def main():
car_train_dst = vision.datasets.StanfordCars('../data/StanfordCars',
split='train')
car_val_dst = vision.datasets.StanfordCars('../data/StanfordCars',
split='test')
aircraft_train_dst = vision.datasets.FGVCAircraft('../data/FGVCAircraft',
split='trainval')
aircraft_val_dst = vision.datasets.FGVCAircraft('../data/FGVCAircraft',
split='test')
car_teacher = vision.models.classification.resnet18(num_classes=196,
pretrained=False)
aircraft_teacher = vision.models.classification.resnet18(num_classes=102,
pretrained=False)
student = vision.models.classification.resnet18(num_classes=196 + 102,
pretrained=False)
car_teacher.load_state_dict(torch.load(args.car_ckpt))
aircraft_teacher.load_state_dict(torch.load(args.aircraft_ckpt))
train_transform = sT.Compose([
sT.RandomResizedCrop(224),
sT.RandomHorizontalFlip(),
sT.ToTensor(),
sT.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
val_transform = sT.Compose([
sT.Resize(256),
sT.CenterCrop(224),
sT.ToTensor(),
sT.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
car_train_dst.transform = aircraft_train_dst.transform = train_transform
car_val_dst.transform = aircraft_val_dst.transform = val_transform
car_metric = metrics.MetricCompose(
metric_dict={
'car_acc': metrics.Accuracy(attach_to=lambda o, t: (o[:, :196], t))
})
aircraft_metric = metrics.MetricCompose(metric_dict={
'aircraft_acc':
metrics.Accuracy(attach_to=lambda o, t: (o[:, 196:], t))
})
train_dst = torch.utils.data.ConcatDataset(
[car_train_dst, aircraft_train_dst])
train_loader = torch.utils.data.DataLoader(train_dst,
batch_size=32,
shuffle=True,
num_workers=4)
car_loader = torch.utils.data.DataLoader(car_val_dst,
batch_size=32,
shuffle=False,
num_workers=4)
aircraft_loader = torch.utils.data.DataLoader(aircraft_val_dst,
batch_size=32,
shuffle=False,
num_workers=4)
car_evaluator = engine.evaluator.BasicEvaluator(car_loader, car_metric)
aircraft_evaluator = engine.evaluator.BasicEvaluator(
aircraft_loader, aircraft_metric)
if args.ckpt is not None:
student.load_state_dict(torch.load(args.ckpt))
print("Load student model from %s" % args.ckpt)
if args.test_only:
results_car = car_evaluator.eval(student)
results_aircraft = aircraft_evaluator.eval(student)
print("Stanford Cars: %s" % (results_car))
print("FGVC Aircraft: %s" % (results_aircraft))
return
TOTAL_ITERS = len(train_loader) * 100
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
optim = torch.optim.Adam(student.parameters(),
lr=args.lr,
weight_decay=1e-4)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(optim,
T_max=TOTAL_ITERS)
trainer = amalgamation.CommonFeatureAmalgamator(
logger=utils.logger.get_logger('cfl'),
tb_writer=SummaryWriter(log_dir='run/cfl-%s' %
(time.asctime().replace(' ', '_'))))
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss', 'loss_kd',
'loss_amal', 'loss_recons',
'lr')))
trainer.add_callback(engine.DefaultEvents.AFTER_EPOCH,
callbacks=[
callbacks.EvalAndCkpt(model=student,
evaluator=car_evaluator,
metric_name='car_acc',
ckpt_prefix='cfl_car'),
callbacks.EvalAndCkpt(
model=student,
evaluator=aircraft_evaluator,
metric_name='aircraft_acc',
ckpt_prefix='cfl_aircraft'),
])
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
layer_groups = [(student.fc, car_teacher.fc, aircraft_teacher.fc)]
layer_channels = [(512, 512, 512)]
trainer.setup(student=student,
teachers=[car_teacher, aircraft_teacher],
layer_groups=layer_groups,
layer_channels=layer_channels,
dataloader=train_loader,
optimizer=optim,
device=device,
on_layer_input=True,
weights=[1., 10., 10.])
