def __init__(self, dataset_type, config):
super(MultiTask, self).__init__()
self.config = config
self.dataset_type = dataset_type
self.task_names = map(lambda x: x.strip(),
self.config["tasks"].split(","))
self.tasks = []
self.tasks_lens = []
for task_name in self.task_names:
task_class = registry.get_task_class(task_name)
if task_class is None:
print("[Error] %s not present in our mapping" % task_name)
return
if task_name not in self.config["task_attributes"]:
print("[Error] No attributes present for task %s in config."
" Skipping" % task_name)
task_attributes = self.config["task_attributes"][task_name]
task_attributes["dataset_type"] = self.dataset_type
task = task_class()
task.load(**task_attributes)
self.tasks.append(task)
self.tasks_lens.append(len(task))
self.task_probabilities = [1 for _ in self.tasks]
self.num_tasks = len(self.tasks)
training_parameters = self.config["training_parameters"]
if training_parameters["task_size_proportional_sampling"]:
self.task_probabilities = self.tasks_lens[:]
len_sum = sum(self.tasks_lens)
self.task_probabilities = [
prob / len_sum for prob in self.task_probabilities
]
self.change_task()
def update_task_args(self):
args = sys.argv
task_names = None
for index, item in enumerate(args):
if item == "--tasks":
task_names = args[index + 1]
if task_names is None:
return
task_names = map(lambda x: x.strip(), task_names.split(","))
for task_name in task_names:
task_class = registry.get_task_class(task_name)
if task_class is None:
return
task_object = task_class()
task_object.init_args(self.parser)