def __init__(self, dim_state: int, dim_action: int, hidden_sizes: List[int], normalizer: GaussianNormalizer,
init_std=1.):
super().__init__()
self.dim_state = dim_state
self.dim_action = dim_action
self.hidden_sizes = hidden_sizes
self.init_std = init_std
self.normalizer = normalizer
with self.scope:
self.op_states = tf.placeholder(tf.float32, shape=[None, dim_state], name='states')
self.op_actions_ = tf.placeholder(tf.float32, shape=[None, dim_action], name='actions')
layers = []
# note that the placeholder has size 105.
all_sizes = [dim_state, *self.hidden_sizes]
for i, (in_features, out_features) in enumerate(zip(all_sizes[:-1], all_sizes[1:])):
layers.append(nn.Linear(in_features, out_features, weight_initializer=normc_initializer(1)))
layers.append(nn.Tanh())
layers.append(nn.Linear(all_sizes[-1], dim_action, weight_initializer=normc_initializer(0.01)))
self.net = nn.Sequential(*layers)
self.op_log_std = nn.Parameter(
tf.constant(np.log(self.init_std), shape=[self.dim_action], dtype=tf.float32), name='log_std')
self.distribution = self(self.op_states)
self.op_actions = self.distribution.sample()
self.op_actions_mean = self.distribution.mean()
self.op_actions_std = self.distribution.stddev()
self.op_nlls_ = -self.distribution.log_prob(self.op_actions_).reduce_sum(axis=1)
def __init__(self, dim_state: int, dim_action: int,
normalizers: Normalizers, *, arch: FLAGS.arch):
super().__init__()
initializer = tf.truncated_normal_initializer(mean=0.0, stddev=1e-5)
self.dim_state = dim_state
self.dim_action = dim_action
self.op_states = tf.placeholder(tf.float32,
shape=[None, self.dim_state],
name='states')
self.op_actions = tf.placeholder(tf.float32,
shape=[None, self.dim_action],
name='actions')
self.mlp = nn.Sequential(
nn.Linear(dim_state + dim_action,
arch.n_units,
weight_initializer=initializer),
nn.ReLU(),
make_blocks(FixupResBlock, arch.n_units, arch.n_blocks,
arch.n_blocks),
nn.Linear(arch.n_units, dim_state, weight_initializer=initializer),
)
self.normalizers = normalizers
self.build()
def __init__(self,
blocks,
activation=nn.ReLU,
squeeze=False,
weight_initializer=None,
build=True):
super().__init__()
self._blocks = blocks
if build:
self.op_inputs = tf.placeholder(tf.float32,
[None, self._blocks[0]])
with self.scope:
kwargs = {}
if weight_initializer is not None:
kwargs['weight_initializer'] = weight_initializer
layers = []
for in_features, out_features in zip(blocks[:-1], blocks[1:]):
if layers:
layers.append(activation())
layers.append(nn.Linear(in_features, out_features, **kwargs))
if squeeze:
layers.append(nn.Squeeze(axis=1))
self.net = nn.Sequential(*layers)
self._squeeze = squeeze
self._activation = activation
if build:
self.build()
def __init__(self,
blocks,
activation,
squeeze=False,
weight_initializer=None,
output_activation=None):
super().__init__()
self._blocks = blocks
with self.scope:
kwargs = {}
if weight_initializer is not None:
kwargs['weight_initializer'] = weight_initializer
layers = []
for in_features, out_features in zip(blocks[:-1], blocks[1:]):
if layers:
layers.append(activation())
layers.append(nn.Linear(in_features, out_features, **kwargs))
if squeeze:
layers.append(nn.Squeeze(axis=-1))
if output_activation:
layers.append(output_activation())
self._modules = {i: module for i, module in enumerate(layers)}
self._squeeze = squeeze
self._activation = activation
self._built = False
def __init__(self, x, n_total_blocks):
super().__init__()
std = np.sqrt(2. / x / n_total_blocks)
self.bias1a = nn.Parameter(tf.zeros(1), name='bias1a')
self.fc1 = nn.Linear(x,
x,
bias=False,
weight_initializer=tf.initializers.random_normal(
0, stddev=std))
self.bias1b = nn.Parameter(tf.zeros(1), name='bias1b')
self.relu = nn.ReLU()
self.bias2a = nn.Parameter(tf.zeros(1), name='bias2a')
self.fc2 = nn.Linear(x,
x,
bias=False,
weight_initializer=tf.initializers.zeros())
self.scale = nn.Parameter(tf.ones(1), name='scale')
self.bias2b = nn.Parameter(tf.zeros(1), name='bias2b')
def __init__(self, n_params):
'''
self.goal_velocity: [-1,1]
'''
super().__init__()
with self.scope:
layers = []
layers.append(nn.Linear(1, n_params, bias=False, weight_initializer=normc_initializer(1.0)))
if FLAGS.task.scaler == 'tanh':
layers.append(nn.Tanh())
self.net = nn.Sequential(*layers)
c = tf.constant(1, shape=[1,1], dtype=tf.float32)
#self.net(c) -> [None, n_params]
if n_params == 1:
self.goal_velocity = self.net(c)[0]
else:
self.goal_velocity = self.net(c)[0]
print (self.goal_velocity)