trainer.run(start_iter=0, max_iter=TOTAL_ITERS)
def main():
# PyTorch Part
num_classes = 11
model = vision.models.segmentation.deeplabv3_resnet50(num_classes=num_classes, pretrained_backbone=True)
train_dst = vision.datasets.CamVid(
'data/CamVid11', split='trainval', transforms=sT.Compose([
sT.Multi( sT.Resize(240), sT.Resize(240, interpolation=Image.NEAREST)),
sT.Sync( sT.RandomRotation(5), sT.RandomRotation(5)),
sT.Multi( sT.ColorJitter(0.2, 0.2, 0.2), None),
sT.Sync( sT.RandomCrop(240), sT.RandomCrop(240)),
sT.Sync( sT.RandomHorizontalFlip(), sT.RandomHorizontalFlip() ),
sT.Multi( sT.ToTensor(), sT.ToTensor( normalize=False, dtype=torch.long) ),
sT.Multi( sT.Normalize( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] ), sT.Lambda(lambd=lambda x: x.squeeze()) )
]) )
val_dst = vision.datasets.CamVid(
'data/CamVid11', split='test', transforms=sT.Compose([
sT.Multi( sT.Resize(240), sT.Resize(240, interpolation=Image.NEAREST)),
sT.Multi( sT.ToTensor(), sT.ToTensor( normalize=False, dtype=torch.long ) ),
sT.Multi( sT.Normalize( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] ), sT.Lambda(lambd=lambda x: x.squeeze()) )
]) )
train_loader = torch.utils.data.DataLoader( train_dst, batch_size=16, shuffle=True, num_workers=4 )
val_loader = torch.utils.data.DataLoader( val_dst, batch_size=16, num_workers=4 )
TOTAL_ITERS=len(train_loader) * 200
device = torch.device( 'cuda' if torch.cuda.is_available() else 'cpu' )
optim = torch.optim.SGD( model.parameters(), lr=0.01, momentum=0.9, weight_decay=1e-4 )
sched = torch.optim.lr_scheduler.CosineAnnealingLR( optim, T_max=TOTAL_ITERS )
# KAE Part
metric = kamal.tasks.StandardMetrics.segmentation(num_classes=num_classes)
evaluator = engine.evaluator.BasicEvaluator( dataloader=val_loader, metric=metric, progress=False )
task = kamal.tasks.StandardTask.segmentation()
trainer = engine.trainer.BasicTrainer(
logger=kamal.utils.logger.get_logger('camvid_seg_deeplab'),
tb_writer=SummaryWriter( log_dir='run/camvid_seg_deeplab-%s'%( time.asctime().replace( ' ', '_' ) ) )
)
trainer.setup( model=model,
task=task,
dataloader=train_loader,
optimizer=optim,
device=device )
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss', 'lr')))
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
trainer.add_callback(
engine.DefaultEvents.AFTER_EPOCH,
callbacks=[
callbacks.EvalAndCkpt(model=model, evaluator=evaluator, metric_name='miou', ckpt_prefix='camvid_seg_deeplabv3_resnet50'),
callbacks.VisualizeSegmentation(
model=model,
dataset=val_dst,
idx_list_or_num_vis=10,
normalizer=kamal.utils.Normalizer( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], reverse=True),
)])
#import matplotlib.pyplot as plt
#lr_finder = kamal.engine.lr_finder.LRFinder()
#best_lr = lr_finder.find( optim, model, trainer, lr_range=[1e-8, 1.0], max_iter=100, smooth_momentum=0.9 )
#fig = lr_finder.plot(polyfit=4)
#plt.savefig('lr_finder_deeplab.png')
#lr_finder.adjust_learning_rate(optim, best_lr)
trainer.run( start_iter=0, max_iter=TOTAL_ITERS )
def main():
# PyTorch Part
model = vision.models.segmentation.deeplabv3_resnet50(
num_classes=1, pretrained_backbone=True)
train_dst = vision.datasets.NYUv2(
'data/NYUv2',
split='train',
target_type='depth',
transforms=sT.Compose([
sT.Multi(sT.Resize(240), sT.Resize(240)),
sT.Sync(sT.RandomRotation(5), sT.RandomRotation(5)),
sT.Multi(sT.ColorJitter(0.2, 0.2, 0.2), None),
sT.Sync(sT.RandomCrop(240), sT.RandomCrop(240)),
sT.Sync(sT.RandomHorizontalFlip(), sT.RandomHorizontalFlip()),
sT.Multi(sT.ToTensor(),
sT.ToTensor(normalize=False, dtype=torch.float)),
