def pack_graph_embeddings(self, graph_features, ge):
# Pack graph embeddings to be transported from actors to learners.
# For now we only handle BipartiteGraphs
if set(graph_features.keys()) != set(
gn.graphs.BipartiteGraphsTuple._fields):
raise Exception(f'Not handled {graph_features.keys()}')
graph_features = gn.graphs.BipartiteGraphsTuple(**graph_features)
left_nodes = tf.scatter_nd(
gn.utils_tf.sparse_to_dense_indices(ge.n_left_nodes),
ge.left_nodes,
infer_shape(graph_features.left_nodes)[:2] +
[infer_shape(ge.left_nodes)[-1]])
right_nodes = tf.scatter_nd(
gn.utils_tf.sparse_to_dense_indices(ge.n_right_nodes),
ge.right_nodes,
infer_shape(graph_features.right_nodes)[:2] +
[infer_shape(ge.right_nodes)[-1]])
bs = infer_shape(graph_features.left_nodes)[0]
dummy_tensor = lambda: tf.fill(tf.expand_dims(bs, 0), 0)
return gn.graphs.BipartiteGraphsTuple(
left_nodes=left_nodes,
right_nodes=right_nodes,
globals=ge.globals,
n_left_nodes=ge.n_left_nodes,
n_right_nodes=ge.n_right_nodes,
# edges is dummy tensor -- not used
edges=tf.expand_dims(tf.expand_dims(dummy_tensor(), -1), -1),
senders=dummy_tensor(),
receivers=dummy_tensor(),
n_edge=dummy_tensor(),
)
def get_value(self, graph_features: gn.graphs.GraphsTuple):
"""
graph_embeddings: Message propagated graph embeddings.
Use self.compute_graph_embeddings to compute and cache
these to use with different network heads for value, policy etc.
"""
with tf.variable_scope('value_network'):
left_nodes = graph_features.left_nodes
right_nodes = graph_features.right_nodes
num_graphs = infer_shape(graph_features.n_left_nodes)[0]
# Apply mlp to left and right nodes.
left_nodes = self.value_torso_1_left(left_nodes)
right_nodes = self.value_torso_1_right(right_nodes)
# aggregate left and right nodes seperately.
left_indices = gn.utils_tf.repeat(tf.range(num_graphs), graph_features.n_left_nodes, axis=0)
left_agg = tf.unsorted_segment_mean(left_nodes[:infer_shape(left_indices)[0]], left_indices,
num_graphs)
right_indices = gn.utils_tf.repeat(tf.range(num_graphs),
graph_features.n_right_nodes,
axis=0)
right_agg = tf.unsorted_segment_mean(right_nodes[:infer_shape(right_indices)[0]],
right_indices, num_graphs)
value = tf.concat([left_agg, right_agg, graph_features.globals], axis=-1)
return tf.squeeze(self.value_torso_2(value), axis=-1)
def _add_stop_action(self, graph_features):
with tf.variable_scope('switch_network'):
# aggregate left and right nodes seperately.
left_nodes = graph_features.left_nodes
num_graphs = infer_shape(graph_features.n_left_nodes)[0]
left_indices = gn.utils_tf.repeat(tf.range(num_graphs), graph_features.n_left_nodes, axis=0)
left_agg = tf.unsorted_segment_mean(left_nodes[:infer_shape(left_indices)[0]], left_indices,
num_graphs)
switch_logits = tf.concat([left_agg, graph_features.globals], axis=-1)
# Now calculate the logits for the switch binary action.
switch_logits = tf.squeeze(self.switch_torso(switch_logits), axis=-1)
return switch_logits
def get_node_embeddings(self, obs, graph_embeddings):
# Returns embeddings of the nodes in the shape (B, N_max, d)
graph_features = graph_embeddings
broadcasted_globals = gn.utils_tf.repeat(graph_features.globals,
graph_features.n_left_nodes,
axis=0)
left_nodes = graph_features.left_nodes[:infer_shape(broadcasted_globals)[0]]
left_nodes = tf.concat([left_nodes, broadcasted_globals], axis=-1)
indices = gn.utils_tf.sparse_to_dense_indices(graph_features.n_left_nodes)
left_nodes = tf.scatter_nd(indices, left_nodes,
infer_shape(obs['node_mask']) + [infer_shape(left_nodes)[-1]])
return left_nodes, graph_features.n_left_nodes
def get_value(self, graph_embeddings, obs):
# Give this the optimal solution as well in the inputs?
if self.config.use_mlp_value_func:
xs, _ = self.get_node_embeddings(obs,
graph_embeddings) # (N, L1, d)
xs = tf.reshape(xs, [infer_shape(xs)[0], -1])
with tf.variable_scope('value_network'):
self.value = snt.nets.MLP(
[32, 32, 1],
initializers=dict(w=glorot_uniform(self.seed),
b=initializers.init_ops.Constant(0.0)),
activate_final=False,
activation=get_activation_from_str('relu'))
return tf.squeeze(self.value(xs), axis=-1)
else:
return self._gcn_model.get_value(graph_embeddings)
def get_auxiliary_loss(self, graph_features: gn.graphs.GraphsTuple, obs):
"""
Returns a prediction for each node.
