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
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 _build_draw_initial(self):
""" Builds the draw model (initial step) """
from diplomacy_research.utils.tensorflow import tf
from diplomacy_research.models.layers.graph_convolution import GraphConvolution, preprocess_adjacency
from diplomacy_research.utils.tensorflow import to_float
if not self.placeholders:
self.placeholders = self.get_placeholders()
else:
self.placeholders.update(self.get_placeholders())
# 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
sigmoid = tf.nn.sigmoid
# Training loop
with tf.variable_scope('draw', reuse=tf.AUTO_REUSE):
with tf.device(self.cluster_config.worker_device if self.
cluster_config else None):
# Features
board_state = to_float(
self.features['board_state']
) # tf.float32 - (b, NB_NODES, NB_FEATURES)
current_power = self.features[
'current_power'] # tf.int32 - (b,)
draw_target = self.features['draw_target'] # tf.float32 - (b,)
# Placeholders
stop_gradient_all = pholder('stop_gradient_all')
# Norm Adjacency
batch_size = tf.shape(board_state)[0]
norm_adjacency = preprocess_adjacency(get_adjacency_matrix())
norm_adjacency = tf.tile(
tf.expand_dims(norm_adjacency, axis=0), [batch_size, 1, 1])
# Graph embeddings
with tf.variable_scope('graph_conv_scope'):
board_state_h0 = board_state # (b, 81, 35)
board_state_h1 = GraphConvolution(
input_dim=NB_FEATURES,
output_dim=hps('draw_gcn_1_output_size'),
norm_adjacency=norm_adjacency,
activation_fn=relu,
bias=True)(board_state_h0) # (b, 81, 25)
# board_state_h2: (b, 2025)
# board_state_h3: (b, 128)
board_state_h2 = tf.reshape(
board_state_h1,
shape=[-1, NB_NODES * hps('draw_gcn_1_output_size')])
board_state_graph_conv = tf.layers.Dense(
units=hps('draw_embedding_size'),
activation=relu,
use_bias=True)(board_state_h2)
# Calculating draw for all powers
with tf.variable_scope('draw_scope'):
current_power_mask = tf.one_hot(current_power,
NB_POWERS,
dtype=tf.float32)
draw_h0 = board_state_graph_conv # (b, 128)
draw_h1 = tf.layers.Dense(
units=hps('draw_h1_size'), # (b, 64)
activation=relu,
use_bias=True)(draw_h0)
draw_h2 = tf.layers.Dense(
units=hps('draw_h2_size'), # (b, 64)
activation=relu,
use_bias=True)(draw_h1)
draw_probs = tf.layers.Dense(
units=NB_POWERS, # (b, 7)
activation=sigmoid,
use_bias=True)(draw_h2)
draw_prob = tf.reduce_sum(draw_probs * current_power_mask,
axis=1) # (b,)
# Computing draw loss
with tf.variable_scope('draw_loss'):
draw_loss = tf.reduce_mean(
tf.square(draw_target - draw_prob))
draw_loss = tf.cond(
stop_gradient_all,
lambda: tf.stop_gradient(draw_loss), # pylint: disable=cell-var-from-loop
lambda: draw_loss) # pylint: disable=cell-var-from-loop
# Building output tags
outputs = {
'tag/draw/v001_draw_relu': True,
'draw_prob': draw_prob,
'draw_loss': draw_loss
}
# Adding features, placeholders and outputs to graph
self.add_meta_information(outputs)
def _build_policy_initial(self):
""" Builds the policy model (initial step) """
from diplomacy_research.utils.tensorflow import tf
from diplomacy_research.models.layers.initializers import uniform
from diplomacy_research.utils.tensorflow import pad_axis, to_int32, to_float, to_bool
if not self.placeholders:
self.placeholders = self.get_placeholders()
# 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]
# Training loop
with tf.variable_scope('policy', reuse=tf.AUTO_REUSE):
with tf.device(self.cluster_config.worker_device if self.
