def _train_with_cache(cls, cache, task_ids, tier, dqn_load_from,
dev_tasks_ids, max_train_actions, output_dir,
**kwargs):
if dqn_load_from is not None:
model = neural_agent.load_agent_from_folder(dqn_load_from)
else:
train_kwargs, _ = cls._extract_dqn_flags(**kwargs)
# CUSTOM
if os.path.exists(cls.ACTION_PATH_DIR):
with open(cls.ACTION_PATH_DIR + '/channel_paths.pickle',
'rb') as fp:
action_path_dict = pickle.load(fp)
train_action_paths = torch.Tensor([
action_path_dict[task]
if task in action_path_dict else torch.zeros(256, 256)
for task in task_ids
])[:, None]
if dev_tasks_ids is not None:
eval_action_paths = torch.Tensor([
action_path_dict[task]
if task in action_path_dict else torch.zeros(256, 256)
for task in dev_tasks_ids
])[:, None]
else:
eval_action_paths = None
else:
print("can't find action_path_dict!")
exit(-1)
model = neural_agent.train(output_dir,
tier,
task_ids,
cache=cache,
max_train_actions=max_train_actions,
dev_tasks_ids=dev_tasks_ids,
train_action_paths=train_action_paths,
eval_action_paths=eval_action_paths,
**train_kwargs)
if max_train_actions:
num_actions = max_train_actions
else:
num_actions = len(cache)
return dict(model=model, num_actions=num_actions, cache=cache)
def _train_with_cache(cls, cache, task_ids, tier, dqn_load_from,
dev_tasks_ids, max_train_actions, output_dir,
**kwargs):
if dqn_load_from is not None:
model = neural_agent.load_agent_from_folder(dqn_load_from)
else:
train_kwargs, _ = cls._extract_dqn_flags(**kwargs)
model = neural_agent.train(output_dir,
tier,
task_ids,
cache=cache,
max_train_actions=max_train_actions,
dev_tasks_ids=dev_tasks_ids,
**train_kwargs)
if max_train_actions:
num_actions = max_train_actions
else:
num_actions = len(cache)
return dict(model=model, num_actions=num_actions, cache=cache)