def main():
parser = argparse.ArgumentParser(description="yumi_push_qmap")
parser.add_argument("--out", type=str, default="",
help="outs directory that should be evaluated")
parser.add_argument("--policy", type=str, default="",
help="policy directory that should be evaluated")
args = parser.parse_args()
# Load and check directory.
directory, path = parse_path(args)
if not os.path.isdir(os.path.join(path, "checkpoints")):
raise IOError("No model to load in {} !".format(path))
if not os.path.isfile(os.path.join(path, "config.txt")):
raise IOError("No config found in {} !".format(path))
config = load_yaml(os.path.join(path, "config.txt"))
config["visualize"] = False
config["verbose"] = True
config["debug"] = False
config["scene"] = "cubes_fixed"
config["rew_schedule"] = False
config["train_steps"] = config["eval_steps"] = 1
config["working_dir"] = path
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"]+"/"+config['camera_config'])
ckp = Checkpoint(config, tag=directory+"_qmap")
# Task definition.
task, sensor, actuator = init_task(config, cam_config, ckp)
# Get Q values from agent.
act = deepq(task, config, eval=True)
obs, Qvalues, states = task.reset(), [], []
for _ in range(config["task_max_trials_eps"]):
sample_probs = task.get_action_probabilities()
act_out = act(obs,sample_probs)
Qvalues.append(act_out[1][0])
_, segmented, _ = sensor.get_state()
states.append(segmented)
obs = task.step(act_out[0][0])[0]
print("Start creating Q-maps ...")
# Plot Q map by iterating over the whole array of actions, determine
# the pixel by "unnormalizing" and assign the Q value.
for iq, q_values in enumerate(Qvalues):
n_actions = task.action_space.n
directions = actuator.directions()
n_directions = len(directions)
n_steps = int(np.sqrt(task.action_space.n/n_directions))
q_map = np.ones((n_steps, n_steps, n_directions))*(-1000)
for x in range(n_actions):
action = actuator.undiscretize(np.asarray(x, dtype=int))
q_map[action[0],action[1],action[2]] = q_values[x]
progress_bar(x, n_actions)
# Plotting output graphs.
draw_graphs(q_map,states[iq],directions,ckp,name="{}_qvalues".format(iq))
# Clearning up directory.
os.remove(os.path.join(ckp.dir, "logging.txt"))
os.remove(os.path.join(ckp.dir, "log.txt"))
os.remove(os.path.join(ckp.dir, "progress.csv"))
print("... finished creating Q-maps.")
def train(paramset):
config = load_yaml(os.environ["YUMI_PUSH_CONFIG"]+"/simulation.yaml")
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"]+"/"+config['camera_config'])
config["visualize"] = config["debug"] = config["verbose"] = False
config["train_steps"] = 180000
tag = "!"
for param in paramset:
assert len(param.split("#")) == 3
p,v,t = param.split("#")
if t == "str": config[p] = v
elif t == "int": config[p] = int(v)
elif t == "float": config[p] = float(v)
elif t == "bool": config[p] = bool(int(v))
else: raise ValueError("unknown type")
tag = tag + param + "!"
ckp = Checkpoint(config, tag=tag)
config["working_dir"] = ckp.get_path("checkpoints/")
# Execute training in given configuration.
world = sim_world.World(config, ckp)
camera = sim_camera.RGBDCamera(world.physics_client, cam_config)
robot = sim_robot.Robot2DHand(config, ckp, world)
actuator = task_actuators.actuator_factory(config, ckp, robot)
sensor = task_sensors.Sensor(config, ckp, camera=camera)
reward = task_rewards.RewardFn(config,ckp,camera)
task = task_task.Task(sensor,actuator,reward,world,config,ckp)
learn_algo = config.get("train_algo", "ppo2")
if learn_algo == "ppo2": ppo2(task, config)
elif learn_algo == "deepq": deepq(task, config)
else: raise ValueError("Invalid training algorithm {}!".format(learn_algo))
task.close()
def main():
parser = argparse.ArgumentParser(description="yumi_push_eval")
parser.add_argument("--policy",
type=str,
default="",
help="policies directory that should be evaluated")
parser.add_argument("--out",
type=str,
default="",
help="outs directory that should be evaluated")
parser.add_argument("--visualize",
action="store_true",
help="enable visualization (default=False)")
parser.add_argument("--debug",
action="store_true",
help="enable debug mode (default=False)")
parser.add_argument("--activations",
action="store_true",
help="enable cnn layer visualization (default=False)")
parser.add_argument("--eval_steps",
type=int,
default=10,
help="number of evaluation steps")
args = parser.parse_args()
# Load and check directory.
if not ((args.policy == "" and args.out != "") \
or (args.policy != "" and args.out == "")):
raise IOError("Usage: python3 eval.py --policy=... or --out=...")
