def build_transforms(self, config, **kwargs):
if config.transforms.name == 'Compose':
transfms = []
for t in config.transforms.params:
transfms.append(Registry.build_from_config('transform', t))
return transforms.Compose(transfms)
elif config.transforms.name == '':
return None
else:
return Registry.build_from_config('transform', config.transforms),
def build_dataloaders(self, config, **kwargs):
dataloaders = []
# splitter = self.build_logger_splitter(config)
# for fold in range(splitter.n_splits):
# train, val = splitter.get_fold(fold)
# # dataset_config = {'name': config.dataset.name, 'params': config.dataset.params}
# # dataset_config['params'].update(split_config)
# transform = self.build_transforms(config)
# default_args = {}
# if transform is not None:
# default_args['transform'] = transform
# default_args['indices'] = train
# dataset = Registry.build_from_config('dataset', config.dataset, default_args=default_args)
# batch_size = config.train.batch_size
# is_train = True
# dataloader = DataLoader(dataset,
# shuffle=is_train,
# batch_size=batch_size,
# drop_last=is_train,
# num_workers=config.transforms.num_preprocessor,
# pin_memory=True)
# dataloaders.append({'fold': fold, 'mode': is_train,'dataloader': dataloader})
# # validation dataloder
# default_args['indices'] = val
# dataset = Registry.build_from_config('dataset', config.dataset, default_args=default_args)
# batch_size = config.evaluation.batch_size
# is_train = False
# dataloader = DataLoader(dataset,
# shuffle=is_train,
# batch_size=batch_size,
# drop_last=is_train,
# num_workers=config.transforms.num_preprocessor,
# pin_memory=True)
# dataloaders.append({'fold': fold, 'mode': is_train,'dataloader': dataloader})
for split_config in config.dataset.splits:
dataset_config = {'name': config.dataset.name, 'params': config.dataset.params}
dataset_config['params'].update(split_config)
transform = self.build_transforms(config)
default_args = {}
if transform is not None:
default_args['transform'] = transform
dataset = Registry.build_from_config('dataset', config.dataset, default_args=default_args)
is_train = dataset_config['params'].train
if is_train:
batch_size = config.train.batch_size
else:
batch_size = config.evaluation.batch_size
dataloader = DataLoader(dataset,
shuffle=is_train,
batch_size=batch_size,
drop_last=is_train,
num_workers=config.transforms.num_preprocessor,
pin_memory=True)
dataloaders.append({'mode': is_train, 'split': dataset_config['params'].split, 'dataloader': dataloader})
return dataloaders
def build_logger_fn(self, config, **kwargs):
return Registry.build_from_config('hooks', config.logger_hook, kwargs)
def build_splitter(self, config, **kwargs):
return Registry.build_from_config('splitter', config.splitter, kwargs)
def build_post_forward_hook(self, config, **kwargs):
return Registry.build_from_config('hooks', config.post_forward_hook, kwargs)
def build_metric_fn(self, config, **kwargs):
return Registry.build_from_config('hooks', config.metric_hook, kwargs)
def build_scheduler(self, config, **kwargs):
return Registry.build_from_config('scheduler', config.scheduler, kwargs)
def build_optimizer(self, config, **kwargs):
return Registry.build_from_config('optimizer', config.optimizer, kwargs)
def build_loss_fn(self, config, **kwargs):
return Registry.build_from_config('loss', config.loss, kwargs)
def build_model(self, config, **kwargs):
return Registry.build_from_config('model', config.model, kwargs)
def regist_defaults(self):
Registry.add('loss', nn.NLLLoss)
Registry.add('optimizer', optim.Adadelta)
Registry.add('optimizer', optim.AdamW)
Registry.add('scheduler', optim.lr_scheduler.StepLR)
Registry.add('scheduler', optim.lr_scheduler.MultiStepLR)
Registry.add('dataset', datasets.MNIST)
Registry.add('transform', transforms.ToTensor)
Registry.add('transform', transforms.Normalize)
Registry.add('hooks', DefaultPostForwardHook)
Registry.add('hooks', DefaultMetric)
Registry.add('hooks', DefaultLogger)
Registry.add('model', Mnist)