def main():
# set GPU ID
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
cudnn.benchmark = True
# check save path
save_path = args.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# make dataloader
train_loader, test_loader = dataset.get_loader(args)
# set model
if args.model == 'res':
model = resnet.ResNet18().cuda()
# set criterion
criterion = nn.CrossEntropyLoss().cuda()
# set optimizer (default:sgd)
optimizer = optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# set scheduler
scheduler = MultiStepLR(optimizer, milestones=[80, 120], gamma=0.1)
# make logger
train_logger = utils.Logger(os.path.join(save_path, 'train.log'))
test_logger = utils.Logger(os.path.join(save_path, 'test.log'))
# forgetting
forgetting_histroy = utils.Forgetting_Events(
data_size=len(train_loader.dataset))
# Start Train
for epoch in range(1, args.epochs + 1):
# scheduler
scheduler.step()
# Train
train(train_loader, model, criterion, optimizer, epoch,
forgetting_histroy, train_logger)
validate(test_loader, model, criterion, epoch, test_logger, 'test')
# Save Model each epoch
if epoch == int(args.epochs):
torch.save(
model.state_dict(),
os.path.join(save_path, '{0}_{1}.pth'.format('model', epoch)))
# Finish Train
torch.save(forgetting_histroy,
os.path.join(save_path, 'train_forgetting.pth'))
# Draw Plot
plot_curves.draw_plot(save_path)
def Modellist(model_name, num_classes=1000, use_attention=None):
if model_name == 'vgg11':
return vgg.VGG11(num_classes, use_attention=use_attention)
elif model_name == 'vgg13':
return vgg.VGG13(num_classes, use_attention=use_attention)
elif model_name == 'vgg16':
return vgg.VGG16(num_classes, use_attention=use_attention)
elif model_name == 'vgg19':
return vgg.VGG19(num_classes, use_attention=use_attention)
elif model_name == 'resnet18':
return resnet.ResNet18(num_classes, use_attention=use_attention)
elif model_name == 'resnet34':
return resnet.ResNet34(num_classes, use_attention=use_attention)
elif model_name == 'resnet50':
return resnet.ResNet50(num_classes, use_attention=use_attention)
elif model_name == 'resnet101':
return resnet.ResNet101(num_classes, use_attention=use_attention)
elif model_name == 'resnet152':
return resnet.ResNet152(num_classes, use_attention=use_attention)
elif model_name == 'densenet121':
return densenet.DenseNet121(num_classes, use_attention=use_attention)
elif model_name == 'densenet169':
return densenet.DenseNet169(num_classes, use_attention=use_attention)
elif model_name == 'densenet201':
return densenet.DenseNet201(num_classes, use_attention=use_attention)
elif model_name == 'densenet161':
return densenet.DenseNet161(num_classes, use_attention=use_attention)
elif model_name == 'mobilenetv3_small':
return mobilenetv3.mobilenetv3_small(num_classes)
elif model_name == 'mobilenetv3_large':
return mobilenetv3.mobilenetv3_large(num_classes)
elif model_name == 'efficientnet_b0':
return efficientnet.efficientnet_b0(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b1':
return efficientnet.efficientnet_b1(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b2':
return efficientnet.efficientnet_b2(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b3':
return efficientnet.efficientnet_b3(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b4':
return efficientnet.efficientnet_b4(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b5':
return efficientnet.efficientnet_b5(num_classes,
use_attention=use_attention)
elif model_name == 'efficientnet_b6':
return efficientnet.efficientnet_b6(num_classes,
use_attention=use_attention)
else:
raise ValueError("The model_name does not exist.")