cluster_config else None):
# Features
board_state = to_float(
self.features['board_state']
) # tf.flt32 - (b, NB_NODES, NB_FEATURES)
board_alignments = to_float(
self.features['board_alignments']) # (b, NB_NODES * len)
prev_orders_state = to_float(
self.features['prev_orders_state']
) # (b, NB_PRV_OD, NB_ND, NB_OD_FT)
decoder_inputs = self.features[
'decoder_inputs'] # tf.int32 - (b, <= 1 + NB_SCS)
decoder_lengths = self.features[
'decoder_lengths'] # tf.int32 - (b,)
candidates = self.features[
'candidates'] # tf.int32 - (b, nb_locs * MAX_CANDIDATES)
current_power = self.features[
'current_power'] # tf.int32 - (b,)
current_season = self.features[
'current_season'] # tf.int32 - (b,)
dropout_rates = self.features[
'dropout_rate'] # tf.flt32 - (b,)
# Batch size
batch_size = tf.shape(board_state)[0]
# Reshaping board alignments
board_alignments = tf.reshape(board_alignments,
[batch_size, -1, NB_NODES])
board_alignments /= tf.math.maximum(
1., tf.reduce_sum(board_alignments, axis=-1,
keepdims=True))
# Overriding dropout_rates if pholder('dropout_rate') > 0
dropout_rates = tf.cond(
tf.greater(pholder('dropout_rate'), 0.),
true_fn=lambda: tf.zeros_like(dropout_rates) + pholder(
'dropout_rate'),
false_fn=lambda: dropout_rates)
# Padding decoder_inputs and candidates
board_alignments = pad_axis(
board_alignments,
axis=1,
min_size=tf.reduce_max(decoder_lengths))
decoder_inputs = pad_axis(decoder_inputs, axis=-1, min_size=2)
candidates = pad_axis(candidates,
axis=-1,
min_size=MAX_CANDIDATES)
# Making sure all RNN lengths are at least 1
# No need to trim, because the fields are variable length
raw_decoder_lengths = decoder_lengths
decoder_lengths = tf.math.maximum(1, decoder_lengths)
# Placeholders
decoder_type = tf.reduce_max(pholder('decoder_type'))
is_training = pholder('is_training')
# Reshaping candidates
candidates = tf.reshape(candidates,
[batch_size, -1, MAX_CANDIDATES])
candidates = candidates[:, :tf.reduce_max(
decoder_lengths), :] # tf.int32 - (b, nb_locs, MAX_CAN)
# Computing FiLM Gammas and Betas
with tf.variable_scope('film_scope'):
power_embedding = uniform(
name='power_embedding',
shape=[NB_POWERS, hps('power_emb_size')],
scale=1.)
current_power_mask = tf.one_hot(current_power,
NB_POWERS,
dtype=tf.float32)
current_power_embedding = tf.reduce_sum(
power_embedding[None] * current_power_mask[:, :, None],
axis=1) # (b, power_emb)
film_embedding_input = current_power_embedding
# Also conditioning on current_season
season_embedding = uniform(
name='season_embedding',
shape=[NB_SEASONS, hps('season_emb_size')],
scale=1.)
current_season_mask = tf.one_hot(current_season,
NB_SEASONS,
dtype=tf.float32)
current_season_embedding = tf.reduce_sum(
season_embedding[None] # (b,season_emb)
* current_season_mask[:, :, None],
axis=1)
film_embedding_input = tf.concat(
[film_embedding_input, current_season_embedding],
axis=1)
film_output_dims = [hps('gcn_size')] * (
hps('nb_graph_conv') - 1) + [hps('attn_size') // 2]
# For board_state
board_film_weights = tf.layers.Dense(
units=2 * sum(film_output_dims), # (b, 1, 750)
use_bias=True,
activation=None)(film_embedding_input)[:, None, :]
board_film_gammas, board_film_betas = tf.split(
board_film_weights, 2, axis=2) # (b, 1, 750)
board_film_gammas = tf.split(board_film_gammas,
film_output_dims,
axis=2)
board_film_betas = tf.split(board_film_betas,
film_output_dims,
axis=2)
# For prev_orders
prev_ord_film_weights = tf.layers.Dense(
units=2 * sum(film_output_dims), # (b, 1, 750)
use_bias=True,
activation=None)(film_embedding_input)[:, None, :]
prev_ord_film_weights = tf.tile(
prev_ord_film_weights,
[NB_PREV_ORDERS, 1, 1]) # (n_pr, 1, 750)
prev_ord_film_gammas, prev_ord_film_betas = tf.split(
prev_ord_film_weights, 2, axis=2)
prev_ord_film_gammas = tf.split(prev_ord_film_gammas,
film_output_dims,
axis=2)
prev_ord_film_betas = tf.split(prev_ord_film_betas,
film_output_dims,
axis=2)
# Storing as temporary output
self.add_output('_board_state_conv_film_gammas',
board_film_gammas)
self.add_output('_board_state_conv_film_betas',
board_film_betas)
self.add_output('_prev_orders_conv_film_gammas',
prev_ord_film_gammas)
self.add_output('_prev_orders_conv_film_betas',
prev_ord_film_betas)
# Creating graph convolution
with tf.variable_scope('graph_conv_scope'):
assert hps('nb_graph_conv') >= 2
assert hps('attn_size') % 2 == 0
# Encoding board state
board_state_0yr_conv = self.encode_board(
board_state, name='board_state_conv')
# Encoding prev_orders
prev_orders_state = tf.reshape(prev_orders_state, [
batch_size * NB_PREV_ORDERS, NB_NODES,
NB_ORDERS_FEATURES
])
