def real_eval(cls, cache, model, actions, task_ids, tier,
max_attempts_per_task, eval_batch_size, finetune_iterations,
refine_iterations, refine_loss, refine_lr):
# TODO: move to a flag.
finetune_lr = 1e-4
model.cuda()
simulator = phyre.initialize_simulator(task_ids, tier)
observations = simulator.initial_scenes
assert tuple(task_ids) == simulator.task_ids
logging.info('Ranking %d actions and simulating top %d', len(actions),
max_attempts_per_task)
if refine_iterations > 0:
logging.info(
'Will do refining for %d iterations with lr=%e and loss=%s',
refine_iterations, refine_lr, refine_loss)
evaluator = phyre.Evaluator(task_ids)
for task_index in tqdm.trange(len(task_ids)):
task_id = simulator.task_ids[task_index]
if refine_iterations > 0:
refined_actions = neural_agent.refine_actions(
model, actions, observations[task_index], refine_lr,
refine_iterations, eval_batch_size, refine_loss)
else:
refined_actions = actions
scores = neural_agent.eval_actions(model, refined_actions,
eval_batch_size,
observations[task_index])
# Order of descendig scores.
action_order = np.argsort(-scores)
if not refine_iterations > 0:
statuses = cache.load_simulation_states(task_id)
finetune_data = []
for action_id in action_order:
if evaluator.get_attempts_for_task(
task_index) >= max_attempts_per_task:
break
action = refined_actions[action_id]
if refine_iterations > 0:
status = simulator.simulate_action(
task_index,
action,
need_images=False,
need_scenes=False).status
else:
status = phyre.SimulationStatus(statuses[action_id])
finetune_data.append((task_index, status, action))
evaluator.maybe_log_attempt(task_index, status)
if evaluator.get_attempts_for_task(task_index) == 0:
logging.warning('Made 0 attempts for task %s', task_id)
if finetune_iterations > 0:
neural_agent.finetune(model, finetune_data, simulator,
finetune_lr, finetune_iterations)
return evaluator
def real_eval(cls, cache, model, actions, task_ids, tier,
max_attempts_per_task, eval_batch_size, finetune_iterations,
refine_iterations, refine_loss, refine_lr):
# TODO: move to a flag.
finetune_lr = 1e-4
model.cuda()
simulator = phyre.initialize_simulator(task_ids, tier)
observations = simulator.initial_scenes
# 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)
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].cuda()
else:
print("can't find action_path_dict!")
exit(-1)
assert tuple(task_ids) == simulator.task_ids
logging.info('Ranking %d actions and simulating top %d', len(actions),
max_attempts_per_task)
if refine_iterations > 0:
logging.info(
'Will do refining for %d iterations with lr=%e and loss=%s',
refine_iterations, refine_lr, refine_loss)
evaluator = phyre.Evaluator(task_ids)
for task_index in tqdm.trange(len(task_ids)):
task_id = simulator.task_ids[task_index]
if refine_iterations > 0:
refined_actions = neural_agent.refine_actions(
model, actions, observations[task_index], refine_lr,
refine_iterations, eval_batch_size, refine_loss)
else:
refined_actions = actions
scores = neural_agent.eval_actions(
model,
refined_actions,
eval_batch_size,
observations[task_index],
action_path=action_paths[task_index])
# Order of descendig scores.
action_order = np.argsort(-scores)
if not refine_iterations > 0:
statuses = cache.load_simulation_states(task_id)
finetune_data = []
for action_id in action_order:
if evaluator.get_attempts_for_task(
task_index) >= max_attempts_per_task:
break
action = refined_actions[action_id]
if refine_iterations > 0:
status = simulator.simulate_action(
task_index,
action,
need_images=False,
need_scenes=False).status
else:
status = phyre.SimulationStatus(statuses[action_id])
finetune_data.append((task_index, status, action))
evaluator.maybe_log_attempt(task_index, status)
if evaluator.get_attempts_for_task(task_index) == 0:
logging.warning('Made 0 attempts for task %s', task_id)
if finetune_iterations > 0:
neural_agent.finetune(model, finetune_data, simulator,
finetune_lr, finetune_iterations)
return evaluator