directory = args.policy if args.policy != "" else args.out
parent_dir = os.environ["YUMI_PUSH_POLICIES"] if args.policy != "" \
else os.environ["YUMI_PUSH_OUTS"]
path = os.path.join(parent_dir, directory)
if not os.path.isdir(path):
raise IOError("Directory {} not existing !".format(path))
if not os.path.isdir(os.path.join(path, "checkpoints")):
raise IOError("No model to load in {} !".format(path))
if not os.path.isfile(os.path.join(path, "config.txt")):
raise IOError("No config found in {} !".format(path))
config = load_yaml(os.path.join(path, "config.txt"))
config["visualize"] = args.visualize
config["debug"] = args.debug
config["working_dir"] = path
config["eval_steps"] = config["seg_rew_list_len"] = args.eval_steps
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/" +
config['camera_config'])
ckp = Checkpoint(config, tag=directory + "_eval")
# Load task and components.
task, _, _ = init_task(config, cam_config, ckp)
learn_algo = config.get("train_algo", "ppo2")
if learn_algo == "ppo2": ppo2(task, config, eval=True)
elif learn_algo == "deepq": deepq(task, config, eval=True)
else: raise ValueError("Invalid training algorithm {}!".format(learn_algo))
task.close()
def main():
config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/simulation.yaml")
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/" +
config['camera_config'])
ckp = Checkpoint(config)
config["working_dir"] = ckp.get_path("checkpoints/")
task, _, _ = init_task(config, cam_config, ckp)
learn_algo = config.get("train_algo", "ppo2")
if learn_algo == "ppo2": ppo2(task, config)
elif learn_algo == "deepq": deepq(task, config)
else: raise ValueError("Invalid training algorithm {}!".format(learn_algo))
task.close()
def main():
parser = argparse.ArgumentParser(description="yumi_push_qmap")
parser.add_argument("--out",
type=str,
default="",
help="outs directory that should be evaluated")
parser.add_argument("--policy",
type=str,
default="",
help="policy directory that should be evaluated")
args = parser.parse_args()
# Load and check directory.
if not ((args.policy == "" and args.out != "") \
or (args.policy != "" and args.out == "")):
raise IOError("Usage: python3 eval.py --policy=... or --out=...")
directory = args.policy if args.policy != "" else args.out
parent_dir = os.environ["YUMI_PUSH_POLICIES"] if args.policy != "" \
else os.environ["YUMI_PUSH_OUTS"]
path = os.path.join(parent_dir, directory)
if not os.path.isdir(path):
raise IOError("Directory {} not existing !".format(path))
if not os.path.isdir(os.path.join(path, "checkpoints")):
raise IOError("No model to load in {} !".format(path))
if not os.path.isfile(os.path.join(path, "config.txt")):
raise IOError("No config found in {} !".format(path))
config = load_yaml(os.path.join(path, "config.txt"))
config["visualize"] = config["verbose"] = True
config["debug"] = False
config["rew_schedule"] = False
config["working_dir"] = path
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/" +
config['camera_config'])
ckp = Checkpoint(config, tag=directory + "_multistep")
# Task definition.
task, sensor, actuator = new_init_task(config, cam_config, ckp)
learn_algo = config.get("train_algo", "ppo2")
if learn_algo == "ppo2": ppo2(task, config, eval=True)
elif learn_algo == "deepq": deepq(task, config, eval=True)
else: raise ValueError("Invalid training algorithm {}!".format(learn_algo))
task.close()
def from_yaml(file_path):
"""Read a transform from a yaml file.
Example of the content of such a file:
transform:
translation: [1., 2., 3.]
rotation: [0., 0., 0., 1.]
Args:
file_path: The path to the YAML file.
Returns:
A 4x4 homogeneous transformation matrix.
"""
cfg = io_utils.load_yaml(file_path)
return from_dict(cfg["transform"])
def main():
parser = argparse.ArgumentParser(description="yumi_push_reward_map")
parser.add_argument("--out",
type=str,
default="",
help="outs directory that should be evaluated")
parser.add_argument("--policy",
type=str,
default="",
help="policy directory that should be evaluated")
parser.add_argument("--config",
type=str,
default="",
help="config file that should be evaluated")
args = parser.parse_args()
# Load and check directory.
if (int(args.policy == "") + int(args.out == "") +
int(args.config == "")) != 2:
raise IOError(
"Usage: python3 reward_map.py --policy=... or --out=... or --config=..."
)
if args.policy != "":
tag = args.policy
file = os.path.join(args.policy, "config.txt")
file = os.path.join(os.environ["YUMI_PUSH_POLICIES"], file)
elif args.out != "":
tag = args.out
file = os.path.join(args.out, "config.txt")
file = os.path.join(os.environ["YUMI_PUSH_OUTS"], file)
else:
tag = args.config.replace(".yaml", "")
file = os.path.join(os.environ["YUMI_PUSH_CONFIG"], args.config)
if not os.path.isfile(file):
raise IOError("File {} not existing !".format(file))
# Load configuration and create output file (checkpoint).
config = load_yaml(file)
config["visualize"] = False
config["verbose"] = True
config["debug"] = False
config["scene"] = "cubes_fixed"
config["rew_schedule"] = False
config["train_algo"] = ""
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/" +
config['camera_config'])
ckp = Checkpoint(config, tag=tag + "_reward_map")
# Task definition.
task, sensor, actuator = init_task(config, cam_config, ckp)
# Get number of pushing actions (not implemented for all actuators, so will
# throw error for "too continuous" actuators).
directions = actuator.directions()
n_directions = len(directions)
# Iterate over all actions and determine reward
# (assume start_point are 1&2 action and 3 pushing direction).
print("Start creating reward map ...")
