def train_process(self):
"""Train process of parameter sharing."""
self.train_loader = Dataset(mode='train').dataloader
self.valid_loader = Dataset(mode='val').dataloader
self.model = self.model.to(self.device)
self.search_alg = SearchAlgorithm(SearchSpace())
self.set_algorithm_model(self.model)
self.optimizer = self._init_optimizer()
self.lr_scheduler = self._init_lr_scheduler()
self.loss_fn = self._init_loss()
np.random.seed(self.cfg.seed)
torch.manual_seed(self.cfg.seed)
for i in range(self.cfg.epochs):
self._train(self.model)
train_top1, train_top5 = self._valid(self.model, self.train_loader)
valid_top1, valid_top5 = self._valid(self.model, self.valid_loader)
self.lr_scheduler.step()
child_desc_temp = self.search_alg.codec.calc_genotype(
self.model.arch_weights)
logging.info(F.softmax(self.model.alphas_normal, dim=-1))
logging.info(F.softmax(self.model.alphas_reduce, dim=-1))
logging.info('normal = %s', child_desc_temp[0])
logging.info('reduce = %s', child_desc_temp[1])
logging.info('Epoch {}: train top1: {}, train top5: {}'.format(
i, train_top1, train_top5))
logging.info('Epoch {}: valid top1: {}, valid top5: {}'.format(
i, valid_top1, valid_top5))
child_desc = self.search_alg.codec.decode(self.model.arch_weights)
self._save_descript(child_desc)
def __init__(self):
self.step_name = General.step_name
self.search_space = SearchSpace()
self.search_alg = SearchAlgorithm(self.search_space.search_space)
self.report = Report()
self.record = ReportRecord()
self.record.step_name = self.step_name
if hasattr(self.search_alg.config, 'objective_keys'):
self.record.objective_keys = self.search_alg.config.objective_keys
def before_train(self, logs=None):
"""Be called before the training process."""
self.config = self.trainer.config
self.unrolled = self.trainer.config.unrolled
self.device = self.trainer.config.device
self.model = self.trainer.model
self.optimizer = self.trainer.optimizer
self.lr_scheduler = self.trainer.lr_scheduler
self.loss = self.trainer.loss
self.search_alg = SearchAlgorithm(SearchSpace().search_space)
self._set_algorithm_model(self.model)
self.trainer.train_loader = self.trainer._init_dataloader(mode='train')
self.trainer.valid_loader = self.trainer._init_dataloader(mode='val')
def before_train(self, epoch, logs=None):
"""Be called before the training process."""
self.cfg = self.trainer.cfg
self.trainer.auto_save_ckpt = False
self.trainer.auto_save_perf = False
self.unrolled = self.trainer.cfg.get('unrolled', True)
self.device = self.trainer.cfg.get('device', 0)
self.model = self.trainer.model
self.optimizer = self.trainer.optimizer
self.lr_scheduler = self.trainer.lr_scheduler
self.loss = self.trainer.loss
self.search_alg = SearchAlgorithm(SearchSpace())
self._set_algorithm_model(self.model)
self.trainer.train_loader = self.trainer._init_dataloader(mode='train')
self.trainer.valid_loader = self.trainer._init_dataloader(mode='val')
def before_train(self, logs=None):
"""Be called before the training process."""
# Use zero valid_interval to supress default valid step
self.trainer.valid_interval = 0
self.trainer.config.report_on_epoch = True
if vega.is_torch_backend():
cudnn.benchmark = True
cudnn.enabled = True
self.search_alg = SearchAlgorithm(SearchSpace().search_space)
self.alg_policy = self.search_alg.config.policy
self.set_algorithm_model(self.trainer.model)
# setup alphas
n_individual = self.alg_policy.num_individual
self.alphas = np.stack([self.search_alg.random_sample_path()
for i in range(n_individual)], axis=0)
self.trainer.train_loader = self.trainer._init_dataloader(mode='train')
self.trainer.valid_loader = self.trainer._init_dataloader(mode='val')
def before_train(self, epoch, logs=None):
"""Be called before the training process."""
# Use zero valid_freq to supress default valid step
self.trainer.auto_save_ckpt = False
self.trainer.auto_save_perf = False
self.trainer.valid_freq = 0
cudnn.benchmark = True
cudnn.enabled = True
self.search_alg = SearchAlgorithm(SearchSpace())
self.set_algorithm_model(self.trainer.model)
# setup alphas
n_individual = self.alg_policy.num_individual
self.alphas = torch.cat([
self.trainer.model.random_single_path().unsqueeze(0)
for i in range(n_individual)
],
dim=0)
self.trainer.train_loader = self.trainer._init_dataloader(mode='train')
self.trainer.valid_loader = self.trainer._init_dataloader(mode='val')