sT.Multi(
sT.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
sT.Lambda(lambda x: x / 1000))
]))
val_dst = vision.datasets.NYUv2(
'data/NYUv2',
split='test',
target_type='depth',
transforms=sT.Compose([
sT.Multi(sT.Resize(240), sT.Resize(240)),
sT.Multi(sT.ToTensor(),
sT.ToTensor(normalize=False, dtype=torch.float)),
sT.Multi(
sT.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
sT.Lambda(lambda x: x / 1000))
]))
train_loader = torch.utils.data.DataLoader(train_dst,
batch_size=16,
shuffle=True,
num_workers=4)
val_loader = torch.utils.data.DataLoader(val_dst,
batch_size=16,
num_workers=4)
TOTAL_ITERS = len(train_loader) * 200
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
optim = torch.optim.SGD(model.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=1e-4)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(optim,
T_max=TOTAL_ITERS)
# KAE Part
metric = kamal.tasks.StandardMetrics.monocular_depth()
evaluator = engine.evaluator.BasicEvaluator(dataloader=val_loader,
metric=metric,
progress=False)
task = kamal.tasks.StandardTask.monocular_depth()
trainer = engine.trainer.BasicTrainer(
logger=kamal.utils.logger.get_logger('nyuv2_depth_deeplab'),
tb_writer=SummaryWriter(log_dir='run/nyuv2_depth_deeplab-%s' %
(time.asctime().replace(' ', '_'))))
trainer.setup(model=model,
task=task,
dataloader=train_loader,
optimizer=optim,
device=device)
trainer.add_callback(engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss',
'lr')))
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
trainer.add_callback(engine.DefaultEvents.AFTER_EPOCH,
callbacks=[
callbacks.EvalAndCkpt(model=model,
evaluator=evaluator,
metric_name='rmse',
metric_mode='min',
ckpt_prefix='nyuv2_depth'),
callbacks.VisualizeDepth(
model=model,
dataset=val_dst,
idx_list_or_num_vis=10,
max_depth=10,
normalizer=kamal.utils.Normalizer(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
reverse=True),
)
])
trainer.run(start_iter=0, max_iter=TOTAL_ITERS)
def main():
# Seg + Depth
model = MultiTaskSegNet(out_channel_list=[13, 1])
seg_teacher = vision.models.segmentation.segnet_vgg16_bn(
num_classes=13, pretrained_backbone=True)
depth_teacher = vision.models.segmentation.segnet_vgg16_bn(
num_classes=1, pretrained_backbone=True)
seg_teacher.load_state_dict(torch.load(args.seg_ckpt))
depth_teacher.load_state_dict(torch.load(args.depth_ckpt))
seg_train_dst = vision.datasets.NYUv2('../data/NYUv2',
split='train',
target_type='semantic')
seg_val_dst = vision.datasets.NYUv2('../data/NYUv2',
split='test',
target_type='semantic')
depth_train_dst = vision.datasets.NYUv2('../data/NYUv2',
split='train',
target_type='depth')
depth_val_dst = vision.datasets.NYUv2('../data/NYUv2',
split='test',
target_type='depth')
train_dst = vision.datasets.LabelConcatDataset(
datasets=[seg_train_dst, depth_train_dst],
transforms=sT.Compose([
sT.Multi(sT.Resize(240), sT.Resize(240,
interpolation=Image.NEAREST),
sT.Resize(240)),
sT.Sync(sT.RandomCrop(240), sT.RandomCrop(240),
sT.RandomCrop(240)),
sT.Sync(sT.RandomHorizontalFlip(), sT.RandomHorizontalFlip(),
sT.RandomHorizontalFlip()),
sT.Multi(sT.ToTensor(),
sT.ToTensor(normalize=False, dtype=torch.long),
sT.ToTensor(normalize=False, dtype=torch.float)),
sT.Multi(
sT.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
sT.Lambda(lambd=lambda x: x.squeeze()),
sT.Lambda(lambd=lambda x: x / 1e3))
]))
val_dst = vision.datasets.LabelConcatDataset(
datasets=[seg_val_dst, depth_val_dst],
transforms=sT.Compose([
sT.Multi(sT.Resize(240), sT.Resize(240,
interpolation=Image.NEAREST),
sT.Resize(240)),
