def get_model_size(self):
return count_parameters_in_MB(self.graph)
def get_model_size(self):
return count_parameters_in_MB(self.history)
def __init__(self, graph, parser, *args, **kwargs):
self.graph = graph
self.parser = parser
try:
self.config = kwargs.get('config', graph.config)
except:
raise ('No configuration specified in graph or kwargs')
np.random.seed(self.config.seed)
random.seed(self.config.seed)
if torch.cuda.is_available():
torch.manual_seed(self.config.seed)
torch.cuda.set_device(self.config.gpu)
cudnn.benchmark = False
cudnn.enabled = True
cudnn.deterministic = True
torch.cuda.manual_seed_all(self.config.seed)
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
# dataloaders
train_queue, valid_queue, test_queue, train_transform, valid_transform = parser.get_train_val_loaders(
)
self.train_queue = train_queue
self.valid_queue = valid_queue
self.test_queue = test_queue
self.train_transform = train_transform
self.valid_transform = valid_transform
criterion = eval('nn.' + self.config.criterion)()
self.criterion = criterion.cuda()
self.model = self.graph.to(self.device)
n_parameters = utils.count_parameters_in_MB(self.model)
logging.info("param size = %fMB", n_parameters)
optimizer = torch.optim.SGD(self.model.parameters(),
self.config.learning_rate,
momentum=self.config.momentum,
weight_decay=self.config.weight_decay)
self.optimizer = optimizer
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
float(self.config.epochs),
eta_min=self.config.learning_rate_min)
logging.info('Args: {}'.format(self.config))
self.run_kwargs = {}
self.errors_dict = utils.AttrDict({
'train_acc': [],
'train_loss': [],
'valid_acc': [],
'valid_loss': [],
'test_acc': [],
'test_loss': [],
'runtime': [],
'params': n_parameters
})
