def run(config):
train_dir = config.train.dir
student_model = get_model(config, model_type).to(device)
criterion = get_loss(config)
trainable_params = filter(lambda p: p.requires_grad,
student_model.parameters())
optimizer = get_optimizer(config, trainable_params)
checkpoint = utils.checkpoint.get_initial_checkpoint(config,
model_type=model_type)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
student_model, optimizer, checkpoint, model_type=model_type)
else:
last_epoch, step = -1, -1
print('student model from checkpoint: {} last epoch:{}'.format(
checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train['student' + '_dir'])
visualizer = get_visualizer(config)
train(config, student_model, dataloaders, criterion, optimizer, scheduler,
writer, visualizer, last_epoch + 1)
def run(config):
train_dir = config.train.dir
model = get_model(config, model_type).to(device)
print('The nubmer of parameters : %d' % count_parameters(model))
criterion = get_loss(config)
optimizer = get_optimizer(config, model)
checkpoint = utils.checkpoint.get_initial_checkpoint(config,
model_type=model_type)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
model, optimizer, checkpoint, model_type=model_type)
else:
last_epoch, step = -1, -1
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train[model_type + '_dir'])
visualizer = get_visualizer(config)
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
visualizer, last_epoch + 1)
def run(config):
train_dir = config.train.dir
model = get_model(config).cuda()
criterion = get_loss(config)
optimizer = get_optimizer(config, model.parameters())
checkpoint = utils.checkpoint.get_initial_checkpoint(config)
if checkpoint is not None:
last_epoch, step = utils.checkpoint.load_checkpoint(
model, optimizer, checkpoint)
else:
last_epoch, step = -1, -1
print('from checkpoint: {} last epoch:{}'.format(checkpoint, last_epoch))
scheduler = get_scheduler(config, optimizer, last_epoch)
# dataloaders = {split:get_dataloader(config, split, get_transform(config, split))
# for split in ['train', 'val']}
print(config.data)
dataloaders = {
'train': get_train_dataloader(config, get_transform(config)),
'val': get_valid_dataloaders(config)[0]
}
writer = SummaryWriter(train_dir)
train(config, model, dataloaders, criterion, optimizer, scheduler, writer,
last_epoch + 1)
def run(config):
teacher_model = get_model(config, 'teacher').to(device)
criterion = get_loss(config)
# for teacher
trainable_params = filter(lambda p: p.requires_grad,
teacher_model.parameters())
optimizer_t = get_optimizer(config, teacher_model.parameters())
checkpoint_t = utils.checkpoint.get_initial_checkpoint(
config, model_type='teacher')
if checkpoint_t is not None:
last_epoch_t, step_t = utils.checkpoint.load_checkpoint(
teacher_model, optimizer_t, checkpoint_t, model_type='teacher')
else:
last_epoch_t, step_t = -1, -1
print('teacher model from checkpoint: {} last epoch:{}'.format(
checkpoint_t, last_epoch_t))
scheduler_t = get_scheduler(config, optimizer_t, last_epoch_t)
print(config.data)
dataloaders = {
'train': get_train_dataloader(config),
'val': get_valid_dataloader(config),
'test': get_test_dataloader(config)
}
writer = SummaryWriter(config.train['teacher' + '_dir'])
visualizer = get_visualizer(config)
train(config, teacher_model, dataloaders, criterion, optimizer_t,
scheduler_t, writer, visualizer, last_epoch_t + 1)
def run(config):
teacher_model = get_model(config, 'teacher').to(device)
student_model = get_model(config, 'student').to(device)
print('The nubmer of parameters : %d'%count_parameters(student_model))
criterion = get_loss(config)
# for teacher
optimizer_t = None
checkpoint_t = utils.checkpoint.get_initial_checkpoint(config,
model_type='teacher')
if checkpoint_t is not None:
last_epoch_t, step_t = utils.checkpoint.load_checkpoint(teacher_model,
optimizer_t, checkpoint_t, model_type='teacher')
else:
last_epoch_t, step_t = -1, -1
print('teacher model from checkpoint: {} last epoch:{}'.format(
checkpoint_t, last_epoch_t))
# for student
optimizer_s = get_optimizer(config, student_model)
checkpoint_s = utils.checkpoint.get_initial_checkpoint(config,
model_type='student')
if checkpoint_s is not None:
last_epoch_s, step_s = utils.checkpoint.load_checkpoint(student_model,
optimizer_s, checkpoint_s, model_type='student')
else:
last_epoch_s, step_s = -1, -1
print('student model from checkpoint: {} last epoch:{}'.format(
checkpoint_s, last_epoch_s))
scheduler_s = get_scheduler(config, optimizer_s, last_epoch_s)
print(config.data)
dataloaders = {'train':get_train_dataloader(config, get_transform(config)),
'val':get_valid_dataloader(config)}
#'test':get_test_dataloader(config)}
writer = SummaryWriter(config.train['student' + '_dir'])
visualizer = get_visualizer(config)
result = train(config, student_model, teacher_model, dataloaders,
criterion, optimizer_s, scheduler_s, writer,
visualizer, last_epoch_s+1)
print('best psnr : %.3f, best epoch: %d'%(result['best_psnr'], result['best_epoch']))