sT.Multi(sT.ToTensor(),
sT.ToTensor(normalize=False, dtype=torch.long),
sT.ToTensor(normalize=False, dtype=torch.float)),
sT.Multi(
sT.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
sT.Lambda(lambd=lambda x: x.squeeze()),
sT.Lambda(lambd=lambda x: x / 1e3))
]))
train_loader = torch.utils.data.DataLoader(train_dst,
batch_size=16,
shuffle=True,
num_workers=4)
val_loader = torch.utils.data.DataLoader(val_dst,
batch_size=16,
num_workers=4)
TOTAL_ITERS = len(train_loader) * 200
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
optim = torch.optim.Adam(model.parameters(), lr=1e-4, weight_decay=1e-4)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(optim,
T_max=TOTAL_ITERS)
confusion_matrix = metrics.ConfusionMatrix(num_classes=13,
ignore_idx=255,
attach_to=0)
metric = metrics.MetricCompose(
metric_dict={
'acc': metrics.Accuracy(attach_to=0),
'cm': confusion_matrix,
'mIoU': metrics.mIoU(confusion_matrix),
'rmse': metrics.RootMeanSquaredError(attach_to=1)
})
evaluator = engine.evaluator.BasicEvaluator(dataloader=val_loader,
metric=metric,
progress=False)
task = [
kamal.tasks.StandardTask.distillation(attach_to=[0, 0]),
kamal.tasks.StandardTask.monocular_depth(attach_to=[1, 1])
]
trainer = engine.trainer.KDTrainer(
logger=kamal.utils.logger.get_logger('nyuv2_simple_kd'),
tb_writer=SummaryWriter(log_dir='run/nyuv2_simple_kd-%s' %
(time.asctime().replace(' ', '_'))))
trainer.setup(student=model,
teacher=[seg_teacher, depth_teacher],
task=task,
dataloader=train_loader,
optimizer=optim,
device=device)
trainer.add_callback(engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss',
'lr')))
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
trainer.add_callback(
engine.DefaultEvents.AFTER_EPOCH,
callbacks=[
callbacks.EvalAndCkpt(model=model,
evaluator=evaluator,
metric_name='rmse',
metric_mode='min',
ckpt_prefix='nyuv2_simple_kd'),
callbacks.VisualizeSegmentation(
model=model,
dataset=val_dst,
idx_list_or_num_vis=10,
attach_to=0,
normalizer=kamal.utils.Normalizer(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
reverse=True),
),
callbacks.VisualizeDepth(
model=model,
dataset=val_dst,
idx_list_or_num_vis=10,
max_depth=10,
attach_to=1,
normalizer=kamal.utils.Normalizer(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
reverse=True),
),
])
trainer.run(start_iter=0, max_iter=TOTAL_ITERS)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--lr', type=float, default=0.05)
parser.add_argument('--t_model_path', type=str, default='../pretrained/cifar100_wrn_40_2.pth')
parser.add_argument('--T', '--temperature', type=float,
default=4.0, help='temperature for KD distillation')
parser.add_argument('-r', '--gamma', type=float, default=1)
parser.add_argument('-a', '--alpha', type=float, default=None)
parser.add_argument('-b', '--beta', type=float, default=None)
parser.add_argument('--epochs', type=int, default=240)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--num_workers', type=int,
default=8, help='num of workers to use')
parser.add_argument('--distill', type=str, default='kd', choices=[
'kd', 'hint', 'attention', 'sp', 'cc', 'vid', 'svd', 'pkt', 'nst', 'rkd'])
# dataset
parser.add_argument('--data_root', type=str, default='../data/torchdata')
parser.add_argument('--img_size', type=int, default=32,
help='image size of datasets')
# hint layer
parser.add_argument('--hint_layer', default=2,
type=int, choices=[0, 1, 2, 3, 4])
# cc embed dim
parser.add_argument('--embed_dim', default=128,
type=int, help='feature dimension')
args = parser.parse_args()