prev_ord_conv = self.encode_board(prev_orders_state,
name='prev_orders_conv')
# Splitting back into (b, nb_prev, NB_NODES, attn_size // 2)
# Reducing the prev ord conv using avg
prev_ord_conv = tf.reshape(prev_ord_conv, [
batch_size, NB_PREV_ORDERS, NB_NODES,
hps('attn_size') // 2
])
prev_ord_conv = tf.reduce_mean(prev_ord_conv, axis=1)
# Concatenating the current board conv with the prev ord conv
# The final board_state_conv should be of dimension (b, NB_NODE, attn_size)
board_state_conv = self.get_board_state_conv(
board_state_0yr_conv, is_training, prev_ord_conv)
# Creating order embedding vector (to embed order_ix)
# Embeddings needs to be cached locally on the worker, otherwise TF can't compute their gradients
with tf.variable_scope('order_embedding_scope'):
# embedding: (order_vocab_size, 64)
caching_device = self.cluster_config.caching_device if self.cluster_config else None
partitioner = tf.fixed_size_partitioner(
NB_PARTITIONS) if hps('use_partitioner') else None
order_embedding = uniform(
name='order_embedding',
shape=[ORDER_VOCABULARY_SIZE,
hps('order_emb_size')],
scale=1.,
partitioner=partitioner,
caching_device=caching_device)
# Creating candidate embedding
with tf.variable_scope('candidate_embedding_scope'):
# embedding: (order_vocab_size, 64)
caching_device = self.cluster_config.caching_device if self.cluster_config else None
partitioner = tf.fixed_size_partitioner(
NB_PARTITIONS) if hps('use_partitioner') else None
candidate_embedding = uniform(
name='candidate_embedding',
shape=[ORDER_VOCABULARY_SIZE,
hps('lstm_size') + 1],
scale=1.,
partitioner=partitioner,
caching_device=caching_device)
# Trimming to the maximum number of candidates
candidate_lengths = tf.reduce_sum(
to_int32(tf.math.greater(candidates, PAD_ID)),
-1) # int32 - (b,)
max_candidate_length = tf.math.maximum(
1, tf.reduce_max(candidate_lengths))
candidates = candidates[:, :, :max_candidate_length]
# Building output tags
outputs = {
'batch_size':
batch_size,
'board_alignments':
board_alignments,
'decoder_inputs':
decoder_inputs,
'decoder_type':
decoder_type,
'raw_decoder_lengths':
raw_decoder_lengths,
'decoder_lengths':
decoder_lengths,
'board_state_conv':
board_state_conv,
'board_state_0yr_conv':
board_state_0yr_conv,
'prev_ord_conv':
prev_ord_conv,
'order_embedding':
order_embedding,
'candidate_embedding':
candidate_embedding,
'candidates':
candidates,
'max_candidate_length':
max_candidate_length,
'in_retreat_phase':
tf.math.logical_and( # 1) board not empty, 2) disl. units present
tf.reduce_sum(board_state[:], axis=[1, 2]) > 0,
tf.math.logical_not(
to_bool(tf.reduce_min(board_state[:, :, 23], -1))))
}
# Adding to graph
self.add_meta_information(outputs)
def _get_board_value(self,
board_state,
current_power,
name='board_state_value',
reuse=None):
""" Computes the estimated value of a board state
:param board_state: The board state - (batch, NB_NODES, NB_FEATURES)
:param current_power: The power for which we want the board value - (batch,)
:param name: The name to use for the operaton
:param reuse: Whether to reuse or not the weights from another operation
:return: The value of the board state for the specified power - (batch,)
"""
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
# No need to use 'stop_gradient_value' - Because this model does not share parameters.
scope = tf.VariableScope(name='value/%s' % name, reuse=reuse)
with tf.variable_scope(scope):
with tf.variable_scope('graph_conv_scope'):
graph_conv = board_state # (b, NB_NODES, NB_FEAT)
graph_conv = GraphConvolution(
input_dim=graph_conv.shape[-1].
value, # (b, NB_NODES, gcn_1)
output_dim=hps('value_gcn_1_output_size'),
norm_adjacency=norm_adjacency,
activation_fn=relu,
bias=True)(graph_conv)
flat_graph_conv = tf.reshape(
graph_conv,
shape=[-1, NB_NODES * hps('value_gcn_1_output_size')])
flat_graph_conv = tf.layers.Dense(
units=hps('value_embedding_size'),
activation=relu,
use_bias=True)(flat_graph_conv) # (b, value_emb_size)
with tf.variable_scope('value_scope'):
current_power_mask = tf.one_hot(current_power,
NB_POWERS,
dtype=tf.float32)
state_value = flat_graph_conv # (b, value_emb_size)
state_value = tf.layers.Dense(
units=hps('value_h1_size'), # (b, value_h1_size)
activation=relu,
use_bias=True)(state_value)
state_value = tf.layers.Dense(
units=hps('value_h2_size'), # (b, value_h2_size)
activation=relu,
use_bias=True)(state_value)
state_value = tf.layers.Dense(
units=NB_POWERS, # (b, NB_POWERS)
activation=None,
use_bias=True)(state_value)
state_value = tf.reduce_sum(state_value * current_power_mask,
axis=1) # (b,)
# Returning
return state_value
