def main(config_file, task_config_file, output_file, run_name):
config = utils.config_from_yaml_file(config_file)
# # init all the policy
# initial_policies = []
# for policy_name in config["initial_policies"]:
# policy_param = config["initial_policies"][policy_name]
#
# if policy_param['params'] is None:
# policy_param['params'] = dict()
# # add name to the parameter
# policy_param['params']["policy_name"] = policy_name
# policy_class, policy_config = utils.import_class_from_config(policy_param)
# policy = policy_class(policy_config)
# initial_policies.append(policy)
# print(f"find {len(initial_policies)} initial policies")
# init all objects
objs_config = utils.config_from_yaml_file(task_config_file)
initial_policies = []
for policy_name in objs_config["clusters"]:
policy_param = objs_config["clusters"][policy_name]
if policy_param['params'] is None:
policy_param['params'] = dict()
# add name to the parameter
policy_param['params']["policy_name"] = policy_name
policy_class, policy_config = utils.import_class_from_config(
policy_param)
policy = policy_class(policy_config)
initial_policies.append(policy)
print(f"find {len(initial_policies)} initial policies")
obj_list = [item for item in objs_config['objs'].items()]
np.random.shuffle(obj_list)
objects_to_cluster = []
for obj_name, obj_config in obj_list:
obj_config_class = MeshObject.Config().update(obj_config)
obj = MeshObject(obj_config_class, obj_name)
objects_to_cluster.append(obj)
print(f"find {len(objects_to_cluster)} objects to cluster")
updated_config = PolicyCompressor.Config().update(config)
compressor = PolicyCompressor(updated_config,
initial_policies=initial_policies,
output_file=output_file,
run_name=run_name)
nmeshes = len(objects_to_cluster)
for mesh_id, meshobj in enumerate(objects_to_cluster):
print(
f"============ {run_name} processing object {mesh_id}/{nmeshes} ============ "
)
compressor.add_object(meshobj)
if (mesh_id + 1) % 5 == 0:
compressor.output_xml()
compressor.output_xml()
compressor.print_policy_summary()
def main(config_file, task_config_file, output_file, run_name):
config = utils.config_from_yaml_file(config_file)
# build detector
if "use_detector" in config and config["use_detector"]:
detector_param = config["detector"]
detector_class, detector_config = utils.import_class_from_config(
detector_param)
detector = detector_class(detector_config)
else:
detector = None
# init all the policy
initial_policies = []
for policy_name in config["initial_policies"]:
policy_param = config["initial_policies"][policy_name]
if policy_param['params'] is None:
policy_param['params'] = dict()
# add name to the parameter
policy_param['params']["policy_name"] = policy_name
policy_class, policy_config = utils.import_class_from_config(
policy_param)
policy = policy_class(policy_config, detector)
initial_policies.append(policy)
print(f"find {len(initial_policies)} initial policies")
# init all objects
objs_config = utils.config_from_yaml_file(task_config_file)
objects_to_cluster = []
for obj_name, obj_config in objs_config['objs'].items():
obj_config_class = MeshObject.Config().update(obj_config)
obj = MeshObject(obj_config_class, obj_name)
objects_to_cluster.append(obj)
print(f"find {len(objects_to_cluster)} objects to cluster")
updated_config = PolicyCompressor.Config().update(config)
compressor = PolicyCompressor(updated_config,
initial_policies=initial_policies,
output_file=output_file,
run_name=run_name)
nmeshes = len(objects_to_cluster)
for mesh_id, meshobj in enumerate(objects_to_cluster):
print(
f"============ {run_name} processing object {mesh_id}/{nmeshes} ============ "
)
compressor.add_object(meshobj)
if (mesh_id + 1) % 5 == 0:
compressor.output_xml()
compressor.output_xml()
compressor.print_policy_summary()
def main(config_file, prefix, notes, seed):
config = config_from_yaml_file(config_file)
config['prefix'] = prefix
config['notes'] = notes
config['model_params']['batch_size'] = config['batch_size']
np.random.seed(seed)
random.seed(seed)
trainer = BehaviorCloningTrainer(config)
log.info('Finished training.')
def main(config_file, task_config_file, output_file):
config = utils.config_from_yaml_file(config_file)
# object configurations
objs_config = utils.config_from_yaml_file(task_config_file)
objects_to_get_grasps_for = list()
for obj_name, obj_config in objs_config['objs'].items():
obj_config_class = MeshObject.Config().update(obj_config)
obj = MeshObject(obj_config_class, obj_name)
objects_to_get_grasps_for.append(obj)
print(f'found {len(objects_to_get_grasps_for)} objects to find grasps for')
updated_config = Grasp6d_graspnet.Config().update(config)
grasper = Grasp6d_graspnet(updated_config)
##### Prepare the 6dof graspnet network for forward pass ######
cfg = grasp_estimator.joint_config(
grasper.vae_checkpoint_folder,
grasper.evaluator_checkpoint_folder,
)
cfg['threshold'] = grasper.grasp_conf_threshold
cfg['sample_based_improvement'] = 1 - int(
grasper.gradient_based_refinement)
cfg['num_refine_steps'] = 10 if grasper.gradient_based_refinement else 20
estimator = grasp_estimator.GraspEstimator(cfg)
sess = tf.Session()
estimator.build_network()
estimator.load_weights(sess)
##### End of 6dof graspnet network preparation ################
nmeshes = len(objects_to_get_grasps_for)
for mesh_id, meshobj in enumerate(objects_to_get_grasps_for):
print(f'---- processing {mesh_id}/{nmeshes} ------')
obj_pc, pc, pc_colors, env_state = grasper.add_object(meshobj)
# the object is added, data is collected, processed and returned
# now sample the grasps, and save them
latents = estimator.sample_latents()
generated_grasps, generated_scores, _ = estimator.predict_grasps(
sess,
obj_pc,
latents,
num_refine_steps=cfg.num_refine_steps,
)
print(
f'------ number of generated grasps are: {len(generated_grasps)} ---------'
)
save_file_path = os.path.join(grasper.output_grasps_dir,
f'grasps_{meshobj.name}.npy')
print(f'---- saving to {save_file_path} -----')
save_dict = {
'generated_grasps': generated_grasps,
'generated_scores': generated_scores,
'pcd': pc,
'pcd_color': pc_colors,
'obj_pcd': obj_pc,
'env_state': env_state,
}
np.save(save_file_path, save_dict)