# prepare data
cifar100_train = CIFAR100(args.data_root, train=True, download=True,
transform=T.Compose([
T.RandomCrop(32, padding=4),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.5071, 0.4867, 0.4408],
std=[0.2675, 0.2565, 0.2761])])
)
train_loader = torch.utils.data.DataLoader(
cifar100_train, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True
)
val_loader = torch.utils.data.DataLoader(
CIFAR100(args.data_root, train=False, download=True,
transform=T.Compose([
T.ToTensor(),
T.Normalize(mean=[0.5071, 0.4867, 0.4408],
std=[0.2675, 0.2565, 0.2761])])
), batch_size=args.batch_size, num_workers=args.num_workers
)
# prepare model
teacher = vision.models.classification.cifar.wrn.wrn_40_2(num_classes=100)
student = vision.models.classification.cifar.wrn.wrn_16_2(num_classes=100)
teacher.load_state_dict(torch.load(args.t_model_path))
print('[!] teacher loads weights from %s' % (args.t_model_path))
# Teacher eval
evaluator = engine.evaluator.BasicEvaluator(val_loader, metric=metrics.StandardTaskMetrics.classification())
teacher_scores = evaluator.eval(teacher)
print('[TEACHER] Acc=%.4f' % (teacher_scores['acc']))
# hook module feature
out_flags = [True, True, True, True, False]
tea_hooks = []
tea_layers = [teacher.conv1, teacher.block1,
teacher.block2, teacher.block3, teacher.fc]
for module in tea_layers:
hookfeat = engine.hooks.FeatureHook(module)
hookfeat.register()
tea_hooks.append(hookfeat)
stu_hooks = []
stu_layers = [student.conv1, student.block1,
student.block2, student.block3, student.fc]
for module in stu_layers:
hookfeat = engine.hooks.FeatureHook(module)
hookfeat.register()
stu_hooks.append(hookfeat)
# distiller setup
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
logger = utils.logger.get_logger('distill_%s' % (args.distill))
tb_writer = SummaryWriter(log_dir='run/distill_%s-%s' %
(args.distill, time.asctime().replace(' ', '_')))
if args.distill == 'kd':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.KDDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader,
optimizer=optimizer, T=args.T, gamma=args.gamma, alpha=args.alpha, device=device)
if args.distill == 'attention':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.AttentionDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'nst':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.NSTDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'sp':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.SPDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'rkd':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.RKDDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'pkt':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.PKTDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'svd':
optimizer = torch.optim.SGD(student.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.SVDDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'vid':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data = torch.randn(1, 3, args.img_size, args.img_size)
data = data.to(device)
student, teacher = student.to(device), teacher.to(device)
teacher(data)
student(data)
s_n = [f.feat_out.shape[1] if flag else f.feat_in.shape[1]
for (f, flag) in zip(stu_hooks[1:-1], out_flags)]
t_n = [f.feat_out.shape[1] if flag else f.feat_in.shape[1]
for (f, flag) in zip(tea_hooks[1:-1], out_flags)]
train_list = nn.ModuleList([student])
VIDRegressor_l = nn.ModuleList()
for s, t in zip(s_n, t_n):
vid_r = slim.VIDRegressor(s, t, t)
VIDRegressor_l.append(vid_r)
train_list.append(vid_r)
