],
activation_functions=[relu, relu]))
actor.apply(gauss_init(0, weight_init_sigma))
critic.apply(gauss_init(0, weight_init_sigma))
# Training
trainer = DDPGTrainer(env=env,
actor=actor,
critic=critic,
tau=tau,
epsilon=epsilon,
batch_size=batch_size,
depsilon=depsilon,
gamma=gamma,
lr_actor=actor_learning_rate,
lr_critic=critic_learning_rate,
warmup=warmup,
replay_memory=replay_memory)
checkpoint_callback = CheckpointCallback(save_path=config.root_path(),
models={
"actor": actor,
"critic": critic
})
trainer.train(2000,
max_episode_len=500,
verbose=True,
callbacks=[checkpoint_callback])
import roboschool
env_name = 'RoboschoolAnt-v1'
# Interpolation parameter v * ppo_gradient + (1-v) * off_policy_gradient
v = 0.5
np.random.seed(456)
tor.manual_seed(456)
config.set_root('torch_rl_ipgppo_' + env_name.lower().split("-")[0] +
"_v={}".format(v),
force=True)
config.configure_logging(clear=False, output_formats=['tensorboard', 'stdout'])
# config.start_tensorboard()
monitor = Monitor(EnvLogger(NormalisedActionsWrapper(gym.make(env_name))),
directory=os.path.join(config.root_path(), 'stats'),
force=True,
video_callable=False,
write_upon_reset=True)
env = RunningMeanStdNormalize(monitor)
num_observations = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print('Action shape: ', num_actions, 'Observation shape: ', num_observations)
tanh, relu = tor.nn.Tanh(), tor.nn.ReLU()
replay_memory = GeneralisedMemory(1000000)
tt = to_tensor
def save_params(**kwargs):
import json
dir = root_path()
with open(os.path.join(dir, 'params.json'), 'w') as f:
json.dump(kwargs, f, indent=4)