def __init__(self, env, scope=None, **kwargs):
self.obs_shape = env.obs_shape
*other, self.image_height, self.image_width, self.image_depth = self.obs_shape
self.n_frames = other[0] if other else 0
self.network = cfg.build_network(env, self, scope="network")
super(Updater, self).__init__(env, scope=scope, **kwargs)
class Updater:
pass
updater = Updater()
sess = tf.Session()
with config:
with sess.as_default():
if hasattr(cfg, 'prepare_func'):
cfg.prepare_func()
network = cfg.build_network(env, updater, scope="network")
inputs = dict(image=tf.placeholder(tf.float32, (None, *image_shape)), )
network_outputs = network(inputs, is_training=False)
network_tensors = network_outputs["tensors"]
# maybe load weights
if load_path:
# variables = {v.name: v for v in trainable_variables("", for_opt=False)}
# saver = tf.train.Saver(variables)
saver = tf.train.Saver()
saver.restore(tf.get_default_session(),
os.path.realpath(load_path))
tf.train.get_or_create_global_step()