def train(model, train_sampler, train_data_loader, val_data_loader, device, distributed, config, args, ckpt_file_path):
train_config = config['train']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(model, optim_config['type'], optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'], scheduler_config['params'])
best_val_map = 0.0
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path, optimizer=optimizer, lr_scheduler=lr_scheduler)
num_epochs = train_config['num_epochs']
log_freq = train_config['log_freq']
start_time = time.time()
for epoch in range(num_epochs):
if distributed:
train_sampler.set_epoch(epoch)
train_model(model, optimizer, train_data_loader, device, epoch, log_freq)
lr_scheduler.step()
# evaluate after every epoch
coco_evaluator = main_util.evaluate(model, val_data_loader, device=device)
# Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]
val_map = coco_evaluator.coco_eval['bbox'].stats[0]
if val_map > best_val_map:
print('Updating ckpt (Best BBox mAP: {:.4f} -> {:.4f})'.format(best_val_map, val_map))
best_val_map = val_map
save_ckpt(model, optimizer, lr_scheduler, best_val_map, config, args, ckpt_file_path)
lr_scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def train(train_loader, valid_loader, input_shape, config, device, distributed,
device_ids):
ae_without_ddp, ae_type = ae_util.get_autoencoder(config, device)
head_model = ae_util.get_head_model(config, input_shape, device)
module_util.freeze_module_params(head_model)
ckpt_file_path = config['autoencoder']['ckpt']
start_epoch, best_valid_acc = resume_from_ckpt(ckpt_file_path,
ae_without_ddp)
if best_valid_acc is None:
best_valid_acc = 0.0
train_config = config['train']
criterion_config = train_config['criterion']
criterion = func_util.get_loss(criterion_config['type'],
criterion_config['params'])
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(ae_without_ddp, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
interval = train_config['interval']
if interval <= 0:
num_batches = len(train_loader)
interval = num_batches // 20 if num_batches >= 20 else 1
autoencoder = ae_without_ddp
if distributed:
autoencoder = DistributedDataParallel(ae_without_ddp,
device_ids=device_ids)
head_model = DataParallel(head_model, device_ids=device_ids)
elif device.type == 'cuda':
autoencoder = DataParallel(ae_without_ddp)
head_model = DataParallel(head_model)
end_epoch = start_epoch + train_config['epoch']
start_time = time.time()
for epoch in range(start_epoch, end_epoch):
if distributed:
train_loader.sampler.set_epoch(epoch)
train_epoch(autoencoder, head_model, train_loader, optimizer,
criterion, epoch, device, interval)
valid_acc = validate(ae_without_ddp, valid_loader, config, device,
distributed, device_ids)
if valid_acc > best_valid_acc and main_util.is_main_process():
print(
'Updating ckpt (Best top1 accuracy: {:.4f} -> {:.4f})'.format(
best_valid_acc, valid_acc))
best_valid_acc = valid_acc
save_ckpt(ae_without_ddp, epoch, best_valid_acc, ckpt_file_path,
ae_type)
scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
del head_model
def distill(teacher_model, student_model, train_sampler, train_data_loader,
val_data_loader, device, distributed, distill_backbone_only,
config, args):
train_config = config['train']
distillation_box = DistillationBox(teacher_model, student_model,
train_config['criterion'])
ckpt_file_path = config['student_model']['ckpt']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
use_bottleneck_transformer = args.transform_bottleneck
best_val_map = 0.0
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path,
optimizer=optimizer,
lr_scheduler=lr_scheduler)
num_epochs = train_config['num_epochs']
log_freq = train_config['log_freq']
teacher_model_without_dp = teacher_model.module if isinstance(
teacher_model, DataParallel) else teacher_model
student_model_without_ddp =\
student_model.module if isinstance(student_model, DistributedDataParallel) else student_model
start_time = time.time()
for epoch in range(num_epochs):
if distributed:
train_sampler.set_epoch(epoch)
teacher_model.eval()
student_model.train()
teacher_model_without_dp.distill_backbone_only = distill_backbone_only
student_model_without_ddp.distill_backbone_only = distill_backbone_only
student_model_without_ddp.backbone.body.layer1.use_bottleneck_transformer = False
distill_model(distillation_box, train_data_loader, optimizer, log_freq,
device, epoch)
student_model_without_ddp.distill_backbone_only = False
student_model_without_ddp.backbone.body.layer1.use_bottleneck_transformer = use_bottleneck_transformer
coco_evaluator = main_util.evaluate(student_model,
