def run_exp(inputs):
seed, use_goalgan = inputs
env = PandaEnv(seed=seed)
name = f"{'goalgan-' if use_goalgan else ''}her-sac"
agent = HERSACAgent(env=env, rank=seed, experiment_name=name)
if use_goalgan:
agent = GoalGANAgent(env=env, agent=agent)
cb = EvaluateCallback(agent=agent, eval_env=PandaEnv(seed=seed), rank=seed)
agent.train(timesteps=int(300000), callbacks=[cb])
def continuous_viz():
env = PandaEnv(visualize=True)
agent = HERSACAgent(env=env, experiment_name="goalgan-her-sac")
obs = env.reset()
while True:
action = agent(obs)
obs, _, done, info = env.step(action)
if done:
obs = env.reset(reset_agent_pos=not info["is_success"])
def main(task: str, use_gan: bool, do_train: bool, perform_eval: bool):
env_params = {"visualize": not do_train}
env_fn = PandaEnv if task == Task.REACH else PandaPickAndPlace
env = env_fn(**env_params)
agent_params = {"env": env}
if use_gan:
agent_params["experiment_name"] = "goalgan-her-sac"
agent = HERSACAgent(**agent_params)
if use_gan:
agent = GoalGANAgent(env=env, agent=agent)
if do_train:
cbs = [EvaluateCallback(agent=agent, eval_env=env_fn(**env_params))] if perform_eval else []
agent.train(timesteps=50000, callbacks=cbs)
else:
while True:
consume(trajectory(agent, env))
def test_class_instantiation():
env = ToyLab()
a1 = PPOAgent(env=env)
a2 = HERSACAgent(env=env)
a3 = GoalGANAgent(env=env, agent=a1)
for a in [a1, a2, a3]:
consume(trajectory(pi=a, env=env))
cb = EvaluateCallback(agent=a1, eval_env=env)
def main(seed: int):
env = ToyLab(seed=seed)
# π = PPOAgent(env=env, experiment_name="goalgan-ppo-seed-{}".format(seed), rank=seed)
π = HERSACAgent(env=env,
experiment_name="goalgan-her-sac-{}".format(seed),
rank=seed)
agent = GoalGANAgent(env=env, agent=π)
callback = EvaluateCallback(agent=agent,
eval_env=ToyLab(seed=seed),
rank=seed)
agent.train(timesteps=int(1e6), callbacks=[callback])