def train_agent(train, pickle_file, agent_type, env_kwargs, parms):
bin_path = "bin/" + pickle_file
if (path.exists(bin_path)):
if agent_type == "a2c":
print("Loading A2C Agent")
RL_model = A2C.load(
bin_path,
tensorboard_log=f"{config.TENSORBOARD_LOG_DIR}/{agent_type}")
elif agent_type == "ddpg":
print("Loading DDPG Agent")
RL_model = DDPG.load(
bin_path,
tensorboard_log=f"{config.TENSORBOARD_LOG_DIR}/{agent_type}")
elif agent_type == "ppo":
print("Loading PPO2 Agent")
RL_model = PPO2.load(
bin_path,
tensorboard_log=f"{config.TENSORBOARD_LOG_DIR}/{agent_type}")
else:
e_train_gym = ipenv.PortfolioAllocEnv(df=train, **env_kwargs)
env_train, _ = e_train_gym.get_sb_env()
agent = ipagent.IPRLAgent(env=env_train)
model = agent.get_model(model_name=agent_type, model_kwargs=parms)
RL_model = agent.train_model(model=model,
tb_log_name=agent_type,
total_timesteps=1000000)
RL_model.save(bin_path)
return RL_model
def test(env_id, seed, policy):
"""
Train PPO2 model for atari environment, for testing purposes
:param env_id: (str) the environment id string
:param seed: (int) Used to seed the random generator.
:param policy: (Object) The policy model to use (MLP, CNN, LSTM, ...)
"""
# if 'lstm' in policy:
# print('LSTM policies not supported for drawing')
# return 1
env = DummyVecEnv([PadEnvRender for _ in range(1)]) # Need for lstm
# else:
# env = PadEnvRender()
env = VecFrameStack(env, 8)
model = PPO2.load('./pad_5combo_ppo2.pkl', env)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
done = done.any()
episode_rew += rew
time.sleep(1 / 24.)
if done:
print('Episode reward:', rew)
def do_ppo(args, start_theta, parent_this_run_dir, full_space_save_dir):
"""
Runs the test
"""
logger.log(f"#######CMA and then PPO TRAIN: {args}")
this_conti_ppo_run_dir = get_ppo_part(parent_this_run_dir)
log_dir = get_log_dir(this_conti_ppo_run_dir)
conti_ppo_save_dir = get_save_dir(this_conti_ppo_run_dir)
logger.configure(log_dir)
full_param_traj_dir_path = get_full_params_dir(this_conti_ppo_run_dir)
if os.path.exists(full_param_traj_dir_path):
import shutil
shutil.rmtree(full_param_traj_dir_path)
os.makedirs(full_param_traj_dir_path)
if os.path.exists(conti_ppo_save_dir):
import shutil
shutil.rmtree(conti_ppo_save_dir)
os.makedirs(conti_ppo_save_dir)
def make_env():
env_out = gym.make(args.env)
env_out.env.disableViewer = True
env_out.env.visualize = False
env_out = bench.Monitor(env_out, logger.get_dir(), allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
if args.normalize:
env = VecNormalize(env)
model = PPO2.load(f"{full_space_save_dir}/ppo2")
model.set_from_flat(start_theta)
if args.normalize:
env.load_running_average(full_space_save_dir)
model.set_env(env)
run_info = {"run_num": args.run_num,
"env_id": args.env,
"full_param_traj_dir_path": full_param_traj_dir_path}
# model = PPO2(policy=policy, env=env, n_steps=args.n_steps, nminibatches=args.nminibatches, lam=0.95, gamma=0.99,
# noptepochs=10,
# ent_coef=0.0, learning_rate=3e-4, cliprange=0.2, optimizer=args.optimizer)
model.tell_run_info(run_info)
episode_returns = model.learn(total_timesteps=args.ppo_num_timesteps)
model.save(f"{conti_ppo_save_dir}/ppo2")
env.save_running_average(conti_ppo_save_dir)
return episode_returns, full_param_traj_dir_path
def load(self, path, env):
if self.trpo():
return TRPO.load(path, env=env)
elif self.ppo():
return PPO2.load(path, env=env)
else:
return SAC.load(path, env=env)
def __init__(self,
policy_file="../models/controllers/PPO/CartPole-v1.pkl",
mass_prior=None,
length_prior=None,
episodes_per_params=1,
seed=1995,
params=["length", "masspole"],
steps_per_episode=50,
sufficient_stats="Cross-Correlation"):
self.env = CartPoleEnv()
self.seed = seed
self.cached_data = None
self.params_scaler = None
self.params = params
self.steps_per_episode = steps_per_episode
self.sufficient_stats = sufficient_stats
self.policy = PPO2.load(policy_file)
if mass_prior is None:
self.m_low = 0.1
self.m_high = 2.0
self.m_prior = self.sample_mass_from_uniform_prior
if length_prior is None:
self.l_low = 0.1
self.l_high = 2.0
self.l_prior = self.sample_length_from_uniform_prior
def __init__(self,
policy_file="../models/controllers/PPO/Pendulum-v0.pkl",
mass_prior=None,
length_prior=None,
episodes_per_params=1,
seed=1995,
params=["length", "mass"],
steps_per_episode=200,
sufficient_stats="Cross-Correlation",
load_from_file=False,
assets_path=".",
filename=""):
self.env = PendulumEnv()
self.seed = seed
self.cached_data = None
self.params_scaler = None
self.params = params
self.steps_per_episode = steps_per_episode
self.sufficient_stats = sufficient_stats
self.assets_path = assets_path
self.load_from_file = load_from_file
self.data_file = os.path.join(assets_path + filename)
self.policy = PPO2.load(policy_file)
if mass_prior is None:
self.m_low = 0.1
self.m_high = 2.0
self.m_prior = self.sample_mass_from_uniform_prior
if length_prior is None:
self.l_low = 0.1
self.l_high = 2.0
self.l_prior = self.sample_length_from_uniform_prior
def main():
import sys
logger.log(sys.argv)
common_arg_parser = get_common_parser()
args, cma_unknown_args = common_arg_parser.parse_known_args()
this_run_dir = get_dir_path_for_this_run(args)
plot_dir_alg = get_plot_dir(args)
traj_params_dir_name = get_full_params_dir(this_run_dir)
intermediate_data_dir = get_intermediate_data_dir(this_run_dir,
params_scope="pi")
save_dir = get_save_dir(this_run_dir)
if not os.path.exists(intermediate_data_dir):
os.makedirs(intermediate_data_dir)
if not os.path.exists(plot_dir_alg):
os.makedirs(plot_dir_alg)
final_file = get_full_param_traj_file_path(traj_params_dir_name,
"pi_final")
final_params = pd.read_csv(final_file, header=None).values[0]
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
if args.normalize:
env = VecNormalize(env)
model = PPO2.load(f"{save_dir}/ppo2") # this also loads V function
model.set_pi_from_flat(final_params)
if args.normalize:
env.load_running_average(save_dir)
obz_tensor = model.act_model.fake_input_tensor
some_neuron = model.act_model.policy_neurons[2][-1]
grads = tf.gradients(tf.math.negative(some_neuron), obz_tensor)
grads = list(zip(grads, obz_tensor))
trainer = tf.train.AdamOptimizer(learning_rate=0.01, epsilon=1e-5)
train_op = trainer.apply_gradients(grads)
for i in range(10000):
obz, _ = model.sess.run([obz_tensor, train_op])
def neuron_values_generator(args, save_dir, pi_theta, eval_timesteps):
# logger.log(f"#######EVAL: {args}")
neuron_values_list = []
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
if args.normalize:
env = VecNormalize(env)
# policy = MlpPolicy
# # model = PPO2.load(f"{save_dir}/ppo2") # this also loads V function
# model = PPO2(policy=policy, env=env, n_steps=args.n_steps, nminibatches=args.nminibatches, lam=0.95, gamma=0.99, noptepochs=10,
# ent_coef=0.0, learning_rate=3e-4, cliprange=0.2, optimizer=args.optimizer)
model = PPO2.load(f"{save_dir}/ppo2") # this also loads V function
if pi_theta is not None:
model.set_pi_from_flat(pi_theta)
if args.normalize:
env.load_running_average(save_dir)
obs = np.zeros((env.num_envs, ) + env.observation_space.shape)
obs[:] = env.reset()
env.render()
ep_infos = []
while 1:
neuron_values, actions, _, _, _ = model.step_with_neurons(obs)
# neuron_values = model.give_neuron_values(obs)
# neuron_values_list.append( neuron_values )
yield neuron_values
obs, rew, done, infos = env.step(actions)
env.render()
# time.sleep(1)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
# env.render()
done = done.any()
if done:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
obs = env.reset()
def cont_learn():
print('Continue learning....')