val_data_loader,
device=device)
# Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ]
val_map = coco_evaluator.coco_eval['bbox'].stats[0]
if val_map > best_val_map and misc_util.is_main_process():
print('Updating ckpt (Best BBox mAP: {:.4f} -> {:.4f})'.format(
best_val_map, val_map))
best_val_map = val_map
save_ckpt(student_model_without_ddp, optimizer, lr_scheduler,
best_val_map, config, args, ckpt_file_path)
lr_scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def distill(teacher_model, student_model, train_data_loader, val_data_loader,
device, distributed, start_epoch, config, args):
print('Start knowledge distillation')
train_config = config['train']
distillation_box = DistillationBox(teacher_model, student_model,
train_config['criterion'])
ckpt_file_path = config['mimic_model']['ckpt']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
best_val_top1_accuracy = 0.0
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path,
optimizer=optimizer,
lr_scheduler=lr_scheduler)
interval = train_config['interval']
if interval <= 0:
num_batches = len(train_data_loader)
interval = num_batches // 20 if num_batches >= 20 else 1
student_model_without_ddp = \
student_model.module if isinstance(student_model, DistributedDataParallel) else student_model
start_time = time.time()
for epoch in range(start_epoch, train_config['epoch']):
if distributed:
train_data_loader.sampler.set_epoch(epoch)
teacher_model.eval()
student_model.train()
distill_one_epoch(distillation_box, train_data_loader, optimizer,
device, epoch, interval, args.apex)
val_top1_accuracy =\
evaluate(student_model, val_data_loader, device=device, interval=interval, split_name='Validation')
if val_top1_accuracy > best_val_top1_accuracy and main_util.is_main_process(
):
print(
'Updating ckpt (Best top1 accuracy: {:.4f} -> {:.4f})'.format(
best_val_top1_accuracy, val_top1_accuracy))
best_val_top1_accuracy = val_top1_accuracy
save_ckpt(student_model_without_ddp, optimizer, lr_scheduler,
best_val_top1_accuracy, config, args, ckpt_file_path)
lr_scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def distill(teacher_model, student_model, train_sampler, train_data_loader, val_data_loader,
device, distributed, distill_backbone_only, config, args):
train_config = config['train']
student_config = config['student_model']
distillation_box = DistillationBox(teacher_model, student_model,
train_config['criterion'], student_config)
ckpt_file_path = config['student_model']['ckpt']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'], optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'], scheduler_config['params'])
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path, optimizer=optimizer, lr_scheduler=lr_scheduler)
save_ckpt(student_model, optimizer, lr_scheduler, best_val_map, config, args, ckpt_file_path)
def train(model, train_loader, valid_loader, best_valid_acc, criterion, device,
distributed, device_ids, train_config, num_epochs, start_epoch,
init_lr, ckpt_file_path, model_type):
model_without_ddp = model
if distributed:
model = DistributedDataParallel(model_without_ddp,
device_ids=device_ids)
elif device.type == 'cuda':
model = DataParallel(model_without_ddp)
optim_config = train_config['optimizer']
if init_lr is not None:
optim_config['params']['lr'] = init_lr
optimizer = func_util.get_optimizer(model, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
interval = train_config['interval']
if interval <= 0:
num_batches = len(train_loader)
interval = num_batches // 20 if num_batches >= 20 else 1
end_epoch = start_epoch + train_config[
'epoch'] if num_epochs is None else start_epoch + num_epochs
start_time = time.time()
for epoch in range(start_epoch, end_epoch):
if distributed:
train_loader.sampler.set_epoch(epoch)
train_epoch(model, train_loader, optimizer, criterion, epoch, device,
interval)
valid_acc = validate(model, valid_loader, device)
if valid_acc > best_valid_acc and main_util.is_main_process():
print(
'Updating ckpt (Best top1 accuracy: {:.4f} -> {:.4f})'.format(
best_valid_acc, valid_acc))
best_valid_acc = valid_acc
save_ckpt(model_without_ddp, best_valid_acc, epoch, ckpt_file_path,
model_type)
scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def train(model, ext_classifier, train_sampler, train_data_loader,
val_data_loader, device, distributed, config, args, ckpt_file_path):
train_config = config['train']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(ext_classifier, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
best_val_roc_auc = 0.0
if file_util.check_if_exists(ckpt_file_path):
best_val_roc_auc, _, _ =\