optimizer = torch.optim.SGD(train_list.parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.VIDDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, regressor_l=VIDRegressor_l, T=args.T,
gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'hint':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data = torch.randn(1, 3, args.img_size, args.img_size)
data = data.to(device)
student, teacher = student.to(device), teacher.to(device)
teacher(data), student(data)
feat_t = [f.feat_out if flag else f.feat_in for (
f, flag) in zip(tea_hooks, out_flags)]
feat_s = [f.feat_out if flag else f.feat_in for (
f, flag) in zip(stu_hooks, out_flags)]
fitnet = slim.Regressor(
feat_s[args.hint_layer].shape, feat_t[args.hint_layer].shape)
optimizer = torch.optim.SGD(nn.ModuleList(
[student, fitnet]).parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.HintDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, regressor=fitnet, hint_layer=args.hint_layer,
T=args.T, gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
if args.distill == 'cc':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data = torch.randn(1, 3, args.img_size, args.img_size)
data = data.to(device)
student, teacher = student.to(device), teacher.to(device)
student, teacher = student.to(device), teacher.to(device)
teacher(data), student(data)
feat_t = [f.feat_out if flag else f.feat_in for (
f, flag) in zip(tea_hooks, out_flags)]
feat_s = [f.feat_out if flag else f.feat_in for (
f, flag) in zip(stu_hooks, out_flags)]
embed_s = slim.LinearEmbed(feat_s[-1].shape[1], args.embed_dim)
embed_t = slim.LinearEmbed(feat_t[-1].shape[1], args.embed_dim)
optimizer = torch.optim.SGD(nn.ModuleList([student, embed_s, embed_t]).parameters(
), lr=args.lr, momentum=0.9, weight_decay=5e-4)
distiller = slim.CCDistiller(logger, tb_writer)
distiller.setup(student=student, teacher=teacher, dataloader=train_loader, optimizer=optimizer, embed_s=embed_s, embed_t=embed_t,
T=args.T, gamma=args.gamma, alpha=args.alpha, beta=args.beta, stu_hooks=stu_hooks, tea_hooks=tea_hooks, out_flags=out_flags, device=device)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, len(train_loader)*args.epochs)
distiller.add_callback(
engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss', 'loss_kld', 'loss_ce', 'loss_additional', 'lr')))
distiller.add_callback(
engine.DefaultEvents.AFTER_EPOCH,
callbacks=callbacks.EvalAndCkpt(model=student, evaluator=evaluator, metric_name='acc', ckpt_prefix=args.distill) )
distiller.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[scheduler]))
distiller.run(start_iter=0, max_iter=len(train_loader)*args.epochs)
def main():
# Pytorch Part
if args.dataset == 'stanford_dogs':
num_classes = 120
train_dst = vision.datasets.StanfordDogs('data/StanfordDogs',
split='train')
val_dst = vision.datasets.StanfordDogs('data/StanfordDogs',
split='test')
elif args.dataset == 'cub200':
num_classes = 200
train_dst = vision.datasets.CUB200('data/CUB200', split='train')
val_dst = vision.datasets.CUB200('data/CUB200', split='test')
elif args.dataset == 'fgvc_aircraft':
num_classes = 102
train_dst = vision.datasets.FGVCAircraft('data/FGVCAircraft/',
split='trainval')
val_dst = vision.datasets.FGVCAircraft('data/FGVCAircraft/',
split='test')
elif args.dataset == 'stanford_cars':
num_classes = 196
train_dst = vision.datasets.StanfordCars('data/StanfordCars/',
split='train')
val_dst = vision.datasets.StanfordCars('data/StanfordCars/',
split='test')
else:
raise NotImplementedError
model = vision.models.classification.resnet18(num_classes=num_classes,
pretrained=args.pretrained)
train_dst.transform = sT.Compose([
sT.RandomResizedCrop(224),