env = gym.make('CarRacing-v0')
env = Monitor(env, log_dir, allow_early_resets=True)
env = DummyVecEnv([lambda: env])
trained_model = PPO2.load("CarRacing_model_PPO2.pkl")
trained_model.set_env(env)
trained_model.learn(300000)
print("Saving model to CarRacing_model.pkl")
trained_model.save("CarRacing_model_PPO2.pkl")
plot_results(log_dir)
def visualize_neurons(args, save_dir, pi_theta, eval_timesteps):
# logger.log(f"#######EVAL: {args}")
def make_env():
env_out = gym.make(args.env)
env_out.env.disableViewer = True
env_out.env.visualize = False
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
if args.normalize:
env = VecNormalize(env)
model = PPO2.load(f"{save_dir}/ppo2") # this also loads V function
if pi_theta is not None:
model.set_pi_from_flat(pi_theta)
if args.normalize:
env.load_running_average(save_dir)
obs = np.zeros((env.num_envs, ) + env.observation_space.shape)
obs[:] = env.reset()
ep_infos = []
for _ in range(eval_timesteps):
actions = model.step(obs)[0]
neuron_values = model.give_neuron_values(obs)
obs, rew, done, infos = env.step(actions)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
# env.render()
done = done.any()
if done:
if pi_theta is None:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
obs = env.reset()
return safe_mean([ep_info['r'] for ep_info in ep_infos])
def test_cnn_lstm_policy(request, policy):
model_fname = './test_model_{}.zip'.format(request.node.name)
try:
env = make_env(0)
model = PPO2(policy, env, nminibatches=1)
model.learn(total_timesteps=15)
env = model.get_env()
evaluate_policy(model, env, n_eval_episodes=5)
# saving
model.save(model_fname)
del model, env
# loading
_ = PPO2.load(model_fname, policy=policy)
finally:
if os.path.exists(model_fname):
os.remove(model_fname)
def run():
"""
Run a trained model for the pong problem
"""
env = gym.make('CarRacing-v0')
env = DummyVecEnv([lambda: env])
# model = PPO2.load("CarRacing_model_PPO1_"+ str(5) +".pkl", env)
model = PPO2.load("CarRacing_model_PPO2_5.pkl", env)
avg_rew = evaluate(model=model, env=env, num_steps=10000)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
def main(env, load_path, fig_path):
# skip over 1-baxter-no-penalty (no log monitor.csv)
if load_path == "1-baxter-no-penalty":
plot = False
else:
plot = True
# arguments
print("env %s; load_path %s; fig_path %s;" % (env, load_path, fig_path))
log_path = os.getcwd() + "/log/" + load_path
os.makedirs(os.getcwd() + "/figs/" + "/", exist_ok=True)
fig_path = os.getcwd() + "/figs/" + "/" + fig_path
load_path = os.getcwd() + "/models/" + load_path
# make environment, flattened environment, vectorized environment
env = gym.make(env)
env = gym.wrappers.FlattenDictWrapper(env, ['observation', 'achieved_goal', 'desired_goal'])
env = DummyVecEnv([lambda: env])
# load model
model = PPO2.load(load_path, env=env)
obs_initial = env.reset()
obs = obs_initial
# plot results
if plot:
plot_results(fig_path, log_path)
# initializations
niter = 10
counter = 0
timestep = 0
results = [[[0,0,0] for i in range(100)], [[0,0,0,0] for i in range(100)]]
current = [[[0,0,0] for i in range(100)], [[0,0,0,0] for i in range(100)]]
print("==============================")
# check initial positions and quaternions
print("grip", env.envs[0].env.env.sim.data.get_site_xpos('grip'))
print("box", env.envs[0].env.env.sim.data.get_site_xpos('box'))
print("tool", env.envs[0].env.env.sim.data.get_site_xpos('tool'))
print("mocap", env.envs[0].env.env.sim.data.mocap_pos)
print("quat", env.envs[0].env.env.sim.data.mocap_quat)
print("==============================")
# mocap quaternion check
for i in range(5):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
quat = env.envs[0].env.env.sim.data.mocap_quat
print("obs", obs)
print("quat", quat)
print("==============================")
# start rendering
dists = []
box_goal_pos = np.array([0.6, 0.05, -0.17])
while True:
if counter == niter:
break
action, _states = model.predict(obs)
obs_old = obs
obs, rewards, dones, info = env.step(action)
quaternion = env.envs[0].env.env.sim.data.mocap_quat
if obs.all() == obs_initial.all():
if counter % 10 == 0:
xyzs = current[0]
quats = current[1]
print(xyzs)
print(quats)
filename = log_path + "/" + "results_" + str(counter) + ".txt"
os.makedirs(log_path + "/", exist_ok=True)
file = open(filename, 'w+')
for xyz, quat in zip(xyzs, quats):
for coord in xyz:
file.write(str(coord) + " ")
for quat_coord in quat:
file.write(str(quat_coord) + " ")
file.write("\n")
file.close()
box_end_pos = np.array(obs_old[0][3:6].tolist())
print(box_end_pos)
print(np.shape(box_end_pos))
print(box_goal_pos)
print(np.shape(box_goal_pos))
dists.append(np.linalg.norm(box_goal_pos - box_end_pos))
current = [[[0,0,0] for i in range(100)], [[0,0,0,0] for i in range(100)]]
timestep = 0
counter += 1
print(timestep)
print("obs", obs)
print("quat", quaternion)
# for average trajectory, smoothed
for i in range(3):
results[0][timestep][i] += obs[0][:3].tolist()[i]
for j in range(4):
results[1][timestep][j] += quaternion[0].tolist()[j]
# for current trajectory
for i in range(3):
current[0][timestep][i] += obs[0][:3].tolist()[i]
for j in range(4):
current[1][timestep][j] += quaternion[0].tolist()[j]
timestep += 1
env.render()
# smooth paths by taking average, and calculate mean distance to goal state
for timestep in range(100):
for i in range(3):
results[0][timeste][i] /= niter
for j in range(4):
results[0][timestep][j] /= niter
dist = np.mean(dists)
# print and write to file
xyzs = results[0]
quats = results[1]
filename = log_path + "/" + "results_avg.txt"
os.makedirs(log_path + "/", exist_ok=True)
file = open(filename, 'w+')
for xyz, quat in zip(xyzs, quats):
for coord in xyz:
file.write(str(coord) + " ")
for quat_coord in quat:
file.write(str(quat_coord) + " ")
file.write("\n")
file.close()
# print average distances
print("average distance of box from end goal: %f" % dist)
def __init__(self):
self.model = PPO2.load(os.path.expanduser(
"~/Code/drl_local_planner_ros_stable_baselines/example_agents/ppo2_1_raw_data_cont_0/ppo2_1_raw_data_cont_0.pkl")) # noqa
# qvel[i,:] = env_in.sim.data.qvel[[7,8,10,12,14,16,17,19,21,23]]
qvel[i, :] = env_in.GetMotorVelocities()
# torque[i,:] = env_in.sim.data.actuator_force
torque[i, :] = env_in.GetMotorTorques()
i = (i + 1) % total_data_length
return True
###############################################################################
# # Use this code for testing the basic controller
# Create the stoch mujoco environment
# env = stoch2_gym_mjc_env.Stoch2Env()
env = stoch2_gym_env.Stoch2Env(render=True)
model_test = PPO2.load(
dir_name +
"/tflow_log/model_trot") # model_trot_200kiter_0pen_notermination
obs = env.reset()
print("Render mode...")
