def __init__(self, config, name='discriminator'):
super().__init__(name)
config = config.copy()
self._obs_action = config.pop('obs_action', True)
self._g = mlp(
**config, out_size=1, name='g'
) # the AIRL paper use a linear model here, we use the same model for g and h for simplicity
self._h = mlp(**config, out_size=1, name='h')
def _add_layer(self, config):
""" Network definition """
if self._duel:
self._v_layers = mlp(
**config,
out_size=1,
name=self.name+'/v',
out_dtype='float32')
self._layers = mlp(
**config,
out_size=self.action_dim,
name=self.name,
out_dtype='float32')
def _add_layer(self, config):
""" Network definition """
if getattr(self, '_duel', False):
self._v_layers = mlp(
**config,
out_size=1,
name=self.name+'/v',
out_dtype='float32')
# we do not define the phi net here to make it consistent with the CNN output size
self._layers = mlp(
**config,
out_size=self.action_dim,
name=self.name,
out_dtype='float32')
def __init__(self, name='rssm'):
super().__init__(name)
self._embed_layer = layers.Dense(self._hidden_size,
activation=self._activation,
name='embed')
self._cell = layers.GRUCell(self._deter_size)
self._img_layers = mlp([self._hidden_size],
out_size=2 * self._stoch_size,
activation=self._activation,
name='img')
self._obs_layers = mlp([self._hidden_size],
out_size=2 * self._stoch_size,
activation=self._activation,
name='obs')
def test_mlp(self):
units_list = [10, 5]
activation = 'relu'
kernel_initializer = 'he_uniform'
out_dim = 3
layer_seed = 10
tf.random.set_seed(0)
x = tf.random.normal([1, 2])
tf.random.set_seed(layer_seed)
plain_layers = [tf.keras.layers.Dense(
u, activation=activation, kernel_initializer=kernel_initializer)
for u in units_list]
plain_layers.append(tf.keras.layers.Dense(out_dim))
plain_y = x
for l in plain_layers:
plain_y = l(plain_y)
tf.random.set_seed(layer_seed)
mlp_layers = mlp(units_list, out_dim, activation=activation, kernel_initializer=kernel_initializer)
mlp_y = mlp_layers(x)
np.testing.assert_allclose(plain_y.numpy(), mlp_y.numpy())
plain_vars = []
for l in plain_layers:
plain_vars += l.variables
for pv, mv in zip(plain_vars, mlp_layers.variables):
np.testing.assert_allclose(pv.numpy(), mv.numpy())
def __init__(self, config, n_agents, name='qmixer'):
super().__init__(name=name)
config = config.copy()
self.n_agents = n_agents
self.hidden_dim = config.pop('hidden_dim')
self.w1 = mlp(**config,
out_size=n_agents * self.hidden_dim,
name=f'{self.name}/w1')
self.w2 = mlp(**config,
out_size=self.hidden_dim,
name=f'{self.name}/w2')
self.b = mlp([], self.hidden_dim, name=f'{self.name}/b')
config['units_list'] = [self.hidden_dim]
self.v = mlp(**config, out_size=1, name=f'{self.name}/v')
def __init__(self, name='fqn'):
super().__init__(name=name)
kernel_initializer = tf.keras.initializers.VarianceScaling(
1./np.sqrt(3.), distribution='uniform')
self._layers = mlp(
out_size=self.N,
name=f'{self.name}/fpn',
kernel_initializer=kernel_initializer)
def __init__(self, config, name='value'):
super().__init__(name=name)
config = config.copy()
config.setdefault('out_gain', 1)
self._layers = mlp(**config,
out_size=1,
out_dtype='float32',
name=name)
def __init__(self, name='encoder'):
super().__init__(name=name)
if getattr(self, '_has_cnn', True):
self._layers = ConvEncoder(time_distributed=True)
else:
self._layers = mlp(self._units_list,
activation=self._activation,
name=name)
def __init__(self, action_dim, is_action_discrete, name='actor'):
super().__init__(name=name)
""" Network definition """
out_size = action_dim if is_action_discrete else 2 * action_dim
self._layers = mlp(self._units_list,
out_size=out_size,
activation=self._activation,
name=name)
self._is_action_discrete = is_action_discrete
def __init__(self, out_size=1, dist='normal', name='decoder'):
super().__init__(name=name)
self._dist = dist
if getattr(self, '_has_cnn', None):
self._layers = ConvDecoder(time_distributed=True)
else:
self._layers = mlp(self._units_list,
out_size=out_size,
activation=self._activation,
name=name)
def __init__(self, config, action_dim, name='actor'):
super().__init__(name=name)
config = config.copy()
self._action_dim = action_dim
self.LOG_STD_MIN = config.pop('LOG_STD_MIN', -20)
self.LOG_STD_MAX = config.pop('LOG_STD_MAX', 2)
self._tsallis_q = config.pop('tsallis_q', 1)
out_size = 2*action_dim
self._layers = mlp(**config, out_size=out_size, name=name)
def __init__(self, config, action_dim, name='actor'):
super().__init__(name=name)
config = config.copy()
self._action_dim = action_dim
prior = np.ones(action_dim, dtype=np.float32)
prior /= np.sum(prior)
self.prior = tf.Variable(prior, trainable=False, name='prior')
self._epsilon_scaled_logits = config.pop('epsilon_scaled_logits',
False)
self._layers = mlp(**config,
out_size=action_dim,
out_dtype='float32',
name=name)
def __init__(self, config, action_dim, is_action_discrete, name='actor'):
super().__init__(name=name)
config = config.copy()
self.action_dim = action_dim
self.is_action_discrete = is_action_discrete
self.eval_act_temp = config.pop('eval_act_temp', 1)
assert self.eval_act_temp >= 0, self.eval_act_temp
self._init_std = config.pop('init_std', 1)
if not self.is_action_discrete:
self.logstd = tf.Variable(
initial_value=np.log(self._init_std)*np.ones(action_dim),
dtype='float32',
trainable=True,
name=f'actor/logstd')
config.setdefault('out_gain', .01)
self._layers = mlp(**config,
out_size=action_dim,
out_dtype='float32',
name=name)
def __init__(self, config, name='q'):
super().__init__(name=name)
config = config.copy()
self._layers = mlp(**config, out_size=1, name=name)
def mlp(self, x, *args, name, **kwargs):
if not hasattr(self, f'_{name}'):
from nn.func import mlp
setattr(self, f'_{name}', mlp(*args, name=name, **kwargs))
return getattr(self, f'_{name}')(x)
def __init__(self, config, name='discriminator'):
super().__init__(name)
config = config.copy()
self._layers = mlp(**config, out_size=1)