sT.RandomHorizontalFlip(),
sT.ToTensor(),
sT.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
val_dst.transform = sT.Compose([
sT.Resize(256),
sT.CenterCrop(224),
sT.ToTensor(),
sT.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
train_loader = torch.utils.data.DataLoader(train_dst,
batch_size=32,
shuffle=True,
num_workers=4)
val_loader = torch.utils.data.DataLoader(val_dst,
batch_size=32,
num_workers=4)
TOTAL_ITERS = len(train_loader) * args.epochs
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
optim = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(optim,
T_max=TOTAL_ITERS)
# KAE Part
# Predefined task & metrics
task = kamal.tasks.StandardTask.classification()
metric = kamal.tasks.StandardMetrics.classification()
# Evaluator and Trainer
evaluator = engine.evaluator.BasicEvaluator(val_loader,
metric=metric,
progress=True)
trainer = engine.trainer.BasicTrainer(
logger=kamal.utils.logger.get_logger(args.dataset),
tb_writer=SummaryWriter(
log_dir='run/%s-%s' %
(args.dataset, time.asctime().replace(' ', '_'))))
# setup trainer
trainer.setup(model=model,
task=task,
dataloader=train_loader,
optimizer=optim,
device=device)
trainer.add_callback(engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss',
'lr')))
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
ckpt_callback = trainer.add_callback(engine.DefaultEvents.AFTER_EPOCH,
callbacks=callbacks.EvalAndCkpt(
model=model,
evaluator=evaluator,
metric_name='acc',
ckpt_prefix=args.dataset))
trainer.run(start_iter=0, max_iter=TOTAL_ITERS)
ckpt_callback.callback.final_ckpt(ckpt_dir='pretrained', add_md5=True)
def main():
car_train_dst = vision.datasets.StanfordCars( '../data/StanfordCars', split='train')
car_val_dst = vision.datasets.StanfordCars( '../data/StanfordCars', split='test')
aircraft_train_dst = vision.datasets.FGVCAircraft( '../data/FGVCAircraft', split='trainval')
aircraft_val_dst = vision.datasets.FGVCAircraft( '../data/FGVCAircraft', split='test')
dog_train_dst = vision.datasets.StanfordDogs( '../data/StanfordDogs', split='train')
dog_val_dst = vision.datasets.StanfordDogs( '../data/StanfordDogs', split='test')
cub_train_dst = vision.datasets.CUB200( '../data/CUB200', split='train')
cub_val_dst = vision.datasets.CUB200( '../data/CUB200', split='test')
#car_teacher = vision.models.classification.resnet18( num_classes=196, pretrained=False )
#aircraft_teacher = vision.models.classification.resnet18( num_classes=102, pretrained=False )
#dog_teacher = vision.models.classification.resnet18( num_classes=120, pretrained=False )
#cub_teacher = vision.models.classification.resnet18( num_classes=200, pretrained=False )
student = vision.models.classification.resnet18( num_classes=196+102+120+200, pretrained=False )
#car_teacher.load_state_dict( torch.load( args.car_ckpt ) )
#aircraft_teacher.load_state_dict( torch.load( args.aircraft_ckpt ) )
#dog_teacher.load_state_dict( torch.load( args.dog_ckpt ) )
#cub_teacher.load_state_dict( torch.load( args.cub_ckpt ) )
train_transform = sT.Compose( [
sT.RandomResizedCrop(224),
sT.RandomHorizontalFlip(),
sT.ToTensor(),
sT.Normalize( mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225] )
] )
val_transform = sT.Compose( [
sT.Resize(256),
sT.CenterCrop( 224 ),
sT.ToTensor(),
sT.Normalize( mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225] )
] )
cub_train_dst.transform = dog_train_dst.transform = car_train_dst.transform = aircraft_train_dst.transform = train_transform
cub_val_dst.transform = dog_val_dst.transform = car_val_dst.transform = aircraft_val_dst.transform = val_transform
