
len(test_loader.sampler))
args.trainset_print_period = parser.getPrintPeriod(
args, len(train_loader.sampler), args.batch_size)
args.validset_print_period = parser.getPrintPeriod(args,
len(val_loader.sampler),
args.batch_size)
args.testset_print_period = parser.getPrintPeriod(args,
len(test_loader.sampler),
args.batch_size)
if args.sensitivity is not None:
sensitivities = np.arange(args.sensitivity_range[0],
args.sensitivity_range[1],
args.sensitivity_range[2])
return sensitivity_analysis(model, criterion, test_loader, pylogger,
args, sensitivities)
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger,
activations_collectors, args,
compression_scheduler)
if args.compress:
# The main use-case for this sample application is CNN compression. Compression
# requires a compression schedule configuration file in YAML.
compression_scheduler = distiller.file_config(
model, optimizer, args.compress, compression_scheduler,
(start_epoch - 1) if
(args.resume and not args.reset_optimizer) else None)
# Model is re-transferred to GPU in case parameters were added (e.g. PACTQuantizer)
model.to(args.device)
elif compression_scheduler is None:
compression_scheduler = distiller.CompressionScheduler(model)
if args.thinnify:
#zeros_mask_dict = distiller.create_model_masks_dict(model)
assert args.resume is not None, "You must use --resume to provide a checkpoint file to thinnify"
distiller.remove_filters(model,
compression_scheduler.zeros_mask_dict,
args.arch,
args.dataset,
optimizer=None)
apputils.save_checkpoint(0,
args.arch,
model,
optimizer=None,
scheduler=compression_scheduler,
name="{}_thinned".format(
args.resume.replace(".pth.tar", "")),
dir=msglogger.logdir)
print(
"Note: your model may have collapsed to random inference, so you may want to fine-tune"
)
return
args.kd_policy = None
if args.kd_teacher:
teacher = create_model(args.kd_pretrained,
args.dataset,
args.kd_teacher,
device_ids=args.gpus)
if args.kd_resume:
teacher = apputils.load_checkpoint(teacher,
chkpt_file=args.kd_resume)[0]
dlw = distiller.DistillationLossWeights(args.kd_distill_wt,
args.kd_student_wt,
args.kd_teacher_wt)
args.kd_policy = distiller.KnowledgeDistillationPolicy(
model, teacher, args.kd_temp, dlw)
compression_scheduler.add_policy(
args.kd_policy, range(args.kd_start_epoch, args.epochs, 1))
msglogger.info('\nStudent-Teacher knowledge distillation enabled:')
msglogger.info('\tTeacher Model: %s', args.kd_teacher)
msglogger.info('\tTemperature: %s', args.kd_temp)
msglogger.info('\tLoss Weights (distillation | student | teacher): %s',
' | '.join(['{:.2f}'.format(val) for val in dlw]))
msglogger.info('\tStarting from Epoch: %s', args.kd_start_epoch)
if getattr(compression_scheduler, 'global_policy_end_epoch',
None) is not None:
if compression_scheduler.global_policy_end_epoch >= (start_epoch +
args.epochs):
msglogger.warning(
'scheduler requires at least {} epochs, but only {} are sanctioned'
.format(compression_scheduler.global_policy_end_epoch,
args.epochs))
accumulated_training_steps = resumed_training_steps if resumed_training_steps is not None else 0
for epoch in range(start_epoch, start_epoch + args.epochs):