for _ in range(15):
action, _states = model_test.predict(obs, deterministic=True)
obs, reward, done, _ = env.step(action, callback=render_callback)
# if done:
# break
pickle.dump(qpos[0:total_data_length:int(total_data_length /
total_data_length_for_saving)],
open(input_file, "wb")) # save it into a file named save.p
pickle.dump(action, open(action_file, "wb"))
def visualize_policy_and_collect_COM(
augment_num_timesteps, top_num_to_include_slice, augment_seed,
augment_run_num, network_size, policy_env, policy_num_timesteps,
policy_run_num, policy_seed, eval_seed, eval_run_num, learning_rate,
additional_note, metric_param):
result_dir = get_result_dir(policy_env, policy_num_timesteps,
policy_run_num, policy_seed, eval_seed,
eval_run_num, additional_note, metric_param)
args = AttributeDict()
args.normalize = True
args.num_timesteps = augment_num_timesteps
args.run_num = augment_run_num
args.alg = "ppo2"
args.seed = augment_seed
logger.log(f"#######VISUALIZE: {args}")
# non_linear_global_dict
linear_global_dict, non_linear_global_dict, lagrangian_values, input_values, layers_values, all_weights = read_all_data(
policy_env,
policy_num_timesteps,
policy_run_num,
policy_seed,
eval_seed,
eval_run_num,
additional_note=additional_note)
timestamp = get_time_stamp('%Y_%m_%d_%H_%M_%S')
experiment_label = f"learning_rate_{learning_rate}timestamp_{timestamp}_augment_num_timesteps{augment_num_timesteps}" \
f"_top_num_to_include{top_num_to_include_slice.start}_{top_num_to_include_slice.stop}" \
f"_augment_seed{augment_seed}_augment_run_num{augment_run_num}_network_size{network_size}" \
f"_policy_num_timesteps{policy_num_timesteps}_policy_run_num{policy_run_num}_policy_seed{policy_seed}" \
f"_eval_seed{eval_seed}_eval_run_num{eval_run_num}_additional_note_{additional_note}"
entry_point = 'gym.envs.dart:DartWalker2dEnv_aug_input'
this_run_dir = get_experiment_path_for_this_run(
entry_point,
args.num_timesteps,
args.run_num,
args.seed,
learning_rate=learning_rate,
top_num_to_include=top_num_to_include_slice,
result_dir=result_dir,
network_size=network_size,
metric_param=metric_param)
traj_params_dir_name = get_full_params_dir(this_run_dir)
save_dir = get_save_dir(this_run_dir)
aug_plot_dir = get_aug_plot_dir(this_run_dir) + "_vis"
final_file = get_full_param_traj_file_path(traj_params_dir_name,
"pi_final")
final_params = pd.read_csv(final_file, header=None).values[0]
args.env = f'{experiment_label}_{entry_point}-v1'
register(id=args.env,
entry_point=entry_point,
max_episode_steps=1000,
kwargs={
'linear_global_dict': linear_global_dict,
'non_linear_global_dict': non_linear_global_dict,
'top_to_include_slice': top_num_to_include_slice,
'aug_plot_dir': aug_plot_dir,
"lagrangian_values": lagrangian_values,
"layers_values": layers_values
})
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
walker_env = env.envs[0].env.env
walker_env.disableViewer = False
if args.normalize:
env = VecNormalize(env)
set_global_seeds(args.seed)
walker_env.seed(args.seed)
model = PPO2.load(f"{save_dir}/ppo2", seed=augment_seed)
model.set_pi_from_flat(final_params)
if args.normalize:
env.load_running_average(save_dir)
sk = env.venv.envs[0].env.env.robot_skeleton
lagrangian_values = {}
obs = np.zeros((env.num_envs, ) + env.observation_space.shape)
obs[:] = env.reset()
env = VecVideoRecorder(env,
aug_plot_dir,
record_video_trigger=lambda x: x == 0,
video_length=3000,
name_prefix="vis_this_policy")
lagrangian_values["M"] = [sk.M.reshape((-1, 1))]
lagrangian_values["COM"] = [sk.C.reshape((-1, 1))]
lagrangian_values["Coriolis"] = [sk.c.reshape((-1, 1))]
lagrangian_values["q"] = [sk.q.reshape((-1, 1))]
lagrangian_values["dq"] = [sk.dq.reshape((-1, 1))]
contact_values = {}
neuron_values = model.give_neuron_values(obs)
raw_layer_values_list = [[neuron_value.reshape((-1, 1))]
for neuron_value in neuron_values]
env.render()
ep_infos = []
steps_to_first_done = 0
first_done = False
# epi_rew = 0
for _ in range(3000):
actions = model.step(obs)[0]
# yield neuron_values
obs, rew, done, infos = env.step(actions)
# epi_rew+= rew[0]
if done and not first_done:
first_done = True
if not first_done:
steps_to_first_done += 1
neuron_values = model.give_neuron_values(obs)
for i, layer in enumerate(neuron_values):
raw_layer_values_list[i].append(layer.reshape((-1, 1)))
# fill_contacts_jac_dict(infos[0]["contacts"], contact_dict=contact_values, neuron_values=neuron_values)
lagrangian_values["M"].append(sk.M.reshape((-1, 1)))
lagrangian_values["q"].append(sk.q.reshape((-1, 1)))
lagrangian_values["dq"].append(sk.dq.reshape((-1, 1)))
lagrangian_values["COM"].append(sk.C.reshape((-1, 1)))
lagrangian_values["Coriolis"].append(sk.c.reshape((-1, 1)))
# env.render()
# time.sleep(1)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
env.render()
done = done.any()
if done:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
# print(f'episode_rew={epi_rew}')
# epi_rew = 0
obs = env.reset()
#Hstack into a big matrix
lagrangian_values["M"] = np.hstack(lagrangian_values["M"])
lagrangian_values["COM"] = np.hstack(lagrangian_values["COM"])
lagrangian_values["Coriolis"] = np.hstack(lagrangian_values["Coriolis"])
lagrangian_values["q"] = np.hstack(lagrangian_values["q"])
lagrangian_values["dq"] = np.hstack(lagrangian_values["dq"])
# for contact_body_name, l in contact_values.items():
# body_contact_dict = contact_values[contact_body_name]
# for name, l in body_contact_dict.items():
# body_contact_dict[name] = np.hstack(body_contact_dict[name])
input_values = np.hstack(raw_layer_values_list[0])
layers_values = [
np.hstack(layer_list) for layer_list in raw_layer_values_list
][1:-2] # drop variance and inputs
for i, com in enumerate(lagrangian_values["COM"]):
plt.figure()
plt.plot(np.arange(len(com)), com)
plt.xlabel("time")
plt.ylabel(f"COM{i}")
plt.savefig(f"{aug_plot_dir}/COM{i}.jpg")
plt.close()
return env
set_global_seeds(seed)
return _init
# Make Environments
envs = [make_env(rank=0, seed=0, framerange=(start_frame, start_frame))]
env = SubprocVecEnv(envs)
'''
env = PyBulletEnvironment(max_steps, GUI, data_path = data_path, reward_func = reward_func)
env = DummyVecEnv(env)
'''
# Load Training Agent
agent = PPO2.load(model_path, env=env)
# Run Agent on Environment
reward_log = []
for i in range(repeats):
state = env.reset()
deterministic = False
sum_reward = 0
for _ in range(max_steps):
action, _ = agent.predict(state, deterministic=deterministic)
#print(action)
state, reward, done, info = env.step(action)
time.sleep(sleep_time)
sum_reward += reward[0]
reward_log.append(sum_reward / max_steps)
print('Reward Fraction Achieved', np.mean(reward_log))
def train_agent_ppo2(config,
agent_name,
total_timesteps,
policy,
gamma=0.99,
n_steps=128,
ent_coef=0.01,
learning_rate=0.00025,
vf_coef=0.5,
max_grad_norm=0.5,
lam=0.95,
nminibatches=4,
noptepochs=4,