rewards = None
# For discrete action space the actions simply are a integer number.
if config["act_discrete"]:
n_actions = task.action_space.n
n_action_steps = int(np.sqrt(task.action_space.n / n_directions))
rewards = np.ones(
(n_action_steps, n_action_steps, n_directions)) * (-1000)
for x in range(n_actions):
task.reset()
action = actuator.undiscretize(np.asarray(x, dtype=int))
_, reward, done, _ = task.step(np.asarray(x, dtype=int))
rewards[action[0], action[1], action[2]] = reward
progress_bar(x, n_actions)
# For continuous action space discretize the actions (assumes action to be
# [starting point, pushing_direction]).
else:
res_cont = 0.1 #[m]
n_action_steps = int(2.0 / res_cont)
n_actions = n_directions * (n_action_steps**2)
rewards = np.ones(
(n_action_steps, n_action_steps, n_directions)) * (-1000)
i = 0
for ix in range(n_action_steps):
for iy in range(n_action_steps):
for id in range(n_directions):
x, y = -1.0 + ix * res_cont, -1.0 + iy * res_cont,
d = -0.9 + id * (2.0 / n_directions)
task.reset()
_, reward, done, _ = task.step(np.asarray([x, y, d]))
rewards[ix, iy, id] = reward
i = i + 1
progress_bar(i, n_actions)
# Plotting output graphs.
_, segmented, _ = sensor.get_state()
draw_graphs(rewards, segmented, directions, ckp, name="rewards")
# Clearning up directory.
os.remove(os.path.join(ckp.dir, "log_num_steps.pdf"))
os.remove(os.path.join(ckp.dir, "log_reward.pdf"))
os.remove(os.path.join(ckp.dir, "log_success_rate.pdf"))
os.remove(os.path.join(ckp.dir, "logging.txt"))
print("... finished creating reward map.")
def main():
parser = argparse.ArgumentParser(description="yumi_push_eval")
parser.add_argument("--policy",
type=str,
default="",
help="policies directory that should be evaluated")
parser.add_argument("--out",
type=str,
default="",
help="outs directory that should be evaluated")
parser.add_argument("--visualize",
action="store_true",
help="enable visualization (default=False)")
parser.add_argument("--debug",
action="store_true",
help="enable debug mode (default=False)")
parser.add_argument("--activations",
action="store_true",
help="enable cnn layer visualization (default=False)")
args = parser.parse_args()
# Load and check directory.
directory, path = parse_path(args)
if not os.path.isdir(os.path.join(path, "checkpoints")):
raise IOError("No model to load in {} !".format(path))
if not os.path.isfile(os.path.join(path, "config.txt")):
raise IOError("No config found in {} !".format(path))
config = load_yaml(os.path.join(path, "config.txt"))
config["visualize"] = args.visualize
config["debug"] = args.debug
config["working_dir"] = path
cam_config = load_yaml(os.environ["YUMI_PUSH_CONFIG"] + "/" +
config['camera_config'])
ckp = Checkpoint(config, tag=directory + "_eval")
# Load task and components.
world = sim_world.World(config, ckp)
camera = sim_camera.RGBDCamera(world.physics_client, cam_config)
robot = sim_robot.Robot2DHand(config, ckp, world)
actuator = task_actuators.actuator_factory(config, ckp, robot)
sensor = task_sensors.Sensor(config, ckp, camera=camera)
reward = task_rewards.RewardFn(config, ckp, camera)
task = task_task.Task(sensor, actuator, reward, world, config, ckp)
learn_algo = config.get("train_algo", "ppo2")
if learn_algo == "ppo2": ppo2(task, config, eval=True)
elif learn_algo == "deepq": deepq(task, config, eval=True)
else: raise ValueError("Invalid training algorithm {}!".format(learn_algo))
# Plot activations (if activated).
if args.activations:
fig = plt.figure() # an empty figure with no axes
fig.suptitle(
'No axes on this figure') # Add a title so we know which it is
fig, ax_lst = plt.subplots(2, 2)
# Access layer activation of last evaluation by accessing the
# tensorflow namespace, in which the network and all its parameters
# are saved.
session = model.sess
imageToUse = sensor.get_state()
graph = session.graph
# First convolution.
layer = graph.get_tensor_by_name("ppo2_model/pi/c1/Conv2D:0")
tf_placeholder = session.graph.get_tensor_by_name("ppo2_model/Ob:0")
getActivations(session, layer, tf_placeholder, imageToUse)
# Second convolution.
layer = graph.get_tensor_by_name("ppo2_model/pi/c2/Conv2D:0")
tf_placeholder = session.graph.get_tensor_by_name("ppo2_model/Ob:0")
getActivations(session, layer, tf_placeholder, imageToUse)
task.close()