def main(rank, world_size):
init_process(rank, world_size)
# make dataloader
train_loader, test_loader = dataset.get_loader(args, rank, world_size)
# set model
if args.model == 'res':
model = resnet.ResNet18()
model = model.cuda(rank)
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = DDP(model, device_ids=[rank])
cudnn.benchmark = True
# set criterion
criterion = nn.CrossEntropyLoss().cuda(rank)
# set optimizer (default:sgd)
optimizer = optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=0.0001,
nesterov=True)
# set scheduler
scheduler = MultiStepLR(optimizer, milestones=[60, 80], gamma=0.1)
# make logger
train_logger = utils.Logger(os.path.join(args.save_path, 'train.log'))
test_logger = utils.Logger(os.path.join(args.save_path, 'test.log'))
# Start Train
for epoch in range(1, args.epochs + 1):
# scheduler
if dist.get_rank() == 0:
scheduler.step()
# Train
train(train_loader, model, criterion, optimizer, epoch, train_logger)
validate(test_loader, model, criterion, epoch, test_logger, 'test')
# Save Model each epoch
if dist.get_rank() == 0:
if epoch == int(args.epochs):
torch.save(
model.state_dict(),
os.path.join(args.save_path,
'{0}_{1}.pth'.format('model', epoch)))
# Finish Train
# Draw Plot
if dist.get_rank() == 0:
plot_curves.draw_plot(args.save_path)
dist.destroy_process_group()
def __init__(self, model, num_workers, lr, job_name):
self.lock = threading.Lock()
self.logger = Logger(job_name=job_name, file_dir=f"./measurement/logs/{job_name}_ps.log").logger
self.cm_t1_start = np.zeros(num_workers)
self.future_model = torch.futures.Future()
self.batch_update_size = num_workers
self.curr_update_size = 0
self.stop_flag = False
if model == 'resnet20':
self.model = resnet3.resnet20()
elif model == 'resnet56':
self.model = resnet3.resnet56()
elif model == 'resnet18':
self.model = resnet.ResNet18()
elif model == 'resnet50':
self.model = resnet.ResNet50()
elif model == 'vgg13':
self.model = vgg.VGG13()
elif model == 'vgg16':
self.model = vgg.VGG16()
elif model == 'densenet121':
self.model = densenet.DenseNet121()
elif model == 'alexnet':
self.model = alexnet.AlexNet()
elif model == 'googlenet':
self.model = googlenet.GoogLeNet()
elif model == 'mobilenet':
self.model = mobilenetv2.MobileNetV2()
self.lr = lr
for p in self.model.parameters():
p.grad = torch.zeros_like(p)
self.optimizer = optim.SGD(self.model.parameters(), lr=lr, momentum=0.9)
self.info_socketm = znet.SocketMsger.tcp_connect(DORKER0_IP, INFO_PORT)
self.info_socketm.send("PS")
self.info_socketm.send(f"1.0\n/home/ubuntu/measurement/logs/{job_name}_info0.log\n{job_name}")
self.ps_launched_lock = threading.Lock()
self.ps_launched = False
"""Based on the model_version, determine model/optimizer and KD training mode
WideResNet and DenseNet were trained on multi-GPU; need to specify a dummy
nn.DataParallel module to correctly load the model parameters
"""
if "distill" in params.model_version:
student_model_load_start = time.time()
# train a 5-layer CNN or a 18-layer ResNet with knowledge distillation
if params.model_version == "cnn_distill":
model = net.Net(params).cuda() if params.cuda else net.Net(params)
optimizer = optim.Adam(model.parameters(), lr=params.learning_rate)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = net.metrics
elif params.model_version == 'resnet18_distill':
model = resnet.ResNet18().cuda() if params.cuda else resnet.ResNet18()
optimizer = optim.SGD(model.parameters(), lr=params.learning_rate,
momentum=0.9, weight_decay=5e-4)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = resnet.metrics
student_model_load_time = time.time() - student_model_load_start
logging.info("student_model_load_time: {}".format(student_model_load_time))
"""
Specify the pre-trained teacher models for knowledge distillation
Important note: wrn/densenet/resnext/preresnet were pre-trained models using multi-GPU,
therefore need to call "nn.DaraParallel" to correctly load the model weights
Trying to run on CPU will then trigger errors (too time-consuming anyway)!
"""
return sum(p.numel() for p in model.parameters() if p.requires_grad)
if __name__ == '__main__':
model_size = 0
args = parser.parse_args()
cnn_dir = 'experiments/cnn_distill'
json_path = os.path.join(cnn_dir, 'params.json')
assert os.path.isfile(
json_path), "No json configuration file found at {}".format(json_path)
params = utils.Params(json_path)
if args.model == "resnet18":
model = resnet.ResNet18()
model_checkpoint = 'experiments/base_resnet18/best.pth.tar'
elif args.model == "wrn":
model = wrn.wrn(depth=28,
num_classes=10,
widen_factor=10,
dropRate=0.3)
model_checkpoint = 'experiments/base_wrn/best.pth.tar'
elif args.model == "distill_resnext":
model = resnet.ResNet18()
model_checkpoint = 'experiments/resnet18_distill/resnext_teacher/best.pth.tar'
elif args.model == "distill_densenet":
model = resnet.ResNet18()
"""Based on the model_version, determine model/optimizer and KD training mode
WideResNet and DenseNet were trained on multi-GPU; need to specify a dummy
nn.DataParallel module to correctly load the model parameters
"""
if "distill" in params.model_version:
# train a 5-layer CNN or a 18-layer ResNet with knowledge distillation
if params.model_version == "cnn_distill":
model = net.Net(params).cuda() if params.cuda else net.Net(params)
optimizer = optim.Adam(model.parameters(), lr=params.learning_rate)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = net.metrics
elif params.model_version == 'resnet18_distill':
model = resnet.ResNet18().cuda(
) if params.cuda else resnet.ResNet18()
optimizer = optim.SGD(model.parameters(),
lr=params.learning_rate,
momentum=0.9,
weight_decay=5e-4)
# fetch loss function and metrics definition in model files
loss_fn_kd = net.loss_fn_kd
metrics = resnet.metrics
"""
Specify the pre-trained teacher models for knowledge distillation
Important note: wrn/densenet/resnext/preresnet were pre-trained models using multi-GPU,
therefore need to call "nn.DaraParallel" to correctly load the model weights
Trying to run on CPU will then trigger errors (too time-consuming anyway)!