cliprange=0.2,
num_envs=1,
robot_radius=0.46,
rew_fnc=3,
num_stacks=1,
stack_offset=15,
disc_action_space=False,
debug=False,
normalize=False,
stage=0,
pretrained_model_name="",
task_mode="static"):
# Setting seed
seed = random.randint(0, 1000)
np.random.seed(seed)
tf.random.set_random_seed(seed)
random.seed(seed)
# Define pathes to store things
path_to_tensorboard_log = config['PATHES']['path_to_tensorboard_log']
global path_to_models
path_to_models = config['PATHES']['path_to_models']
agent_dir = '%s/%s' % (path_to_models, agent_name)
if not os.path.exists(agent_dir):
os.makedirs(agent_dir)
# Loading simulation environment
env = load_train_env(num_envs, robot_radius, rew_fnc, num_stacks,
stack_offset, debug, task_mode, policy,
disc_action_space, normalize)
if stage == 0:
model = PPO2(eval(policy),
env,
gamma=gamma,
n_steps=n_steps,
ent_coef=ent_coef,
learning_rate=learning_rate,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
lam=lam,
nminibatches=nminibatches,
noptepochs=noptepochs,
cliprange=cliprange,
verbose=1,
tensorboard_log='%s' % (path_to_tensorboard_log))
else:
# Pretrained model is loaded to continue training.
model = PPO2.load(
"%s/%s/%s.pkl" %
(path_to_models, pretrained_model_name, pretrained_model_name),
env,
tensorboard_log='%s' % (path_to_tensorboard_log))
# Document agent
print("Starting PPO2 Training of agent: %s" % (agent_name))
print("------------------------------------------------------")
print("gamma \t\t\t\t %f" % model.gamma)
print("n_steps \t\t\t %d" % model.n_steps)
print("ent_coef \t\t\t %f" % model.ent_coef)
print("learning_rate \t\t\t %f" % learning_rate)
print("vf_coef \t\t\t %f" % model.vf_coef)
print("max_grad_norm \t\t\t %f" % model.max_grad_norm)
print("lam \t\t\t\t %f" % model.lam)
print("nminibatches \t\t\t %d" % model.nminibatches)
print("noptepochs \t\t\t %d" % model.noptepochs)
print("cliprange \t\t\t %f" % cliprange)
print("total_timesteps \t\t %d" % total_timesteps)
print("Policy \t\t\t\t %s" % policy)
print("reward_fnc \t\t\t %d" % rew_fnc)
print("Normalized state: %d" % normalize)
print("discrete action space %d" % disc_action_space)
print("Number of stacks: %d, stack offset: %d" %
(num_stacks, stack_offset))
print("\n")
# Starting training
reset_num_timesteps = False
if stage == 0:
reset_num_timesteps = True
model.learn(total_timesteps=total_timesteps,
log_interval=100,
callback=train_callback,
tb_log_name=agent_name,
reset_num_timesteps=reset_num_timesteps)
# Saving final model
model.save("%s/%s/%s" % (path_to_models, agent_name, "%s_stage_%d" %
(agent_name, stage)))
print("Training finished.")
env.close()
def visualize_policy_and_collect_COM(seed, run_num, policy_env,
policy_num_timesteps, policy_seed,
policy_run_num):
logger.log(sys.argv)
common_arg_parser = get_common_parser()
args, cma_unknown_args = common_arg_parser.parse_known_args()
args.env = policy_env
args.seed = policy_seed
args.num_timesteps = policy_num_timesteps
args.run_num = policy_run_num
this_run_dir = get_dir_path_for_this_run(args)
traj_params_dir_name = get_full_params_dir(this_run_dir)
save_dir = get_save_dir(this_run_dir)
final_file = get_full_param_traj_file_path(traj_params_dir_name,
"pi_final")
final_params = pd.read_csv(final_file, header=None).values[0]
def make_env():
env_out = gym.make(args.env)
env_out.env.disableViewer = False
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
env_out.seed(seed)
return env_out
env = DummyVecEnv([make_env])
if args.normalize:
env = VecNormalize(env)
model = PPO2.load(f"{save_dir}/ppo2", seed=seed)
model.set_pi_from_flat(final_params)
if args.normalize:
env.load_running_average(save_dir)
sk = env.venv.envs[0].env.env.robot_skeleton
lagrangian_values = {}
obs = np.zeros((env.num_envs, ) + env.observation_space.shape)
obs[:] = env.reset()
plot_dir = get_plot_dir(policy_env=args.env,
policy_num_timesteps=policy_num_timesteps,
policy_run_num=policy_run_num,
policy_seed=policy_seed,
eval_seed=seed,
eval_run_num=run_num,
additional_note="")
if os.path.exists(plot_dir):
shutil.rmtree(plot_dir)
os.makedirs(plot_dir)
env = VecVideoRecorder(env,
plot_dir,
record_video_trigger=lambda x: x == 0,
video_length=3000,
name_prefix="3000000agent-{}".format(args.env))
lagrangian_values["M"] = [sk.M.reshape((-1, 1))]
lagrangian_values["COM"] = [sk.C.reshape((-1, 1))]
lagrangian_values["Coriolis"] = [sk.c.reshape((-1, 1))]
lagrangian_values["q"] = [sk.q.reshape((-1, 1))]
lagrangian_values["dq"] = [sk.dq.reshape((-1, 1))]
contact_values = {}
neuron_values = model.give_neuron_values(obs)
raw_layer_values_list = [[neuron_value.reshape((-1, 1))]
for neuron_value in neuron_values]
env.render()
ep_infos = []
steps_to_first_done = 0
first_done = False
# epi_rew = 0
for _ in range(3000):
actions = model.step(obs)[0]
# yield neuron_values
obs, rew, done, infos = env.step(actions)
# epi_rew+= rew[0]
if done and not first_done:
first_done = True
if not first_done:
steps_to_first_done += 1
neuron_values = model.give_neuron_values(obs)
for i, layer in enumerate(neuron_values):
raw_layer_values_list[i].append(layer.reshape((-1, 1)))
# fill_contacts_jac_dict(infos[0]["contacts"], contact_dict=contact_values, neuron_values=neuron_values)
lagrangian_values["M"].append(sk.M.reshape((-1, 1)))
lagrangian_values["q"].append(sk.q.reshape((-1, 1)))
lagrangian_values["dq"].append(sk.dq.reshape((-1, 1)))
lagrangian_values["COM"].append(sk.C.reshape((-1, 1)))
lagrangian_values["Coriolis"].append(sk.c.reshape((-1, 1)))
# env.render()
# time.sleep(1)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
env.render()
done = done.any()
if done:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
# print(f'episode_rew={epi_rew}')
# epi_rew = 0
obs = env.reset()
#Hstack into a big matrix
lagrangian_values["M"] = np.hstack(lagrangian_values["M"])
lagrangian_values["COM"] = np.hstack(lagrangian_values["COM"])
lagrangian_values["Coriolis"] = np.hstack(lagrangian_values["Coriolis"])
lagrangian_values["q"] = np.hstack(lagrangian_values["q"])
lagrangian_values["dq"] = np.hstack(lagrangian_values["dq"])
# for contact_body_name, l in contact_values.items():
# body_contact_dict = contact_values[contact_body_name]
# for name, l in body_contact_dict.items():
# body_contact_dict[name] = np.hstack(body_contact_dict[name])
input_values = np.hstack(raw_layer_values_list[0])
layers_values = [
np.hstack(layer_list) for layer_list in raw_layer_values_list
][1:-2] # drop variance and inputs
for i, com in enumerate(lagrangian_values["COM"]):
plt.figure()
plt.plot(np.arange(len(com)), com)
plt.xlabel("time")
plt.ylabel(f"COM{i}")
plt.savefig(f"{plot_dir}/COM{i}.jpg")
plt.close()
model = PPO2(CustomPolicy,
env,
n_steps=int(2048 / 128),
nminibatches=64,
noptepochs=10,
lam=0.98,
verbose=1,
tensorboard_log='/home/xi/model/log')
