def test_keras_pipeline():
class MyEstimator(KerasEstimator):
def build(self):
self.model = tf.keras.models.Sequential()
self.model.add(
tf.keras.layers.Dense(self.observation_size,
activation='relu',
input_shape=self.observation_shape))
self.model.add(tf.keras.layers.Dense(self.action_size))
self.model.compile(
tf.keras.optimizers.Adam(learning_rate=self.learning_rate),
loss='mse')
def preprocess(self, observations, actions):
return observations
env = CatchEnv()
custom_action_value = MyEstimator(
observation_space=env.observation_space,
action_space=env.action_space,
batch_size=64,
freezed_steps=20)
agent = StandardAgent(observation_space=env.observation_space,
action_space=env.action_space,
action_values=custom_action_value)
pg = rl.Playground(env, agent)
pg.fit(1)
import learnrl as rl
import tensorflow as tf
import retro
from agents.a2c import ActorCriticAgent
env = retro.make(game='Airstriker-Genesis')
agent = ActorCriticAgent(env.action_space)
print(env.observation_space, env.action_space)
reward_handler = rl.RewardHandler
pg = rl.Playground(env, agent)
pg.fit(
100,
render=True,
verbose=2,
titles_on_top=False,
)
pg.test(1, titles_on_top=False)