This is useful for supervised node labelling/prediction tasks.
"""
# broadcast globals and attach them to node features
broadcasted_globals = gn.utils_tf.repeat(graph_features.globals,
graph_features.n_left_nodes,
axis=0)
left_nodes = tf.concat([graph_features.left_nodes, broadcasted_globals], axis=-1)
# get logits over nodes
preds = self.supervised_prediction_torso(left_nodes)
# remove the final singleton dimension
preds = tf.squeeze(preds, axis=-1)
indices = gn.utils_tf.sparse_to_dense_indices(graph_features.n_left_nodes)
opt_sol = tf.gather_nd(indices, obs['optimal_solution'], infer_shape(preds))
auxiliary_loss = tf.reduce_mean((preds - opt_sol)**2)
return auxiliary_loss
def body_fn(i, decoder_inputs, logitss, actions, node_mask):
dec = self._trans.decode(decoder_inputs, memory, src_masks,
True) # (N, T2, d_model)
# Q -> (N, d_model)
Q = self._out_projection_mlp(dec[:, -1])
# Q -> (N, 1, d_model)
Q = tf.expand_dims(Q, 1)
# dot product
outputs = tf.matmul(Q, tf.transpose(xs, [0, 2, 1])) # (N, 1, T1)
outputs = tf.squeeze(outputs, 1) # (N, T1)
# scale
outputs /= (Q.get_shape().as_list()[-1]**0.5)
# [N, T1]
logits = tf.where(node_mask, outputs,
tf.fill(tf.shape(node_mask), np.float32(-1e9)))
# dont sample if sampled_actions is provided.
assert sampled_actions is None
act = sample_from_logits(logits, self.seed) # (N,)
logitss = tf.concat(
[logitss, tf.expand_dims(logits, axis=1)], axis=1)
actions = tf.concat([actions, tf.expand_dims(act, axis=1)],
axis=-1)
# update node_masks to remove the current selected node for the next
# decoding iteration.
indices = tf.stack([tf.range(N), act], axis=-1)
action_mask = tf.scatter_nd(indices, tf.ones((N, ), tf.bool),
infer_shape(node_mask))
node_mask = tf.logical_and(node_mask, tf.logical_not(action_mask))
embs = tf.gather_nd(xs, indices) # (N, d)
embs = tf.expand_dims(embs, 1) # (N, 1, d)
decoder_inputs = tf.concat((decoder_inputs, embs), 1)
return i + 1, decoder_inputs, logitss, actions, node_mask
def calc_logits(decoder_inputs, node_mask):
# calculate logits with sampled actions.
dec = self._trans.decode(decoder_inputs, memory, src_masks,
True) # (N, T2, d_model)
# Q -> (N, T2, d_model)
Q = snt.BatchApply(self._out_projection_mlp)(dec)
# dot product
outputs = tf.matmul(Q, tf.transpose(xs,
[0, 2, 1])) # (N, T2, T1)
# scale
outputs /= (Q.get_shape().as_list()[-1]**0.5)
# (N, 1, T1)
node_mask = tf.expand_dims(node_mask, 1)
# (N, T2, T1)
node_mask = tf.tile(node_mask, [1, infer_shape(outputs)[1], 1])
# [N, T2, T1]
logits = tf.where(
node_mask, outputs,
tf.fill(tf.shape(node_mask), np.float32(-1e9)))
return logits
def get_actions(self,
graph_embeddings,
obs,
step_types=None,
sampled_actions=None,
log_features=False):
# step_types: [T + 1, B] of stepTypes
# If actions ar eprovided, then sampling is ommitted.
"""
Difference between src_masks, node_mask and action_mask
src_masks: Indicates that certain node inputs to the encoder
are from padding and should not be considered.
node_masks: Sampling should happen only at these nodes during
decoding phase.
action_mask: Masks where the previous action has taken place.
"""
xs, n_node = self.get_node_embeddings(obs,
graph_embeddings) # (N, L1, d)
node_mask = tf.reshape(obs['node_mask'], infer_shape(xs)[:-1])
node_mask = tf.cast(node_mask, tf.bool)
n_actions = tf.reshape(obs['n_actions'], (infer_shape(xs)[0], ))
if sampled_actions is not None:
# Remove some of the samples from the batch to match the size of the
# actions.
# This happens because the graph_embeddings are calculated for T + 1
# observations whereas actions are only avaiable for T steps.
# (T + 1) * B -> T * B
t_times_b = infer_shape(sampled_actions)[0]
xs = xs[:t_times_b]
n_node = n_node[:t_times_b]
node_mask = node_mask[:t_times_b]
n_actions = n_actions[:t_times_b]
N = infer_shape(xs)[0]
T1 = infer_shape(xs)[1] # T1 and L1 used interchangebly.
