def _make_trainer(self, train_env, eval_env, output_dir, model=None):
if model is None:
model = lambda: [layers.Dense(1)]
return ppo_trainer.PPO(
train_env=train_env,
eval_env=eval_env,
policy_and_value_model=model,
n_optimizer_steps=1,
output_dir=output_dir,
random_seed=0,
boundary=2,
)
def _run_training_loop(self, train_env, eval_env, output_dir, model=None):
if model is None:
model = lambda: [layers.Dense(1)]
n_epochs = 2
# Run the training loop.
trainer = ppo_trainer.PPO(
train_env=train_env,
eval_env=eval_env,
policy_and_value_model=model,
n_optimizer_steps=1,
output_dir=output_dir,
random_seed=0,
boundary=2,
)
trainer.training_loop(n_epochs=n_epochs)
def _make_trainer(self,
train_env,
eval_env,
output_dir,
model=None,
**kwargs):
if model is None:
model = lambda: layers.Serial(layers.Dense(1))
return ppo_trainer.PPO(train_env=train_env,
eval_env=eval_env,
policy_and_value_model=model,
n_optimizer_steps=1,
output_dir=output_dir,
random_seed=0,
max_timestep=3,
boundary=2,
save_every_n=1,
**kwargs)
