def cli_train(triple_source, config, model_class, optimizer_class):
data_loader = _build_data_loader_based_on_model(model_class, triple_source, config)
model = estimate.train(
triple_source,
config,
data_loader,
model_class,
optimizer_class,
drawer=create_drawer(config))
def cli_profile(triple_source, config, model_class, optimizer_class):
with torch.autograd.profiler.profile() as prof:
model = estimate.train(
triple_source,
config,
model_class,
optimizer_class,
drawer=create_drawer(config))
print(prof)
def cli_train_and_validate(triple_source, config, model_class, optimizer_class,
validation_evaluator):
data_loader = _build_data_loader_based_on_model(model_class, triple_source, config)
model = estimate.train_and_validate(
triple_source,
config,
data_loader,
model_class,
optimizer_class,
validation_evaluator,
drawer=create_drawer(config))
d.update(changed_config)
if os.path.exists(
os.path.join('model_states', changed_config['name'],
'checkpoint.pth.tar_1000')):
continue
search_config = utils.build_config_with_dict(d)
utils.seed_modules(config,
numpy_seed=10000,
torch_seed=20000,
torcu_cuda_seed_all=2192,
cuda_deterministic=True,
kgegrok_base_seed=30000,
cuda_benchmark=config.cudnn_benchmark)
optimizer_class = utils.load_class_from_module(search_config.optimizer,
'torch.optim')
data_loader = data.create_dataloader(triple_source, search_config,
model_class.require_labels())
drawer = stats.create_drawer(search_config)
model = estimate.train_and_validate(
triple_source,
search_config,
data_loader,
model_class,
optimizer_class,
evaluator=evaluator,
stat_gather=evaluation.build_stat_gather_from_config(
config, drawer),
drawer=drawer)