# model = PPO2.load("ppo2_ipadgame")
# model.set_env(env)
# model.tensorboard_log='/home/xi/model/log'
# env.load_running_average("/home/xi/model/")
model.learn(total_timesteps=50000)
# model.save("ppo2_ipadgame")
# env.save_running_average("/home/xi/model/")
# print ('done')
env = gym.make(env_id)
env = DummyVecEnv([lambda: env])
env = VecNormalize(env)
obs = env.reset()
model = PPO2.load("ppo2_ipadgame")
env.load_running_average("/home/xi/model/")
for i in range(1000):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
env.render()
def run_experiment_with_trained(augment_num_timesteps, linear_co_threshold, augment_seed, augment_run_num, network_size,
policy_env, policy_num_timesteps, policy_run_num, policy_seed, eval_seed, eval_run_num, learning_rate,
additional_note, result_dir, keys_to_include, metric_param, linear_top_vars_list=None,
linear_correlation_neuron_list=None, visualize=False, lagrangian_inds_to_include=None,
neurons_inds_to_include=None, use_lagrangian=True):
trained_model = None
if not use_lagrangian:
with tf.variable_scope("trained_model"):
common_arg_parser = get_common_parser()
trained_args, cma_unknown_args = common_arg_parser.parse_known_args()
trained_args.env = policy_env
trained_args.seed = policy_seed
trained_args.num_timesteps = policy_num_timesteps
trained_args.run_num = policy_run_num
trained_this_run_dir = get_dir_path_for_this_run(trained_args)
trained_traj_params_dir_name = get_full_params_dir(trained_this_run_dir)
trained_save_dir = get_save_dir(trained_this_run_dir)
trained_final_file = get_full_param_traj_file_path(trained_traj_params_dir_name, "pi_final")
trained_final_params = pd.read_csv(trained_final_file, header=None).values[0]
trained_model = PPO2.load(f"{trained_save_dir}/ppo2", seed=augment_seed)
trained_model.set_pi_from_flat(trained_final_params)
args = AttributeDict()
args.normalize = True
args.num_timesteps = augment_num_timesteps
args.run_num = augment_run_num
args.alg = "ppo2"
args.seed = augment_seed
logger.log(f"#######TRAIN: {args}")
# non_linear_global_dict
timestamp = get_time_stamp('%Y_%m_%d_%H_%M_%S')
experiment_label = f"learning_rate_{learning_rate}timestamp_{timestamp}_augment_num_timesteps{augment_num_timesteps}" \
f"_top_num_to_include{linear_co_threshold.start}_{linear_co_threshold.stop}" \
f"_augment_seed{augment_seed}_augment_run_num{augment_run_num}_network_size{network_size}" \
f"_policy_num_timesteps{policy_num_timesteps}_policy_run_num{policy_run_num}_policy_seed{policy_seed}" \
f"_eval_seed{eval_seed}_eval_run_num{eval_run_num}_additional_note_{additional_note}"
if policy_env == "DartWalker2d-v1":
entry_point = 'gym.envs.dart:DartWalker2dEnv_aug_input'
elif policy_env == "DartHopper-v1":
entry_point = 'gym.envs.dart:DartHopperEnv_aug_input'
elif policy_env == "DartHalfCheetah-v1":
entry_point = 'gym.envs.dart:DartHalfCheetahEnv_aug_input'
elif policy_env == "DartSnake7Link-v1":
entry_point = 'gym.envs.dart:DartSnake7LinkEnv_aug_input'
else:
raise NotImplemented()
this_run_dir = get_experiment_path_for_this_run(entry_point, args.num_timesteps, args.run_num,
args.seed, learning_rate=learning_rate, top_num_to_include=linear_co_threshold,
result_dir=result_dir, network_size=network_size)
full_param_traj_dir_path = get_full_params_dir(this_run_dir)
log_dir = get_log_dir(this_run_dir)
save_dir = get_save_dir(this_run_dir)
create_dir_remove(this_run_dir)
create_dir_remove(full_param_traj_dir_path)
create_dir_remove(save_dir)
create_dir_remove(log_dir)
logger.configure(log_dir)
linear_top_vars_list_wanted_to_print = []
if (use_lagrangian and lagrangian_inds_to_include is None) or (not use_lagrangian and neurons_inds_to_include is None):
# note this is only linear
if linear_top_vars_list is None or linear_correlation_neuron_list is None:
linear_top_vars_list, linear_correlation_neuron_list = read_linear_top_var(policy_env, policy_num_timesteps, policy_run_num, policy_seed, eval_seed,
eval_run_num, additional_note, metric_param=metric_param)
lagrangian_inds_to_include, neurons_inds_to_include, linear_top_vars_list_wanted_to_print = \
get_wanted_lagrangians_and_neurons(keys_to_include, linear_top_vars_list, linear_correlation_neuron_list, linear_co_threshold)
with open(f"{log_dir}/lagrangian_inds_to_include.json", 'w') as fp:
json.dump(lagrangian_inds_to_include, fp)
with open(f"{log_dir}/linear_top_vars_list_wanted_to_print.json", 'w') as fp:
json.dump(linear_top_vars_list_wanted_to_print, fp)
with open(f"{log_dir}/neurons_inds_to_include.json", 'w') as fp:
json.dump(neurons_inds_to_include, fp)
args.env = f'{experiment_label}_{entry_point}-v1'
if not use_lagrangian:
register(
id=args.env,
entry_point=entry_point,
max_episode_steps=1000,
kwargs={"lagrangian_inds_to_include": None, "trained_model": trained_model,
"neurons_inds_to_include": neurons_inds_to_include}
)
else:
register(
id=args.env,
entry_point=entry_point,
max_episode_steps=1000,
kwargs={"lagrangian_inds_to_include": lagrangian_inds_to_include, "trained_model": None,
"neurons_inds_to_include": None}
)
def make_env():
env_out = gym.make(args.env)
env_out.env.visualize = visualize
env_out = bench.Monitor(env_out, logger.get_dir(), allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
walker_env = env.envs[0].env.env
walker_env.disableViewer = not visualize
if args.normalize:
env = VecNormalize(env)
policy = MlpPolicy
set_global_seeds(args.seed)
walker_env.seed(args.seed)
num_dof = walker_env.robot_skeleton.ndofs
show_M_matrix(num_dof, lagrangian_inds_to_include, linear_co_threshold, log_dir)
# extra run info I added for my purposes
run_info = {"run_num": args.run_num,
"env_id": args.env,
"full_param_traj_dir_path": full_param_traj_dir_path}
layers = [network_size, network_size]
policy_kwargs = {"net_arch" : [dict(vf=layers, pi=layers)]}
model = PPO2(policy=policy, env=env, n_steps=4096, nminibatches=64, lam=0.95, gamma=0.99,
noptepochs=10,
ent_coef=0.0, learning_rate=learning_rate, cliprange=0.2, optimizer='adam', policy_kwargs=policy_kwargs,
seed=args.seed)
model.tell_run_info(run_info)
model.learn(total_timesteps=args.num_timesteps, seed=args.seed)
model.save(f"{save_dir}/ppo2")
if args.normalize:
env.save_running_average(save_dir)
return log_dir
qpos[i, :] = env_in.GetMotorAngles()
# qvel[i,:] = env_in.sim.data.qvel[[7,8,10,12,14,16,17,19,21,23]]
qvel[i, :] = env_in.GetMotorVelocities()
# torque[i,:] = env_in.sim.data.actuator_force
torque[i, :] = env_in.GetMotorTorques()
i = (i + 1) % total_data_length
return True
###############################################################################
# # Use this code for testing the basic controller
# Create the stoch mujoco environment
# env = stoch2_gym_mjc_env.Stoch2Env()
env = vision60_gym_bullet_env.Vision60BulletEnv(render=True)
model_test = PPO2.load(dir_name + "/model_trot")
obs = env.reset()
print("Render mode...")