# This is the main training loop.
msglogger.info('\n')
if compression_scheduler:
compression_scheduler.on_epoch_begin(epoch)
# Train for one epoch
with collectors_context(activations_collectors["train"]) as collectors:
try:
train(train_loader,
model,
criterion,
optimizer,
epoch,
accumulated_training_steps,
compression_scheduler,
loggers=[tflogger, pylogger],
args=args)
except RuntimeError as e:
if ('cuda out of memory' in str(e).lower()):
msglogger.error(
'CUDA memory failure has been detected.\n'
'Sometimes it helps to decrease batch size.\n'
'e.g. Add the following flag to your call: --batch-size={}'
.format(args.batch_size // 10))
raise
distiller.log_weights_sparsity(model,
epoch,
loggers=[tflogger, pylogger])
distiller.log_activation_statsitics(
epoch,
"train",
loggers=[tflogger],
collector=collectors["sparsity"])
if args.masks_sparsity:
msglogger.info(
distiller.masks_sparsity_tbl_summary(
model, compression_scheduler))
accumulated_training_steps += math.ceil(
len(train_loader.sampler) / train_loader.batch_size)
# evaluate on validation set
with collectors_context(activations_collectors["valid"]) as collectors:
top1, top5, vloss = validate(val_loader, model, criterion,
[pylogger], args, epoch)
distiller.log_activation_statsitics(
epoch,
"valid",
loggers=[tflogger],
collector=collectors["sparsity"])
save_collectors_data(collectors, msglogger.logdir)
stats = ('Performance/Validation/',
OrderedDict([('Loss', vloss), ('Top1', top1),
('Top5', top5)]))
tflogger.log_training_progress(stats, epoch, None)
if compression_scheduler:
compression_scheduler.on_epoch_end(epoch, optimizer)
if getattr(compression_scheduler, 'global_policy_end_epoch',
None) is None or (
compression_scheduler.global_policy_end_epoch <= epoch):
# Update the list of top scores achieved since all policies have concluded
if top1 > 0:
best_epochs.append(
distiller.MutableNamedTuple({
'top1': top1,
'top5': top5,
'epoch': epoch
}))
# Keep best_epochs sorted from best to worst
# Sort by top1 first, secondary sort by top5, and so forth
best_epochs.sort(key=operator.attrgetter('top1', 'top5', 'epoch'),
reverse=True)
for score in best_epochs[:args.num_best_scores]:
msglogger.info('==> Best Top1: %.3f Top5: %.3f on epoch: %d',
score.top1, score.top5, score.epoch)
is_best = best_epochs and (epoch == best_epochs[0].epoch)
apputils.save_checkpoint(epoch, args.arch, model, optimizer,
compression_scheduler,
best_epochs[0].top1 if best_epochs else None,
is_best, args.name, msglogger.logdir,
accumulated_training_steps)
# Finally run results on the test set
test(test_loader,
model,
criterion, [pylogger],
activations_collectors,
args=args)
if args.qe_calibration:
msglogger.info('Quantization calibration stats collection enabled:')
msglogger.info('\tStats will be collected for {:.1%} of test dataset'.format(args.qe_calibration))
msglogger.info('\tSetting constant seeds and converting model to serialized execution')
distiller.set_deterministic()
model = distiller.make_non_parallel_copy(model)
activations_collectors.update(create_quantization_stats_collector(model))
args.evaluate = True
args.effective_test_size = args.qe_calibration
# Load the datasets: the dataset to load is inferred from the model name passed
# in args.arch. The default dataset is ImageNet, but if args.arch contains the
# substring "_cifar", then cifar10 is used.
train_loader, val_loader, test_loader, _ = apputils.load_data(
args.dataset, os.path.expanduser(args.data), args.batch_size,
args.workers, args.validation_split, args.deterministic,
args.effective_train_size, args.effective_valid_size, args.effective_test_size)
msglogger.info('Dataset sizes:\n\ttraining=%d\n\tvalidation=%d\n\ttest=%d',
len(train_loader.sampler), len(val_loader.sampler), len(test_loader.sampler))
if args.sensitivity is not None:
sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2])
return sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities)
if args.evaluate:
return evaluate_model(model, criterion, test_loader, pylogger, activations_collectors, args,
compression_scheduler)
if args.compress:
# The main use-case for this sample application is CNN compression. Compression
# requires a compression schedule configuration file in YAML.