"""
teacher_model = get_vgg()
def main():
# Load the parameters from json file
args = parser.parse_args()
json_path = os.path.join(args.model_dir, 'params.json')
assert os.path.isfile(
json_path), "No json configuration file found at {}".format(json_path)
params = utils.Params(json_path)
# Set the random seed for reproducible experiments
random.seed(230)
torch.manual_seed(230)
np.random.seed(230)
torch.cuda.manual_seed(230)
warnings.filterwarnings("ignore")
# Set the logger
utils.set_logger(os.path.join(args.model_dir, 'train.log'))
# Create the input data pipeline
logging.info("Loading the datasets...")
# fetch dataloaders, considering full-set vs. sub-set scenarios
if params.subset_percent < 1.0:
train_dl = data_loader.fetch_subset_dataloader('train', params)
else:
train_dl = data_loader.fetch_dataloader('train', params)
dev_dl = data_loader.fetch_dataloader('dev', params)
logging.info("- done.")
"""
Load student and teacher model
"""
if "distill" in params.model_version:
# Specify the student models
if params.model_version == "cnn_distill": # 5-layers Plain CNN
print("Student model: {}".format(params.model_version))
model = net.Net(params).cuda()
elif params.model_version == "shufflenet_v2_distill":
print("Student model: {}".format(params.model_version))
model = shufflenet.shufflenetv2(class_num=args.num_class).cuda()
elif params.model_version == "mobilenet_v2_distill":
print("Student model: {}".format(params.model_version))
model = mobilenet.mobilenetv2(class_num=args.num_class).cuda()
elif params.model_version == 'resnet18_distill':
print("Student model: {}".format(params.model_version))
model = resnet.ResNet18(num_classes=args.num_class).cuda()
elif params.model_version == 'resnet50_distill':
print("Student model: {}".format(params.model_version))
model = resnet.ResNet50(num_classes=args.num_class).cuda()
elif params.model_version == "alexnet_distill":
print("Student model: {}".format(params.model_version))
model = alexnet.alexnet(num_classes=args.num_class).cuda()
elif params.model_version == "vgg19_distill":
print("Student model: {}".format(params.model_version))
model = models.vgg19_bn(num_classes=args.num_class).cuda()
elif params.model_version == "googlenet_distill":
print("Student model: {}".format(params.model_version))
model = googlenet.GoogleNet(num_class=args.num_class).cuda()
elif params.model_version == "resnext29_distill":
print("Student model: {}".format(params.model_version))
model = resnext.CifarResNeXt(cardinality=8,
depth=29,
num_classes=args.num_class).cuda()
elif params.model_version == "densenet121_distill":
print("Student model: {}".format(params.model_version))
model = densenet.densenet121(num_class=args.num_class).cuda()
# optimizer
if params.model_version == "cnn_distill":
optimizer = optim.Adam(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128))
else:
optimizer = optim.SGD(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128),
momentum=0.9,
weight_decay=5e-4)
iter_per_epoch = len(train_dl)
warmup_scheduler = utils.WarmUpLR(
optimizer, iter_per_epoch *
args.warm) # warmup the learning rate in the first epoch
# specify loss function
if args.self_training:
print(
'>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>self training>>>>>>>>>>>>>>>>>>>>>>>>>>>>>'
)
loss_fn_kd = loss_kd_self
else:
loss_fn_kd = loss_kd
"""
Specify the pre-trained teacher models for knowledge distillation
Checkpoints can be obtained by regular training or downloading our pretrained models
For model which is pretrained in multi-GPU, use "nn.DaraParallel" to correctly load the model weights.