for _ in range(10):
action, _states = model_test.predict(obs, deterministic=True)
obs, reward, done, _ = env.step(action, callback=render_callback)
# if done:
# break
pickle.dump(qpos[0:total_data_length:int(total_data_length / 100)],
open("save.p", "wb")) # save it into a file named save.p
# print(np.shape(qpos[0:total_data_length:int(total_data_length/100)]))
# print(np.shape(qpos))
def main(env,
load,
save_path,
load_path=None,
train_timesteps=1.25e6,
eval_timesteps=5e3):
# arguments
print(
"env %s; load %s; save_path %s; load_path %s; train_timesteps %s; eval_timesteps %s;"
% (env, load, save_path, load_path, train_timesteps, eval_timesteps))
train_timesteps = int(float(train_timesteps))
eval_timesteps = int(float(eval_timesteps))
# models path
model_dir = os.getcwd() + "/models/"
os.makedirs(model_dir, exist_ok=True)
# logging path
log_dir = os.getcwd() + "/log/" + save_path
os.makedirs(log_dir, exist_ok=True)
# absolute save path and models path
save_path = model_dir + save_path
if load and not load_path:
print("no load path given, exiting...")
sys.exit()
elif load:
load_path = model_dir + load_path
# make environment, flattened environment, monitor, vectorized environment
env = gym.make(env)
env = gym.wrappers.FlattenDictWrapper(
env, ['observation', 'achieved_goal', 'desired_goal'])
env = Monitor(env, log_dir, allow_early_resets=True)
env = DummyVecEnv([lambda: env])
# load model, or start from scratch
if load:
print("loading model from: " + load_path)
model = PPO2.load(load_path, env=env)
else:
print("training model from scratch")
model = PPO2(MlpPolicy, env, verbose=1)
# evaluate current model
mean_reward_before_train = evaluate(model, env, num_steps=eval_timesteps)
# train model
global best_mean_reward, n_steps
best_mean_reward, n_steps = -np.inf, 0
model.learn(total_timesteps=train_timesteps, callback=None)
# save model
print("saving model to:" + save_path)
model.save(save_path)
# evaluate post training model
mean_reward_after_train = evaluate(model, env, num_steps=eval_timesteps)
# results
print("reward before training:" + str(mean_reward_before_train))
print("reward after training:" + str(mean_reward_after_train))
print("done")
env.close()
# Curriculum implemented is extremely basic one. (Need to improvise later on)
# Curriculum is a newly made folder. In google drive, read the note
all_arenas = sorted(glob.glob('configs/Curriculum/*.yaml'))
print(all_arenas)
model_name = 'ppo_model_after_bc'
all_frames_vec = [200000, 200000, 100000, 200000, 200000, 200000, 200000, 200000] #no of tsteps per curriculum
for i in range(len(all_arenas)):
# create arena
env = create_env_fn(num_actors = 1, inference=False, config=all_arenas[i], seed=0)
env = make_vec_env(env, n_envs=4)
print('####################')
print("## Curriculum {} ##".format(i))
print('####################')
model = PPO2.load(model_name, env)
frames_idx = all_frames_vec[i]
print('{} arena is used for training for {} timesteps'.format(all_arenas[i], frames_idx))
model.learn(total_timesteps=frames_idx)
model_name = "ppo_model_after_training_arena_{}".format(i)
model.save(model_name)
env.close()
del model
del env
print('Training complete!!')
def eval_trained_policy_and_collect_data(eval_seed, eval_run_num, policy_env, policy_num_timesteps, policy_seed, policy_run_num, additional_note):
logger.log(sys.argv)
common_arg_parser = get_common_parser()
args, cma_unknown_args = common_arg_parser.parse_known_args()
args.env = policy_env
args.seed = policy_seed
args.num_timesteps = policy_num_timesteps
args.run_num = policy_run_num
this_run_dir = get_dir_path_for_this_run(args)
traj_params_dir_name = get_full_params_dir(this_run_dir)
save_dir = get_save_dir( this_run_dir)
final_file = get_full_param_traj_file_path(traj_params_dir_name, "pi_final")
final_params = pd.read_csv(final_file, header=None).values[0]
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out, logger.get_dir(), allow_early_resets=True)
env_out.seed(eval_seed)
return env_out
env = DummyVecEnv([make_env])
running_env = env.envs[0].env.env
set_global_seeds(eval_seed)
running_env.seed(eval_seed)
if args.normalize:
env = VecNormalize(env)
model = PPO2.load(f"{save_dir}/ppo2", seed=eval_seed)
model.set_pi_from_flat(final_params)
if args.normalize:
env.load_running_average(save_dir)
