def __init__(self, **kwargs):
super().__init__(**kwargs)
self._layers = [
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(64, name='fc1'),
ReLU(name='relu1'),
tf.keras.layers.Dense(64, name='fc2'),
ReLU(name='relu2'),
]
def __init__(self, **kwargs):
super().__init__(**kwargs)
self._concat = tf.keras.layers.Concatenate()
self._layers = [
tf.keras.layers.Dense(400, name='fc1'),
ReLU(name='relu1'),
tf.keras.layers.Dense(300, name='fc2'),
ReLU(name='relu2'),
tf.keras.layers.Dense(1, name='fc3')
]
def __init__(self, action_space, **kwargs):
super().__init__(**kwargs)
assert isinstance(action_space, gym.spaces.Box)
self.action_space = action_space
self._layers = [
tf.keras.layers.Dense(400, name='fc1'),
ReLU(name='relu1'),
tf.keras.layers.Dense(300, name='fc2'),
ReLU(name='relu2'),
tf.keras.layers.Dense(action_space.shape[0], name='fc3'),
tf.keras.layers.Activation(activation='tanh', name='tanh')
]
def __init__(self, **kwargs):
'''Nature CNN originated from
"Playing Atari with Deep Reinforcement Learning"
'''
super().__init__(**kwargs)
self._layers = [
tf.keras.layers.Conv2D(32, 8, 4, name='conv1'),
ReLU(name='relu1'),
tf.keras.layers.Conv2D(64, 4, 2, name='conv2'),
ReLU(name='relu2'),
tf.keras.layers.Conv2D(64, 3, 1, name='conv3'),
ReLU(name='relu3'),
tf.keras.layers.Flatten(name='flatten'),
tf.keras.layers.Dense(512, name='fc'),
ReLU(name='relu4')
]
def __init__(self, **kwargs):
super().__init__(**kwargs)
self._h_layers = [
tf.keras.layers.Conv2D(32, 8, 4, name='conv1'),
ReLU(name='relu1'),
tf.keras.layers.Conv2D(64, 4, 2, name='conv2'),
ReLU(name='relu2'),
tf.keras.layers.Conv2D(64, 3, 1, name='conv3'),
ReLU(name='relu3'),
tf.keras.layers.Flatten()
]
self._t_layers = [
tf.keras.layers.Dense(512, name='fc'),
ReLU(name='relu4')
]
def __init__(self, latent_size, embed_size=64, **kwargs):
'''Phi net (quantile distortion function)
Args:
latent_size (int): The size of latent vectors from
feature extraction nets.
embed_size (int, optional): The embedding dimensions
(number of cosine samples) Defaults to 64.
'''
super().__init__(**kwargs)
self.embed_size = embed_size
self._layers = [tf.keras.layers.Dense(latent_size), ReLU(name='relu1')]