aircraft_train_dst.target_transform = lambda t: t+196
dog_train_dst.target_transform = lambda t: t+196+102
cub_train_dst.target_transform = lambda t: t+196+102+120
car_metric = metrics.MetricCompose(metric_dict={ 'car_acc': metrics.Accuracy(attach_to=lambda o, t: (o[:, :196],t) ) })
aircraft_metric = metrics.MetricCompose(metric_dict={ 'aircraft_acc': metrics.Accuracy(attach_to=lambda o, t: (o[:, 196:196+102],t) ) })
dog_metric = metrics.MetricCompose(metric_dict={ 'dog_acc': metrics.Accuracy(attach_to=lambda o, t: (o[:, 196+102:196+102+120],t) ) })
cub_metric = metrics.MetricCompose(metric_dict={ 'cub_acc': metrics.Accuracy(attach_to=lambda o, t: (o[:, 196+102+120:196+102+120+200],t) ) })
train_dst = torch.utils.data.ConcatDataset( [car_train_dst, aircraft_train_dst, dog_train_dst, cub_train_dst] )
train_loader = torch.utils.data.DataLoader( train_dst, batch_size=32, shuffle=True, num_workers=4 )
car_loader = torch.utils.data.DataLoader( car_val_dst, batch_size=32, shuffle=False, num_workers=2 )
aircraft_loader = torch.utils.data.DataLoader( aircraft_val_dst, batch_size=32, shuffle=False, num_workers=2 )
dog_loader = torch.utils.data.DataLoader( dog_val_dst, batch_size=32, shuffle=False, num_workers=2 )
cub_loader = torch.utils.data.DataLoader( cub_val_dst, batch_size=32, shuffle=False, num_workers=2 )
car_evaluator = engine.evaluator.BasicEvaluator( car_loader, car_metric )
aircraft_evaluator = engine.evaluator.BasicEvaluator( aircraft_loader, aircraft_metric )
dog_evaluator = engine.evaluator.BasicEvaluator( dog_loader, dog_metric )
cub_evaluator = engine.evaluator.BasicEvaluator( cub_loader, cub_metric )
if args.ckpt is not None:
student.load_state_dict( torch.load( args.ckpt ) )
print("Load student model from %s"%args.ckpt)
if args.test_only:
results_car = car_evaluator.eval( student )
results_aircraft = aircraft_evaluator.eval( student )
results_dog = dog_evaluator.eval( student )
results_cub = cub_evaluator.eval( student )
print("Stanford Cars: %s"%( results_car ))
print("FGVC Aircraft: %s"%( results_aircraft ))
print("Stanford Dogs: %s"%( results_dog ))
print("CUB200: %s"%( results_cub ))
return
TOTAL_ITERS=len(train_loader) * 100
device = torch.device( 'cuda' if torch.cuda.is_available() else 'cpu' )
optim = torch.optim.Adam( student.parameters(), lr=args.lr, weight_decay=1e-4)
sched = torch.optim.lr_scheduler.CosineAnnealingLR( optim, T_max=TOTAL_ITERS )
task = tasks.StandardTask.classification()
trainer = engine.trainer.BasicTrainer(
logger=utils.logger.get_logger('scratch-4'),
tb_writer=SummaryWriter( log_dir='run/scratch-4-%s'%( time.asctime().replace( ' ', '_' ) ) )
)
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP(every=10),
callbacks=callbacks.MetricsLogging(keys=('total_loss', 'lr')))
trainer.add_callback(
engine.DefaultEvents.AFTER_EPOCH,
callbacks=[
callbacks.EvalAndCkpt(model=student, evaluator=car_evaluator, metric_name='car_acc', ckpt_prefix='cfl_car'),
callbacks.EvalAndCkpt(model=student, evaluator=aircraft_evaluator, metric_name='aircraft_acc', ckpt_prefix='cfl_aircraft'),
callbacks.EvalAndCkpt(model=student, evaluator=dog_evaluator, metric_name='dog_acc', ckpt_prefix='cfl_dog'),
callbacks.EvalAndCkpt(model=student, evaluator=cub_evaluator, metric_name='cub_acc', ckpt_prefix='cfl_cub'),
] )
trainer.add_callback(
engine.DefaultEvents.AFTER_STEP,
callbacks=callbacks.LRSchedulerCallback(schedulers=[sched]))
trainer.setup( model=student,
task=task,
dataloader=train_loader,
optimizer=optim,
device=device )
trainer.run(start_iter=0, max_iter=TOTAL_ITERS)