# is it necessary?
running_env = env.venv.envs[0].env.env
lagrangian_values = {}
obs = np.zeros((env.num_envs,) + env.observation_space.shape)
obs[:] = env.reset()
# env = VecVideoRecorder(env, "./",
# record_video_trigger=lambda x: x == 0, video_length=3000,
# name_prefix="3000000agent-{}".format(args.env))
#init lagrangian values
for lagrangian_key in lagrangian_keys:
flat_array = running_env.get_lagrangian_flat_array(lagrangian_key)
lagrangian_values[lagrangian_key] = [flat_array]
neuron_values = model.give_neuron_values(obs)
raw_layer_values_list = [[neuron_value.reshape((-1,1))] for neuron_value in neuron_values]
# env.render()
ep_infos = []
steps_to_first_done = 0
first_done = False
for _ in range(30000):
actions = model.step(obs)[0]
# yield neuron_values
obs, rew, done, infos = env.step(actions)
if done and not first_done:
first_done = True
if not first_done:
steps_to_first_done += 1
neuron_values = model.give_neuron_values(obs)
for i, layer in enumerate(neuron_values):
raw_layer_values_list[i].append(layer.reshape((-1,1)))
# fill_contacts_jac_dict(infos[0]["contacts"], contact_dict=contact_values, neuron_values=neuron_values)
# filling lagrangian values
for lagrangian_key in lagrangian_keys:
flat_array = running_env.get_lagrangian_flat_array(lagrangian_key)
lagrangian_values[lagrangian_key].append(flat_array)
# env.render()
# time.sleep(1)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
# env.render()
done = done.any()
if done:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
obs = env.reset()
#Hstack into a big matrix
for lagrangian_key in lagrangian_keys:
lagrangian_values[lagrangian_key] = np.hstack(lagrangian_values[lagrangian_key])
# for contact_body_name, l in contact_values.items():
# body_contact_dict = contact_values[contact_body_name]
# for name, l in body_contact_dict.items():
# body_contact_dict[name] = np.hstack(body_contact_dict[name])
input_values = np.hstack(raw_layer_values_list[0])
layers_values = [np.hstack(layer_list) for layer_list in raw_layer_values_list][1:-2]# drop variance and inputs
data_dir = get_data_dir(policy_env=args.env, policy_num_timesteps=policy_num_timesteps, policy_run_num=policy_run_num
, policy_seed=policy_seed, eval_seed=eval_seed, eval_run_num=eval_run_num, additional_note=additional_note)
if os.path.exists(data_dir):
shutil.rmtree(data_dir)
os.makedirs(data_dir)
lagrangian_values_fn = f"{data_dir}/lagrangian.pickle"
with open(lagrangian_values_fn, 'wb') as handle:
pickle.dump(lagrangian_values, handle, protocol=pickle.HIGHEST_PROTOCOL)
input_values_fn = f"{data_dir}/input_values.npy"
layers_values_fn = f"{data_dir}/layer_values.npy"
np.save(input_values_fn, input_values)
np.save(layers_values_fn, layers_values)
all_weights = model.get_all_weight_values()
for ind, weights in enumerate(all_weights):
fname = f"{data_dir}/weights_layer_{ind}.txt"
np.savetxt(fname, weights)
def main():
import sys
logger.log(sys.argv)
common_arg_parser = get_common_parser()
args, cma_unknown_args = common_arg_parser.parse_known_args()
this_run_dir = get_dir_path_for_this_run(args)
plot_dir_alg = get_plot_dir(args)
traj_params_dir_name = get_full_params_dir(this_run_dir)
save_dir = get_save_dir(this_run_dir)
if not os.path.exists(plot_dir_alg):
os.makedirs(plot_dir_alg)
final_file = get_full_param_traj_file_path(traj_params_dir_name,
"pi_final")
final_params = pd.read_csv(final_file, header=None).values[0]
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
env = DummyVecEnv([make_env])
# env_out = gym.make(args.env)
# env_out = bench.Monitor(env_out, logger.get_dir(), allow_early_resets=True)
if args.normalize:
env = VecNormalize(env)
# policy = MlpPolicy
model = PPO2.load(f"{save_dir}/ppo2") # this also loads V function
# model = PPO2(policy=policy, env=env, n_steps=args.n_steps, nminibatches=args.nminibatches, lam=0.95, gamma=0.99, noptepochs=10,
# ent_coef=0.0, learning_rate=3e-4, cliprange=0.2, optimizer=args.optimizer)
model.set_pi_from_flat(final_params)
if args.normalize:
env.load_running_average(save_dir)
sk = env.venv.envs[0].env.env.robot_skeleton
lagrangian_values = {}
obs = np.zeros((env.num_envs, ) + env.observation_space.shape)
obs[:] = env.reset()
# env = VecVideoRecorder(env, "./",
# record_video_trigger=lambda x: x == 0, video_length=3000,
# name_prefix="3000000agent-{}".format(args.env))
lagrangian_values["M"] = [sk.M.reshape((-1, 1))]
lagrangian_values["COM"] = [sk.C.reshape((-1, 1))]
lagrangian_values["Coriolis"] = [sk.c.reshape((-1, 1))]
lagrangian_values["q"] = [sk.q.reshape((-1, 1))]
lagrangian_values["dq"] = [sk.dq.reshape((-1, 1))]
contact_values = {}
neuron_values = model.give_neuron_values(obs)
layer_values_list = [[neuron_value.reshape((-1, 1))]
for neuron_value in neuron_values]
env.render()
ep_infos = []
steps_to_first_done = 0
first_done = False
for _ in range(3000):
actions = model.step(obs)[0]
# yield neuron_values
obs, rew, done, infos = env.step(actions)
if done and not first_done:
first_done = True
if not first_done:
steps_to_first_done += 1
neuron_values = model.give_neuron_values(obs)
for i, layer in enumerate(neuron_values):
layer_values_list[i].append(layer.reshape((-1, 1)))
fill_contacts_jac_dict(infos[0]["contacts"],
contact_dict=contact_values,
neuron_values=neuron_values)
lagrangian_values["M"].append(sk.M.reshape((-1, 1)))
lagrangian_values["q"].append(sk.q.reshape((-1, 1)))
lagrangian_values["dq"].append(sk.dq.reshape((-1, 1)))
lagrangian_values["COM"].append(sk.C.reshape((-1, 1)))
lagrangian_values["Coriolis"].append(sk.c.reshape((-1, 1)))
env.render()
# time.sleep(1)
for info in infos:
maybe_ep_info = info.get('episode')
if maybe_ep_info is not None:
ep_infos.append(maybe_ep_info)
# env.render()
done = done.any()
if done:
episode_rew = safe_mean([ep_info['r'] for ep_info in ep_infos])
print(f'episode_rew={episode_rew}')
obs = env.reset()
#Hstack into a big matrix
lagrangian_values["M"] = np.hstack(lagrangian_values["M"])
lagrangian_values["COM"] = np.hstack(lagrangian_values["COM"])
lagrangian_values["Coriolis"] = np.hstack(lagrangian_values["Coriolis"])
lagrangian_values["q"] = np.hstack(lagrangian_values["q"])
lagrangian_values["dq"] = np.hstack(lagrangian_values["dq"])
for contact_body_name, l in contact_values.items():
body_contact_dict = contact_values[contact_body_name]
for name, l in body_contact_dict.items():
body_contact_dict[name] = np.hstack(body_contact_dict[name])
layer_values_list = [
np.hstack(layer_list) for layer_list in layer_values_list
][1:-2] # drop variance