load_ckpt(ckpt_file_path, model=ext_classifier, optimizer=optimizer, lr_scheduler=lr_scheduler)
num_epochs = train_config['num_epochs']
log_freq = train_config['log_freq']
start_time = time.time()
for epoch in range(num_epochs):
if distributed:
train_sampler.set_epoch(epoch)
train_model(model, optimizer, train_data_loader, device, epoch,
log_freq)
lr_scheduler.step()
# evaluate after every epoch
val_roc_auc = evaluate(model,
val_data_loader,
device,
min_recall=args.min_recall,
split_name='Validation')
if val_roc_auc > best_val_roc_auc:
print('Updating ckpt (Best ROC-AUC: {:.4f} -> {:.4f})'.format(
best_val_roc_auc, val_roc_auc))
best_val_roc_auc = val_roc_auc
save_ckpt(ext_classifier, optimizer, lr_scheduler,
best_val_roc_auc, config, args, ckpt_file_path)
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main(args):
config = yaml_util.load_yaml_file(args.config)
if args.json is not None:
main_util.overwrite_config(config, args.json)
distributed, device_ids = main_util.init_distributed_mode(args.world_size, args.dist_url)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
teacher_model = get_model(config['teacher_model'], device)
module_util.freeze_module_params(teacher_model)
student_model_config = config['student_model']
student_model = get_model(student_model_config, device)
freeze_modules(student_model, student_model_config)
ckpt_file_path = config['student_model']['ckpt']
train_config = config['train']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'], optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'], scheduler_config['params'])
if file_util.check_if_exists(ckpt_file_path):
best_val_map, _, _ = load_ckpt(ckpt_file_path, optimizer=optimizer, lr_scheduler=lr_scheduler)
save_ckpt(student_model, optimizer, lr_scheduler, best_val_map, config, args, ckpt_file_path)
def distill(train_loader, valid_loader, input_shape, aux_weight, config,
device, distributed, device_ids):
teacher_model_config = config['teacher_model']
teacher_model, teacher_model_type = mimic_util.get_teacher_model(
teacher_model_config, input_shape, device)
module_util.freeze_module_params(teacher_model)
student_model_config = config['student_model']
student_model = mimic_util.get_student_model(teacher_model_type,
student_model_config,
config['dataset']['name'])
student_model = student_model.to(device)
start_epoch, best_valid_acc = mimic_util.resume_from_ckpt(
student_model_config['ckpt'], student_model, is_student=True)
if best_valid_acc is None:
best_valid_acc = 0.0
train_config = config['train']
criterion_config = train_config['criterion']
criterion = func_util.get_loss(criterion_config['type'],
criterion_config['params'])
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
interval = train_config['interval']
if interval <= 0:
num_batches = len(train_loader)
interval = num_batches // 20 if num_batches >= 20 else 1
student_model_without_ddp = student_model
if distributed:
teacher_model = DataParallel(teacher_model, device_ids=device_ids)
student_model = DistributedDataParallel(student_model,
device_ids=device_ids)
student_model_without_ddp = student_model.module
ckpt_file_path = student_model_config['ckpt']
end_epoch = start_epoch + train_config['epoch']
start_time = time.time()
for epoch in range(start_epoch, end_epoch):
if distributed:
train_loader.sampler.set_epoch(epoch)
distill_one_epoch(student_model, teacher_model, train_loader,
optimizer, criterion, epoch, device, interval,
aux_weight)
valid_acc = validate(student_model, valid_loader, config, device,
distributed, device_ids)
if valid_acc > best_valid_acc and main_util.is_main_process():
print(
'Updating ckpt (Best top1 accuracy: {:.4f} -> {:.4f})'.format(
best_valid_acc, valid_acc))
best_valid_acc = valid_acc
save_ckpt(student_model_without_ddp, epoch, best_valid_acc,
ckpt_file_path, teacher_model_type)
scheduler.step()
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
del teacher_model
del student_model
def distill(teacher_model, student_model, train_sampler, train_data_loader,
val_data_loader, device, distributed, distill_backbone_only,
config, args):
train_config = config['train']
student_config = config['student_model']
distillation_box = DistillationBox(teacher_model, student_model,
train_config['criterion'],
student_config)
ckpt_file_path = config['student_model']['ckpt']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
scheduler_config = train_config['scheduler']
lr_scheduler = func_util.get_scheduler(optimizer, scheduler_config['type'],
scheduler_config['params'])
use_bottleneck_transformer = args.transform_bottleneck
best_val_map = 0.0
