def __init__(self, vector_dim, output_shape, name, hidden_units, *, visual_net):
super().__init__(visual_net, name=name)
self.share = mlp(hidden_units['share'], out_layer=False)
self.mu = mlp(hidden_units['mu'], output_shape=output_shape, out_activation=None)
self.log_std = mlp(hidden_units['log_std'], output_shape=output_shape, out_activation='tanh')
self.init_or_run(tf.keras.Input(shape=vector_dim + self.visual_net.hdim))
self.update_vars()
def __init__(self, vector_dim, action_dim, name, hidden_units, *, visual_net):
assert len(hidden_units) > 1, "if you want to use this architecture of critic network, the number of layers must greater than 1"
super().__init__(visual_net, name=name)
self.feature_net = mlp(hidden_units[0:1])
self.net = mlp(hidden_units[1:], output_shape=1, out_activation=None)
self.init_or_run(tf.keras.Input(shape=vector_dim + self.visual_net.hdim), tf.keras.Input(shape=action_dim))
self.update_vars()
def __init__(self, vector_dim, action_dim, quantiles_idx, name, hidden_units, *, visual_net):
super().__init__(visual_net, name=name)
self.action_dim = action_dim
self.q_net_head = mlp(hidden_units['q_net'], out_layer=False) # [B, vector_dim]
self.quantile_net = mlp(hidden_units['quantile'], out_layer=False) # [N*B, quantiles_idx]
self.q_net_tile = mlp(hidden_units['tile'], output_shape=action_dim, out_activation=None) # [N*B, hidden_units['quantile'][-1]]
self.init_or_run(tf.keras.Input(shape=vector_dim + self.visual_net.hdim), tf.keras.Input(shape=quantiles_idx))
self.update_vars()
def __init__(self, vector_dim, action_dim, atoms, name, hidden_units, *, visual_net):
super().__init__(visual_net, name=name)
self.action_dim = action_dim
self.atoms = atoms
self.share = mlp(hidden_units['share'], layer=Noisy, out_layer=False)
self.v = mlp(hidden_units['v'], layer=Noisy, output_shape=atoms, out_activation=None)
self.adv = mlp(hidden_units['adv'], layer=Noisy, output_shape=action_dim * atoms, out_activation=None)
self.init_or_run(tf.keras.Input(shape=vector_dim + self.visual_net.hdim))
self.update_vars()
def __init__(self, vector_dim, action_dim, nums, name, hidden_units, *, visual_net):
super().__init__(visual_net, name=name)
self.action_dim = action_dim
self.nums = nums
self.net = mlp(hidden_units, output_shape=nums * action_dim, out_activation=None)
self.init_or_run(tf.keras.Input(shape=vector_dim + self.visual_net.hdim))
self.update_vars()
def __init__(self, vector_dim, output_shape, name, hidden_units, *, visual_net):
super().__init__(visual_net, name=name)
self.masking = tf.keras.layers.Masking(mask_value=0.)
self.lstm_net = tf.keras.layers.LSTM(hidden_units['lstm'], return_state=True, return_sequences=True)
self.net = mlp(hidden_units['dense'], output_shape=output_shape, out_activation=None)
self.init_or_run(tf.keras.Input(shape=(None, vector_dim + self.visual_net.hdim)))
self.update_vars()
def __init__(self, vector_dim, output_shape, name, hidden_units, *,
visual_net):
super().__init__(visual_net, name=name)
self.logits = mlp(hidden_units,
output_shape=output_shape,
out_activation=None)
self.init_or_run(
tf.keras.Input(shape=vector_dim + self.visual_net.hdim))
self.update_vars()
def __init__(self, vector_dim, output_shape, hidden_units, is_continuous):
super().__init__()
self.is_continuous = is_continuous
out_activation = 'tanh' if self.is_continuous else None
self.net = mlp(hidden_units, act_fn='tanh', output_shape=output_shape, out_activation=out_activation, out_layer=True)
self.weights_2dim = [[i, j] for i, j in zip([vector_dim]+hidden_units, hidden_units+[output_shape])]
self.weights_nums = np.asarray(self.weights_2dim).prod(axis=-1).tolist()
self.weights_total_nums = np.asarray(self.weights_2dim).prod(axis=-1).sum() + np.asarray(hidden_units).sum() + output_shape
self(tf.keras.Input(shape=vector_dim)) # 初始化网络权重
def __init__(self, vector_dim, output_shape, name, hidden_units,
visual_net, **kwargs):
super(actor_continuous_gcn, self).__init__(name=name, **kwargs)
self.visual_net = visual_net
self.tv = []
self.tv += self.visual_net.trainable_variables
self.gcn_layers1 = GraphConvolution(input_dim=64,
output_dim=64,
num_features_nonzero=0,
activation=tf.nn.relu,
dropout=0.5,
is_sparse_inputs=False,
bias=False,
featureless=False)
# self.gcn_layers2 = GraphConvolution(input_dim=64,
# output_dim=64,
# num_features_nonzero=0,
# activation=tf.nn.relu,
# dropout=0.5,
# is_sparse_inputs=False,
# bias=False,
# featureless=False)
#
# self.attention_layer1 = MultiHeadAttention(d_model=64, num_heads=8)
# self.attention_layer2 = MultiHeadAttention(d_model=64, num_heads=8)
self.layer_x_embeding = Dense(64, activation='tanh')
# self.share = mlp(hidden_units['share'], out_layer=False)
self.layer_1 = Dense(64, activation='tanh')
self.mu = mlp(hidden_units['mu'],
output_shape=output_shape,
out_activation=None)
self.log_std = mlp(hidden_units['log_std'],
output_shape=output_shape,
out_activation='tanh')
self.init_or_run(tf.keras.Input(shape=(vector_dim[0], vector_dim[0])),
tf.keras.Input(shape=vector_dim))
self.update_vars()
def __init__(self, adj_shape, x_shape, action_dim, name, hidden_units, *,
visual_net):
super().__init__(visual_net, name=name)
self.gcn_layers = []
self.gcn_layers.append(
GraphConvolution(input_dim=x_shape[1],
output_dim=16,
num_features_nonzero=0,
activation=tf.nn.relu,
dropout=0.5,
is_sparse_inputs=False))
self.attention_layer = MultiHeadAttention(d_model=16, num_heads=8)
self.net = mlp(hidden_units, output_shape=1, out_activation=None)
self.init_or_run(tf.keras.Input(shape=(adj_shape[0], adj_shape[1])),
tf.keras.Input(shape=(x_shape[0], x_shape[1])),
tf.keras.Input(shape=(1, adj_shape[0])),
tf.keras.Input(shape=action_dim))
self.update_vars()