# plt.scatter(lagrangian_values["M"][15], layer_values_list[1][2])
# plt.scatter(lagrangian_values["M"][11], layer_values_list[0][63])
out_dir = f"/home/panda-linux/PycharmProjects/low_dim_update_dart/low_dim_update_stable/neuron_vis/plots_{args.env}_{args.num_timesteps}"
if os.path.exists(out_dir):
shutil.rmtree(out_dir)
os.makedirs(out_dir)
all_weights = model.get_all_weight_values()
for ind, weights in enumerate(all_weights):
fname = f"{out_dir}/weights_layer_{ind}.txt"
np.savetxt(fname, weights)
PLOT_CUTOFF = steps_to_first_done
plot_everything(lagrangian_values, layer_values_list, out_dir, PLOT_CUTOFF)
scatter_the_linear_significant_ones(lagrangian_values,
layer_values_list,
threshold=0.6,
out_dir=out_dir)
scatter_the_nonlinear_significant_but_not_linear_ones(
lagrangian_values,
layer_values_list,
linear_threshold=0.3,
nonlinear_threshold=0.6,
out_dir=out_dir)
#
# contact_dicts = {}
# for contact_body_name, l in contact_values.items():
# body_contact_dict = contact_values[contact_body_name]
#
#
# contact_dicts[contact_body_name] = {}
#
# build_dict = contact_dicts[contact_body_name]
#
# build_dict["body"] = {}
# build_dict["layer"] = {}
# for name, l in body_contact_dict.items():
# for i in range(len(l)):
#
# if name == contact_body_name:
# build_dict["body"][f"{contact_body_name}_{i}"] = l[i]
# else:
# build_dict["layer"][f"layer_{name}_neuron_{i}"] = l[i]
#
# body_contact_df = pd.DataFrame.from_dict(build_dict["body"], "index")
# layer_contact_df = pd.DataFrame.from_dict(build_dict["layer"], "index")
# body_contact_df.to_csv(f"{data_dir}/{contact_body_name}_contact.txt", sep='\t')
# layer_contact_df.to_csv(f"{data_dir}/{contact_body_name}_layers.txt", sep='\t')
# #TO CSV format
# data_dir = f"/home/panda-linux/PycharmProjects/low_dim_update_dart/mictools/examples/neuron_vis_data{args.env}_time_steps_{args.num_timesteps}"
# if os.path.exists(data_dir):
# shutil.rmtree(data_dir)
#
# os.makedirs(data_dir)
#
# for contact_body_name, d in contact_dicts.items():
#
# build_dict = d
#
# body_contact_df = pd.DataFrame.from_dict(build_dict["body"], "index")
# layer_contact_df = pd.DataFrame.from_dict(build_dict["layer"], "index")
#
# body_contact_df.to_csv(f"{data_dir}/{contact_body_name}_contact.txt", sep='\t')
# layer_contact_df.to_csv(f"{data_dir}/{contact_body_name}_layers.txt", sep='\t')
#
#
#
# neurons_dict = {}
# for layer_index in range(len(layer_values_list)):
# for neuron_index in range(len(layer_values_list[layer_index])):
# neurons_dict[f"layer_{layer_index}_neuron_{neuron_index}"] = layer_values_list[layer_index][neuron_index]
#
# for i in range(len(lagrangian_values["COM"])):
# neurons_dict[f"COM_index_{i}"] = lagrangian_values["COM"][i]
#
# neuron_df = pd.DataFrame.from_dict(neurons_dict, "index")
#
#
#
# lagrangian_dict = {}
# for k,v in lagrangian_values.items():
# for i in range(len(v)):
# lagrangian_dict[f"{k}_index_{i}"] = v[i]
#
# lagrangian_df = pd.DataFrame.from_dict(lagrangian_dict, "index")
#
#
# neuron_df.to_csv(f"{data_dir}/neurons.txt", sep='\t')
# lagrangian_df.to_csv(f"{data_dir}/lagrangian.txt", sep='\t')
# cor = {}
# best_cor = {}
# cor["M"] = get_correlations(lagrangian_values["M"], layer_values_list)
# best_cor["M"] = [np.max(np.abs(cor_m)) for cor_m in cor["M"]]
#
#
# cor["COM"] = get_correlations(lagrangian_values["COM"], layer_values_list)
# best_cor["COM"] = [np.max(np.abs(cor_m)) for cor_m in cor["COM"]]
#
# cor["Coriolis"] = get_correlations(lagrangian_values["Coriolis"], layer_values_list)
# best_cor["Coriolis"] = [np.max(np.abs(cor_m)) for cor_m in cor["Coriolis"]]
# best_cor["Coriolis_argmax"] = [np.argmax(np.abs(cor_m)) for cor_m in cor["Coriolis"]]
#
#
#
#
# ncor = {}
# nbest_cor = {}
# ncor["M"] = get_normalized_correlations(lagrangian_values["M"], layer_values_list)
# nbest_cor["M"] = [np.max(np.abs(cor_m)) for cor_m in ncor["M"]]
#
#
# ncor["COM"] = get_normalized_correlations(lagrangian_values["COM"], layer_values_list)
# nbest_cor["COM"] = [np.max(np.abs(cor_m)) for cor_m in ncor["COM"]]
#
# ncor["Coriolis"] = get_normalized_correlations(lagrangian_values["Coriolis"], layer_values_list)
# nbest_cor["Coriolis"] = [np.max(np.abs(cor_m)) for cor_m in ncor["Coriolis"]]
# nbest_cor["Coriolis_argmax"] = [np.argmax(np.abs(cor_m)) for cor_m in ncor["Coriolis"]]
#
#
#
#
#
# lin_reg = {"perm_1":{}, "perm_2":{}}
# best_lin_reg = {"perm_1":{}, "perm_2":{}}
# lin_reg["perm_1"]["M"], best_lin_reg["perm_1"]["M"] = get_results("M", lagrangian_values, layer_values_list, perm_num=1)
# lin_reg["perm_2"]["M"], best_lin_reg["perm_2"]["M"] = get_results("M", lagrangian_values, layer_values_list, perm_num=2)
# lin_reg["perm_1"]["COM"], best_lin_reg["perm_1"]["COM"] = get_results("COM", lagrangian_values, layer_values_list, perm_num=1)
# lin_reg["perm_2"]["COM"], best_lin_reg["perm_2"]["COM"] = get_results("COM", lagrangian_values, layer_values_list, perm_num=2)
#
#
# lin_reg_1["M"] = get_linear_regressions_1_perm(lagrangian_values["M"], layer_values_list)
# lin_reg_2["M"] = get_linear_regressions_2_perm(lagrangian_values["M"], layer_values_list)
# best_lin_reg_2["M"] = []
# for lin_l in lin_reg_2["M"]:
# if lin_l == []:
# best_lin_reg_2["M"].append([])
# else:
# best_lin_reg_2["M"].append(lin_l[np.argmin(lin_l[:,0])])
#
# best_lin_reg_1["M"] = []
# for lin_l in lin_reg_1["M"]:
# if lin_l == []:
# best_lin_reg_1["M"].append([])
# else:
# best_lin_reg_1["M"].append(lin_l[np.argmin(lin_l[:,0])])
# best_lin_reg_1["M"] = np.array(best_lin_reg_1["M"])
# best_lin_reg_2["M"] = np.array(best_lin_reg_2["M"])
#
#
# lin_reg_1["M"].dump("lin_reg_1_M.txt")
# lin_reg_2["M"].dump("lin_reg_2_M.txt")
# best_lin_reg_1["M"].dump("best_lin_reg_1_M.txt")
# best_lin_reg_2["M"].dump("best_lin_reg_2_M.txt")
#
# lin_reg_1["COM"] = get_linear_regressions_1_perm(lagrangian_values["COM"], layer_values_list)
# lin_reg_2["COM"] = get_linear_regressions_2_perm(lagrangian_values["COM"], layer_values_list)
# best_lin_reg_2["COM"] = []
# for lin_l in lin_reg_2["COM"]:
# if lin_l == []:
# best_lin_reg_2["COM"].append([])
# else:
# best_lin_reg_2["COM"].append(lin_l[np.argmin(lin_l[:, 0])])
#
# best_lin_reg_1["COM"] = []
# for lin_l in lin_reg_1["COM"]:
# if lin_l == []:
# best_lin_reg_1["COM"].append([])
# else:
# best_lin_reg_1["COM"].append(lin_l[np.argmin(lin_l[:, 0])])
#
#
# best_lin_reg_1["COM"] = np.array(best_lin_reg_1["M"])
# best_lin_reg_2["COM"] = np.array(best_lin_reg_2["M"])
# lin_reg_1["COM"].dump("lin_reg_1_COM.txt")
# lin_reg_2["COM"].dump("lin_reg_2_COM.txt")
# best_lin_reg_1["COM"].dump("best_lin_reg_1_COM.txt")
# best_lin_reg_2["COM"].dump("best_lin_reg_2_COM.txt")
pass
def load_model(load_dir=LOAD_DIR):
return PPO2.load(load_dir)