"""
if params.teacher == "resnet18":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet18(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet18/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet18/0.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "alexnet":
print("Teacher model: {}".format(params.teacher))
teacher_model = alexnet.alexnet(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_alexnet/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "googlenet":
print("Teacher model: {}".format(params.teacher))
teacher_model = googlenet.GoogleNet(num_class=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_googlenet/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "vgg19":
print("Teacher model: {}".format(params.teacher))
teacher_model = models.vgg19_bn(num_classes=args.num_class)
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_vgg19/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "resnet50":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet50(num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet50/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet50/50.pth.tar'
elif params.teacher == "resnet101":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnet.ResNet101(num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnet101/best.pth.tar'
teacher_model = teacher_model.cuda()
elif params.teacher == "densenet121":
print("Teacher model: {}".format(params.teacher))
teacher_model = densenet.densenet121(
num_class=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_densenet121/best.pth.tar'
# teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "resnext29":
print("Teacher model: {}".format(params.teacher))
teacher_model = resnext.CifarResNeXt(
cardinality=8, depth=29, num_classes=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnext29/best.pth.tar'
if args.pt_teacher: # poorly-trained teacher for Defective KD experiments
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_resnext29/50.pth.tar'
teacher_model = nn.DataParallel(teacher_model).cuda()
elif params.teacher == "mobilenet_v2":
print("Teacher model: {}".format(params.teacher))
teacher_model = mobilenet.mobilenetv2(
class_num=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_mobilenet_v2/best.pth.tar'
elif params.teacher == "shufflenet_v2":
print("Teacher model: {}".format(params.teacher))
teacher_model = shufflenet.shufflenetv2(
class_num=args.num_class).cuda()
teacher_checkpoint = 'experiments/pretrained_teacher_models/base_shufflenet_v2/best.pth.tar'
utils.load_checkpoint(teacher_checkpoint, teacher_model)
# Train the model with KD
logging.info("Starting training for {} epoch(s)".format(
params.num_epochs))
train_and_evaluate_kd(model, teacher_model, train_dl, dev_dl,
optimizer, loss_fn_kd, warmup_scheduler, params,
args, args.restore_file)
# non-KD mode: regular training to obtain a baseline model
else:
print("Train base model")
if params.model_version == "cnn":
model = net.Net(params).cuda()
elif params.model_version == "mobilenet_v2":
print("model: {}".format(params.model_version))
model = mobilenet.mobilenetv2(class_num=args.num_class).cuda()
elif params.model_version == "shufflenet_v2":
print("model: {}".format(params.model_version))
model = shufflenet.shufflenetv2(class_num=args.num_class).cuda()
elif params.model_version == "alexnet":
print("model: {}".format(params.model_version))
model = alexnet.alexnet(num_classes=args.num_class).cuda()
elif params.model_version == "vgg19":
print("model: {}".format(params.model_version))
model = models.vgg19_bn(num_classes=args.num_class).cuda()
elif params.model_version == "googlenet":
print("model: {}".format(params.model_version))
model = googlenet.GoogleNet(num_class=args.num_class).cuda()
elif params.model_version == "densenet121":
print("model: {}".format(params.model_version))
model = densenet.densenet121(num_class=args.num_class).cuda()
elif params.model_version == "resnet18":
model = resnet.ResNet18(num_classes=args.num_class).cuda()
elif params.model_version == "resnet50":
model = resnet.ResNet50(num_classes=args.num_class).cuda()
elif params.model_version == "resnet101":
model = resnet.ResNet101(num_classes=args.num_class).cuda()
elif params.model_version == "resnet152":
model = resnet.ResNet152(num_classes=args.num_class).cuda()
elif params.model_version == "resnext29":
model = resnext.CifarResNeXt(cardinality=8,
depth=29,
num_classes=args.num_class).cuda()
# model = nn.DataParallel(model).cuda()
if args.regularization:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Loss of Regularization>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = loss_kd_regularization
elif args.label_smoothing:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Label Smoothing>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = loss_label_smoothing
else:
print(
">>>>>>>>>>>>>>>>>>>>>>>>Normal Training>>>>>>>>>>>>>>>>>>>>>>>>"
)
loss_fn = nn.CrossEntropyLoss()
if args.double_training: # double training, compare to self-KD
print(
">>>>>>>>>>>>>>>>>>>>>>>>Double Training>>>>>>>>>>>>>>>>>>>>>>>>"
)
checkpoint = 'experiments/pretrained_teacher_models/base_' + str(
params.model_version) + '/best.pth.tar'
utils.load_checkpoint(checkpoint, model)
if params.model_version == "cnn":
optimizer = optim.Adam(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128))
else:
optimizer = optim.SGD(model.parameters(),
lr=params.learning_rate *
(params.batch_size / 128),
momentum=0.9,
weight_decay=5e-4)
iter_per_epoch = len(train_dl)
warmup_scheduler = utils.WarmUpLR(optimizer,
iter_per_epoch * args.warm)
# Train the model
logging.info("Starting training for {} epoch(s)".format(
params.num_epochs))
train_and_evaluate(model, train_dl, dev_dl, optimizer, loss_fn, params,
args.model_dir, warmup_scheduler, args,
args.restore_file)