if file_util.check_if_exists(ckpt_file_path):
if args.ignore_optimizer:
best_val_map, _, _ = load_ckpt(ckpt_file_path,
optimizer=None,
lr_scheduler=None)
else:
best_val_map, _, _ = load_ckpt(ckpt_file_path,
optimizer=optimizer,
lr_scheduler=lr_scheduler)
num_epochs = train_config['num_epochs']
log_freq = train_config['log_freq']
teacher_model_without_dp = teacher_model.module if isinstance(
teacher_model, DataParallel) else teacher_model
student_model_without_ddp = \
student_model.module if isinstance(student_model, DistributedDataParallel) else student_model
start_time = time.time()
post_bn = False
if 'post_batch_norm' in config['train']:
post_bn = config['train']['post_batch_norm']
for epoch in range(lr_scheduler.last_epoch, num_epochs):
if distributed:
train_sampler.set_epoch(epoch)
teacher_model.eval()
student_model.train()
teacher_model_without_dp.distill_backbone_only = distill_backbone_only
student_model_without_ddp.distill_backbone_only = distill_backbone_only
set_bottleneck_transformer(student_model_without_ddp, False)
distill_model(distillation_box, train_data_loader, optimizer, log_freq,
device, epoch)
student_model_without_ddp.distill_backbone_only = False
set_bottleneck_transformer(student_model_without_ddp,
use_bottleneck_transformer)
val_map = 0
width_list = [1.0]
if 'slimmable' in student_config['backbone']['params']:
width_list = [0.25, 0.5, 0.75, 1.0]
width_list = [
w for w in width_list
if w in student_config['backbone']['params']['width_mult_list']
]
for width in width_list:
set_width(student_model, width)
if post_bn:
ComputeBN(student_model, train_data_loader)
print('\n[Student model@width={}]'.format(width))
coco_evaluator = main_util.evaluate(student_model,
val_data_loader,
device=device)
val_map += coco_evaluator.coco_eval['bbox'].stats[0]
val_map = val_map / len(width_list)
print('BBox mAP: {:.4f})'.format(val_map))
if val_map > best_val_map and misc_util.is_main_process():
print('Updating ckpt (Best BBox mAP: {:.4f} -> {:.4f})'.format(
best_val_map, val_map))
best_val_map = val_map
save_ckpt(student_model_without_ddp, optimizer, lr_scheduler,
best_val_map, config, args, ckpt_file_path)
lr_scheduler.step()
if distributed:
dist.barrier()
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
def main(args):
if args.apex:
if sys.version_info < (3, 0):
raise RuntimeError(
'Apex currently only supports Python 3. Aborting.')
if amp is None:
raise RuntimeError(
'Failed to import apex. Please install apex from https://www.github.com/nvidia/apex '
'to enable mixed-precision training.')
distributed, device_ids = main_util.init_distributed_mode(
args.world_size, args.dist_url)
print(args)
if torch.cuda.is_available():
torch.backends.cudnn.benchmark = True
config = yaml_util.load_yaml_file(args.config)
device = torch.device(args.device if torch.cuda.is_available() else 'cpu')
dataset_config = config['dataset']
input_shape = config['input_shape']
train_config = config['train']
test_config = config['test']
train_data_loader, val_data_loader, test_data_loader =\
dataset_util.get_data_loaders(dataset_config, batch_size=train_config['batch_size'],
rough_size=train_config['rough_size'], reshape_size=input_shape[1:3],
jpeg_quality=-1, test_batch_size=test_config['batch_size'],
distributed=distributed)
teacher_model_config = config['teacher_model']
teacher_model, teacher_model_type = mimic_util.get_org_model(
teacher_model_config, device)
module_util.freeze_module_params(teacher_model)
student_model = mimic_util.get_mimic_model_easily(config, device)
student_model_config = config['mimic_model']
optim_config = train_config['optimizer']
optimizer = func_util.get_optimizer(student_model, optim_config['type'],
optim_config['params'])
use_apex = args.apex
if use_apex:
student_model, optimizer = amp.initialize(
student_model, optimizer, opt_level=args.apex_opt_level)
if distributed:
teacher_model = DataParallel(teacher_model, device_ids=device_ids)
student_model = DistributedDataParallel(student_model,
device_ids=device_ids)
start_epoch = args.start_epoch
if not args.test_only:
distill(teacher_model, student_model, train_data_loader,
val_data_loader, device, distributed, start_epoch, config,
args)
student_model_without_ddp =\
student_model.module if isinstance(student_model, DistributedDataParallel) else student_model
load_ckpt(student_model_config['ckpt'],
model=student_model_without_ddp,
strict=True)
if not args.student_only:
evaluate(teacher_model,
test_data_loader,
device,
title='[Teacher: {}]'.format(teacher_model_type))
evaluate(student_model,
test_data_loader,
device,
title='[Student: {}]'.format(student_model_config['type']))