# compute src_mask to remove padding nodes interfering.
indices = gn.utils_tf.sparse_to_dense_indices(n_node)
# src_masks -> (N, L1)
src_masks = tf.scatter_nd(indices,
tf.ones(infer_shape(indices)[:1], tf.bool),
infer_shape(xs)[:2])
if sampled_actions is not None:
log_features = dict()
log_features.update(
dict(xs=xs, src_masks=src_masks, step_types=step_types[:-1]))
log_features.update(dict(actions=sampled_actions))
xs = snt.BatchApply(self._node_emb_mlp)(xs) # (N, L1, d_model)
memory = xs
d = self.config.d_model
def cond_fn(i, *_):
# until all samples have n_actions sampled.
return i < tf.reduce_max(n_actions)
def body_fn(i, decoder_inputs, logitss, actions, node_mask):
dec = self._trans.decode(decoder_inputs, memory, src_masks,
True) # (N, T2, d_model)
# Q -> (N, d_model)
Q = self._out_projection_mlp(dec[:, -1])
# Q -> (N, 1, d_model)
Q = tf.expand_dims(Q, 1)
# dot product
outputs = tf.matmul(Q, tf.transpose(xs, [0, 2, 1])) # (N, 1, T1)
outputs = tf.squeeze(outputs, 1) # (N, T1)
# scale
outputs /= (Q.get_shape().as_list()[-1]**0.5)
# [N, T1]
logits = tf.where(node_mask, outputs,
tf.fill(tf.shape(node_mask), np.float32(-1e9)))
# dont sample if sampled_actions is provided.
assert sampled_actions is None
act = sample_from_logits(logits, self.seed) # (N,)
logitss = tf.concat(
[logitss, tf.expand_dims(logits, axis=1)], axis=1)
actions = tf.concat([actions, tf.expand_dims(act, axis=1)],
axis=-1)
# update node_masks to remove the current selected node for the next
# decoding iteration.
indices = tf.stack([tf.range(N), act], axis=-1)
action_mask = tf.scatter_nd(indices, tf.ones((N, ), tf.bool),
infer_shape(node_mask))
node_mask = tf.logical_and(node_mask, tf.logical_not(action_mask))
embs = tf.gather_nd(xs, indices) # (N, d)
embs = tf.expand_dims(embs, 1) # (N, 1, d)
decoder_inputs = tf.concat((decoder_inputs, embs), 1)
return i + 1, decoder_inputs, logitss, actions, node_mask
if sampled_actions is None:
i = tf.constant(0)
logitss = tf.constant([], shape=(N, 0, T1), dtype=tf.float32)
actions = tf.constant([], shape=(N, 0), dtype=tf.int32)
# Feed in zero embedding as the start sentinel.
decoder_inputs = tf.zeros((N, 1, d), tf.float32) # (N, 1, d)
i, _, logits, actions, _ = tf.while_loop(
cond_fn,
body_fn,
(i, decoder_inputs, logitss, actions, node_mask),
shape_invariants=(
i.get_shape(),
tf.TensorShape([N, None, d]),
tf.TensorShape([N, None, T1]),
tf.TensorShape([N, None]),
node_mask.get_shape(),
),
return_same_structure=True,
back_prop=False,
)
else:
def calc_logits(decoder_inputs, node_mask):
# calculate logits with sampled actions.
dec = self._trans.decode(decoder_inputs, memory, src_masks,
True) # (N, T2, d_model)
# Q -> (N, T2, d_model)
Q = snt.BatchApply(self._out_projection_mlp)(dec)
# dot product
outputs = tf.matmul(Q, tf.transpose(xs,
[0, 2, 1])) # (N, T2, T1)
# scale
outputs /= (Q.get_shape().as_list()[-1]**0.5)
# (N, 1, T1)
node_mask = tf.expand_dims(node_mask, 1)
# (N, T2, T1)
node_mask = tf.tile(node_mask, [1, infer_shape(outputs)[1], 1])
# [N, T2, T1]
logits = tf.where(
node_mask, outputs,
tf.fill(tf.shape(node_mask), np.float32(-1e9)))
return logits
decoder_inputs = tf.gather_nd(xs,
tf.expand_dims(sampled_actions, -1),
batch_dims=1)
logits = calc_logits(decoder_inputs, node_mask)
actions = sampled_actions
# Finally logits -> (N, T2, T1)
# Finally actions -> (N, T2)
# now pad actions -> (N, max_k) and logits -> (N, max_k, T1)
pad_d = obs['max_k'][0] - infer_shape(logits)[1]
logits = tf.pad(logits, ([0, 0], [0, pad_d], [0, 0]))
pad_d = obs['max_k'][0] - infer_shape(actions)[1]
actions = tf.pad(actions, ([0, 0], [0, pad_d]))
if log_features:
return logits, actions, log_features
return logits, actions