def sensitivity_analysis(model, criterion, data_loader, loggers, args,
sparsities):
# This sample application can be invoked to execute Sensitivity Analysis on your
# model. The ouptut is saved to CSV and PNG.
msglogger.info("Running sensitivity tests")
if not isinstance(loggers, list):
loggers = [loggers]
test_fnc = partial(
classifier.test,
test_loader=data_loader,
criterion=criterion,
loggers=loggers,
args=args,
activations_collectors=classifier.create_activation_stats_collectors(
model))
which_params = [param_name for param_name, _ in model.named_parameters()]
sensitivity = distiller.perform_sensitivity_analysis(
model,
net_params=which_params,
sparsities=sparsities,
test_func=test_fnc,
group=args.sensitivity)
distiller.sensitivities_to_png(
sensitivity, os.path.join(msglogger.logdir, 'sensitivity.png'))
distiller.sensitivities_to_csv(
sensitivity, os.path.join(msglogger.logdir, 'sensitivity.csv'))
def handle_subapps(model, criterion, optimizer, compression_scheduler, pylogger, args):
def load_test_data(args):
test_loader = classifier.load_data(args, load_train=False, load_val=False, load_test=True)
return test_loader
do_exit = False
if args.greedy:
greedy(model, criterion, optimizer, pylogger, args)
do_exit = True
elif args.summary:
# This sample application can be invoked to produce various summary reports
for summary in args.summary:
distiller.model_summary(model, summary, args.dataset)
do_exit = True
elif args.export_onnx is not None:
distiller.export_img_classifier_to_onnx(model,
os.path.join(msglogger.logdir, args.export_onnx),
args.dataset, add_softmax=True, verbose=False)
do_exit = True
elif args.qe_calibration and not (args.evaluate and args.quantize_eval):
classifier.acts_quant_stats_collection(model, criterion, pylogger, args, save_to_file=True)
do_exit = True
elif args.activation_histograms:
classifier.acts_histogram_collection(model, criterion, pylogger, args)
do_exit = True
elif args.sensitivity is not None:
test_loader = load_test_data(args)
sensitivities = np.arange(*args.sensitivity_range)
sensitivity_analysis(model, criterion, test_loader, pylogger, args, sensitivities)
do_exit = True
elif args.evaluate:
if args.quantize_eval and args.qe_lapq:
from ptq_lapq import image_classifier_ptq_lapq
image_classifier_ptq_lapq(model, criterion, pylogger, args)
else:
test_loader = load_test_data(args)
classifier.evaluate_model(test_loader, model, criterion, pylogger,
classifier.create_activation_stats_collectors(model, *args.activation_stats),
args, scheduler=compression_scheduler)
do_exit = True
elif args.thinnify:
assert args.resumed_checkpoint_path is not None, \
"You must use --resume-from to provide a checkpoint file to thinnify"
distiller.contract_model(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)
msglogger.info("Note: if your model collapsed to random inference, you may want to fine-tune")
do_exit = True
return do_exit
def handle_subapps(model, criterion, optimizer, compression_scheduler,
pylogger, args):
if args.transfer or args.dataset in custom_datasets:
if args.arch == 'resnet34' or 'resnet20_cifar':
model.module.fc = nn.Linear(model.module.fc.in_features,
args.num_classes)
if args.resumed_checkpoint_path:
model, _, _, _ = apputils.load_checkpoint(
model,
args.resumed_checkpoint_path,
model_device=args.device)
model = model.module
model = nn.DataParallel(model, device_ids=args.gpus)
print(model.module.fc)
elif args.arch == 'vgg16':
model.classifier[6] = nn.Linear(model.classifier[6].in_features,
args.num_classes)
if args.resumed_checkpoint_path:
model, _, _, _ = apputils.load_checkpoint(
model,
args.resumed_checkpoint_path,
model_device=args.device)
model.to(0)
print(model.classifier[6])
def load_test_data(args):
test_loader = classifier.load_data(args,
load_train=False,
load_val=False,
load_test=True)
return test_loader
do_exit = False
if args.greedy:
greedy(model, criterion, optimizer, pylogger, args)
do_exit = True
elif args.summary:
# This sample application can be invoked to produce various summary reports
for summary in args.summary:
distiller.model_summary(model, summary, args.dataset)
do_exit = True
elif args.export_onnx is not None:
distiller.export_img_classifier_to_onnx(model,
os.path.join(
msglogger.logdir,
args.export_onnx),
args.dataset,
add_softmax=True,
verbose=False)
do_exit = True
elif args.qe_calibration:
classifier.acts_quant_stats_collection(model, criterion, pylogger,
args)
do_exit = True
elif args.activation_histograms:
classifier.acts_histogram_collection(model, criterion, pylogger, args)
do_exit = True
elif args.sensitivity is not None:
test_loader = load_test_data(args)
#sensitivities = np.arange(args.sensitivity_range[0], args.sensitivity_range[1], args.sensitivity_range[2])
sensitivities = np.arange(*args.sensitivity_range)
sensitivity_analysis(model, criterion, test_loader, pylogger, args,
sensitivities)
do_exit = True
elif args.evaluate:
test_loader = load_test_data(args)
activations_collectors = classifier.create_activation_stats_collectors(
model, *args.activation_stats)
classifier.evaluate_model(model, criterion, test_loader, pylogger,
activations_collectors, args,
compression_scheduler)
do_exit = True
elif 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)
msglogger.info(
"Note: if your model collapsed to random inference, you may want to fine-tune"
)
do_exit = True
return do_exit
