def _encode_board(self, board_state, name, reuse=None):
""" Encodes a board state or prev orders state
:param board_state: The board state / prev orders state to encode - (batch, NB_NODES, initial_features)
:param name: The name to use for the encoding
:param reuse: Whether to reuse or not the weights from another encoding operation
:return: The encoded board state / prev_orders state
"""
from diplomacy_research.utils.tensorflow import tf
from diplomacy_research.models.layers.graph_convolution import film_gcn_res_block, preprocess_adjacency
# Quick function to retrieve hparams and placeholders and function shorthands
hps = lambda hparam_name: self.hparams[hparam_name]
pholder = lambda placeholder_name: self.placeholders[placeholder_name]
relu = tf.nn.relu
# Getting film gammas and betas
film_gammas = self.outputs['_%s_film_gammas' % name]
film_betas = self.outputs['_%s_film_betas' % name]
# Computing norm adjacency
norm_adjacency = preprocess_adjacency(get_adjacency_matrix())
norm_adjacency = tf.tile(tf.expand_dims(norm_adjacency, axis=0),
[tf.shape(board_state)[0], 1, 1])
# Building scope
scope = tf.VariableScope(name='policy/%s' % name, reuse=reuse)
with tf.variable_scope(scope):
# Adding noise to break symmetry
board_state = board_state + tf.random_normal(tf.shape(board_state),
stddev=0.01)
graph_conv = tf.layers.Dense(units=hps('gcn_size'),
activation=relu)(board_state)
# First and intermediate layers
for layer_idx in range(hps('nb_graph_conv') - 1):
graph_conv = film_gcn_res_block(
inputs=graph_conv, # (b, NB_NODES, gcn_size)
gamma=film_gammas[layer_idx],
beta=film_betas[layer_idx],
gcn_out_dim=hps('gcn_size'),
norm_adjacency=norm_adjacency,
is_training=pholder('is_training'),
residual=True)
# Last layer
graph_conv = film_gcn_res_block(
inputs=graph_conv, # (b, NB_NODES, final_size)
gamma=film_gammas[-1],
beta=film_betas[-1],
gcn_out_dim=hps('attn_size') // 2,
norm_adjacency=norm_adjacency,
is_training=pholder('is_training'),
residual=False)
# Returning
return graph_conv
def _encode_board(self, board_state, name, reuse=None):
""" Encodes a board state or prev orders state
:param board_state: The board state / prev orders state to encode - (batch, NB_NODES, initial_features)
:param name: The name to use for the encoding
:param reuse: Whether to reuse or not the weights from another encoding operation
:return: The encoded board state / prev_orders state
"""
from diplomacy_research.utils.tensorflow import tf
from diplomacy_research.models.layers.graph_convolution import GraphConvolution, preprocess_adjacency
from diplomacy_research.utils.tensorflow import batch_norm
# Quick function to retrieve hparams and placeholders and function shorthands
hps = lambda hparam_name: self.hparams[hparam_name]
pholder = lambda placeholder_name: self.placeholders[placeholder_name]
relu = tf.nn.relu
# Computing norm adjacency
norm_adjacency = preprocess_adjacency(get_adjacency_matrix())
norm_adjacency = tf.tile(tf.expand_dims(norm_adjacency, axis=0), [tf.shape(board_state)[0], 1, 1])
# Building scope
scope = tf.VariableScope(name='policy/%s' % name, reuse=reuse)
with tf.variable_scope(scope):
# Adding noise to break symmetry
board_state = board_state + tf.random_normal(tf.shape(board_state), stddev=0.01)
graph_conv = board_state
# First Layer
graph_conv = GraphConvolution(input_dim=graph_conv.shape[-1].value, # (b, NB_NODES, gcn_size)
output_dim=hps('gcn_size'),
norm_adjacency=norm_adjacency,
activation_fn=relu,
bias=True)(graph_conv)
# Intermediate Layers
for _ in range(1, hps('nb_graph_conv') - 1):
graph_conv = GraphConvolution(input_dim=hps('gcn_size'), # (b, NB_NODES, gcn_size)
output_dim=hps('gcn_size'),
norm_adjacency=norm_adjacency,
activation_fn=relu,
bias=True)(graph_conv)
graph_conv = batch_norm(graph_conv, is_training=pholder('is_training'), fused=True)
# Final Layer
graph_conv = GraphConvolution(input_dim=hps('gcn_size'), # (b, NB_NODES, attn_size)
output_dim=hps('attn_size'),
norm_adjacency=norm_adjacency,
activation_fn=None,
bias=True)(graph_conv)
# Returning
return graph_conv
def _encode_board(self, board_state, name, reuse=None):
""" Encodes a board state or prev orders state
:param board_state: The board state / prev orders state to encode - (batch, NB_NODES, initial_features)
:param name: The name to use for the encoding
:param reuse: Whether to reuse or not the weights from another encoding operation
:return: The encoded board state / prev_orders state
"""
from diplomacy_research.utils.tensorflow import tf
from diplomacy_research.models.layers.graph_convolution import GraphConvolution, preprocess_adjacency
# Quick function to retrieve hparams and placeholders and function shorthands
hps = lambda hparam_name: self.hparams[hparam_name]
relu = tf.nn.relu
# Computing norm adjacency
norm_adjacency = preprocess_adjacency(get_adjacency_matrix())
norm_adjacency = tf.tile(tf.expand_dims(norm_adjacency, axis=0), [tf.shape(board_state)[0], 1, 1])
# Building scope
scope = tf.VariableScope(name='policy/%s' % name, reuse=reuse)
with tf.variable_scope(scope):
batch_size = tf.shape(board_state)[0]
# Adding noise to break symmetry
board_state = board_state + tf.random_normal(tf.shape(board_state), stddev=0.01)
# Projecting (if needed) to 'gcn_size'
if board_state.shape[-1].value == NB_FEATURES:
with tf.variable_scope('proj', reuse=tf.AUTO_REUSE):
proj_w = tf.get_variable('W', shape=[1, NB_FEATURES, hps('gcn_size')], dtype=tf.float32)
graph_conv = relu(tf.matmul(board_state, tf.tile(proj_w, [batch_size, 1, 1])))
else:
graph_conv = board_state
# First and intermediate layers
for _ in range(hps('nb_graph_conv') - 1):
graph_conv = GraphConvolution(input_dim=hps('gcn_size'), # (b, NB_NODES, gcn_size)
output_dim=hps('gcn_size'),
norm_adjacency=norm_adjacency,
activation_fn=relu,
residual=True,
bias=True)(graph_conv)
# Last Layer
graph_conv = GraphConvolution(input_dim=hps('gcn_size'), # (b, NB_NODES, final_size)
output_dim=hps('attn_size') // 2,
norm_adjacency=norm_adjacency,
activation_fn=relu,
residual=False,
bias=True)(graph_conv)
# Returning
return graph_conv