def test_num_features(self): assert rcfr.num_features(_GAME) == 13
def __init__(self,
game,
num_hidden_units,
num_hidden_layers=1,
num_hidden_factors=0,
hidden_activation=tf.nn.relu,
use_skip_connections=False,
regularizer=None,
autoencode=False):
"""Creates a new `DeepNeurdModel.
Args:
game: The OpenSpiel game being solved.
num_hidden_units: The number of units in each hidden layer.
num_hidden_layers: The number of hidden layers. Defaults to 1.
num_hidden_factors: The number of hidden factors or the matrix rank of the
layer. If greater than zero, hidden layers will be split into two
separate linear transformations, the first with
`num_hidden_factors`-columns and the second with
`num_hidden_units`-columns. The result is that the logical hidden layer
is a rank-`num_hidden_units` matrix instead of a rank-`num_hidden_units`
matrix. When `num_hidden_units < num_hidden_units`, this is effectively
implements weight sharing. Defaults to 0.
hidden_activation: The activation function to apply over hidden layers.
Defaults to `tf.nn.relu`.
use_skip_connections: Whether or not to apply skip connections (layer
output = layer(x) + x) on hidden layers. Zero padding or truncation is
used to match the number of columns on layer inputs and outputs.
regularizer: A regularizer to apply to each layer. Defaults to `None`.
autoencode: Whether or not to output a reconstruction of the inputs upon
being called. Defaults to `False`.
"""
self._autoencode = autoencode
self._use_skip_connections = use_skip_connections
self._hidden_are_factored = num_hidden_factors > 0
self.layers = []
for _ in range(num_hidden_layers):
if self._hidden_are_factored:
self.layers.append(
tf.keras.layers.Dense(num_hidden_factors,
use_bias=True,
kernel_regularizer=regularizer))
self.layers.append(
tf.keras.layers.Dense(num_hidden_units,
use_bias=True,
activation=hidden_activation,
kernel_regularizer=regularizer))
self.layers.append(
tf.keras.layers.Dense(1 +
self._autoencode * rcfr.num_features(game),
use_bias=True,
kernel_regularizer=regularizer))
# Construct variables for all layers by exercising the network.
x = tf.zeros([1, rcfr.num_features(game)])
for layer in self.layers:
x = layer(x)
self.trainable_variables = sum(
[layer.trainable_variables for layer in self.layers], [])
self.losses = sum([layer.losses for layer in self.layers], [])
