def train():
hp = Hp()
np.random.seed(hp.seed)
env = OpenAIGym(hp.env_name)
nb_inputs = env.observation_space.shape[0]
# if using cnn for inputs
if hp.conv_input:
# create dummy image
test_img = np.ones([env.observation_space.shape[0], env.observation_space.shape[1], env.observation_space.shape[2]]).astype(np.uint8)
# create dummy Normalizer obj
test_n = Normalizer(0, hp)
# pass through cnn
test_output = test_n.image_cnn(test_img)
# get output length
nb_inputs = len(test_output)
nb_outputs = env.action_space.shape[0]
policy = Policy(nb_inputs, nb_outputs, hp)
normalizer = Normalizer(nb_inputs, hp)
if hp.train_from_previous_weights:
policy.load()
normalizer.load()
instance = Run(env, policy, normalizer, hp)
instance.train()
def evaluate(n_steps):
hp = Hp()
np.random.seed(hp.seed)
env = OpenAIGym(hp.env_name)
nb_inputs = env.observation_space.shape[0]
if hp.conv_input:
# create dummy image
test_img = np.ones([
env.observation_space.shape[0], env.observation_space.shape[1],
env.observation_space.shape[2]
]).astype(np.uint8)
# create dummy Normalizer obj
test_n = Normalizer(0, hp)
# pass through cnn
test_output = test_n.image_cnn(test_img)
# get output length
nb_inputs = len(test_output)
nb_outputs = env.action_space.shape[0]
policy = Policy(nb_inputs, nb_outputs, hp)
normalizer = Normalizer(nb_inputs, hp)
normalizer.load() # load normalizer weights
policy.load() # load policy weights
instance = Run(env, policy, normalizer, hp)
instance.evaluate(n_steps)