def load_task(self, task_name, load_path):
splits = ['train', 'val', 'test']
samples = []
save_paths = []
for split in splits:
file_path = os.path.join(load_path,
'{}_{}.pth'.format(task_name, split))
save_paths.append(file_path)
assert os.path.isfile(file_path), file_path
xs, ys = torch.load(file_path)
samples.append((xs, ys))
metadata_file = os.path.join(load_path, '{}.meta'.format(task_name))
if os.path.isfile(metadata_file):
meta = torch.load(metadata_file)
else:
meta = {}
task = Task(task_name,
samples,
loss,
split_names=self.split_names,
id=len(self.task_pool),
**meta)
task.save_path = save_paths
self.task_pool.append(task)
self.contains_loaded_tasks = True
return task
def _create_task(self, task_spec, name, save_path):
concepts = task_spec.src_concepts
attributes = task_spec.attributes
transformation = task_spec.transformation
n_samples_per_class = task_spec.n_samples_per_class
samples = self.get_samples(concepts, attributes, transformation,
n_samples_per_class)
if self.flatten:
samples = [(x.view(x.size(0), -1), y) for x, y in samples]
task = Task(name, samples, loss, transformation, self.split_names,
source_concepts=concepts, attributes=attributes,
creator=self.strat.descr(), generator=self,
n_samples_per_class=n_samples_per_class,
save_path=save_path)
return task