def testFeedPartialShapes(self):
with self.test_session(use_gpu=False):
# Incorporate new rank into shape information if known
sp_input = self._SparseTensorPlaceholder()
sp_output = tf.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[None, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[3])
# Incorporate known shape information about input indices in output
# indices
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
sp_output = tf.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[5, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[3])
# Even if new_shape has no shape information, we know the ranks of
# output indices and shape
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
new_shape = tf.placeholder(tf.int64)
sp_output = tf.sparse_reshape(sp_input, new_shape)
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[5, None])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[None])
def lookup_char_emb(text,c2v_vocab, c2v_emb, dim_c2v_emb):
str_tensor = tf.string_split(text)
str_split = tf.sparse_reshape(str_tensor,[-1])
str_split,text_mask = tf.sparse_fill_empty_rows(str_split,"")
#return str_split
#str_split = tf.sparse_tensor_to_dense(str_split,default_value="")
#char_split = tf.string_split(str_split.values,'')
char_split = tf.string_split(str_split.values,'')
#return char_split
#return tf.SparseTensor(indices=tf.stack([str_split.indices,res.indices],axis=1),values = res.values, dense_shape=tf.stack([str_split.dense_shape[0],str_split.dense_shape[1],res.dense_shape[0], res.dense_shape[1]]))
#return char_split
#char_tensor_indices = tf.transpose(tf.stack([tf.gather(str_split.indices[:,0],char_split.indices[:,0]),tf.gather(str_split.indices[:,1],char_split.indices[:,0])]))
char_tensor = tf.SparseTensor(indices = char_split.indices, values = c2v_vocab.lookup(char_split.values), dense_shape = char_split.dense_shape)
#return char_tensor
#char_tensor = tf.SparseTensor(indices = char_split.indices, values = char_dict.lookup(char_split.values), dense_shape = char_split.dense_shape)
char_tensor_reshape = tf.sparse_reshape(char_tensor,[-1])
char_tensor,term_mask = tf.sparse_fill_empty_rows(char_tensor_reshape,0)
#return char_tensor
char_vecs = tf.nn.embedding_lookup_sparse(c2v_emb, char_tensor, None, combiner='sum')
char_vecs = tf.where(~term_mask, char_vecs, tf.zeros_like(char_vecs))
#return char_vecs
term_char_vecs = tf.reshape(char_vecs, shape = tf.stack([tf.shape(text)[0],tf.cast(tf.reduce_max(str_tensor.indices[:,1])+1,tf.int32),-1,tf.shape(char_vecs)[-1]]))
term_char_mask_tmp = tf.reduce_sum(term_char_vecs,axis=-1)
term_char_mask = ~tf.equal(term_char_mask_tmp,0)
term_char_len = tf.cast(tf.count_nonzero(term_char_mask,axis=-1),tf.int32)
text_mask = ~tf.equal(tf.reduce_sum(term_char_mask_tmp,axis=-1),0)
text_len = tf.cast(tf.count_nonzero(text_mask,axis=-1),tf.int32)
return term_char_vecs, term_char_mask, term_char_len, text_mask, text_len
def sparse_loss_to_node(self, samples, support_size, num_samples):
batch_size = self.batch_size
length = sum(support_size[1:])*batch_size
node_dim = self.loss_node.get_shape().as_list()
#discount = .9
for k in range(1, 2):
#for k in range(1, len(samples)):
#import pdb
#pdb.set_trace()
x = tf.reshape(tf.tile(tf.expand_dims(samples[k-1], -1), [1, tf.cast(support_size[k]/support_size[k-1], tf.int32)]), [-1])
x = tf.cast(x, tf.int64)
y = samples[k]
y = tf.cast(y, tf.int64)
idx = tf.expand_dims(x*node_dim[0] + y,1)
#loss = (discount**(k-1))*tf.reshape(tf.tile(tf.expand_dims(tf.reduce_sum(self.cross_entropy, 1), -1), [1, support_size[k]]), [-1])
loss = tf.reshape(tf.tile(tf.expand_dims(self.loss_node_, -1), [1, support_size[k]]), [-1])
scatter1 = tf.SparseTensor(idx, loss, tf.constant([node_dim[0]*node_dim[1]], dtype=tf.int64))
scatter1 = tf.sparse_reshape(scatter1, tf.constant([node_dim[0], node_dim[1]]))
self.loss_node = tf.sparse_add(self.loss_node, scatter1)
ones = tf.reshape(tf.tile(tf.expand_dims(tf.ones(batch_size), -1), [1, support_size[k]]), [-1])
scatter2 = tf.SparseTensor(idx, ones, tf.constant([node_dim[0]*node_dim[1]], dtype=tf.int64))
scatter2 = tf.sparse_reshape(scatter2, tf.constant([node_dim[0], node_dim[1]]))
self.loss_node_count = tf.sparse_add(self.loss_node_count, scatter2)
def sp_attn_head(seq,
out_sz,
adj_mat_local,
adj_mat_global,
activation,
in_drop=0.0,
coef_drop=0.0,
residual=False):
with tf.name_scope('my_attn'):
if in_drop != 0.0:
seq = tf.nn.dropout(seq, 1.0 - in_drop)
seq_fts = seq
latent_factor_size = 8
nb_nodes = seq_fts.shape[1].value
w_1 = glorot([seq_fts.shape[2].value, latent_factor_size])
w_2 = glorot([3 * seq_fts.shape[2].value, latent_factor_size])
f_1 = tf.layers.conv1d(seq_fts, 1, 1)
f_2 = tf.layers.conv1d(seq_fts, 1, 1)
#local neighbours
logits = tf.add(f_1[0], tf.transpose(f_2[0]))
logits_first = adj_mat_local * logits
lrelu = tf.SparseTensor(indices=logits_first.indices,
values=tf.nn.leaky_relu(logits_first.values),
dense_shape=logits_first.dense_shape)
coefs = tf.sparse_softmax(lrelu)
coefs = tf.sparse_reshape(coefs, [nb_nodes, nb_nodes])
seq_fts = tf.squeeze(seq_fts)
neigh_embs = tf.sparse.sparse_dense_matmul(coefs, seq_fts)
#non-local neighbours
logits_global = adj_mat_global * logits
lrelu_global = tf.SparseTensor(indices=logits_global.indices,
values=tf.nn.leaky_relu(
logits_global.values),
dense_shape=logits_global.dense_shape)
coefs_global = tf.sparse_softmax(lrelu_global)
coefs_global = tf.sparse_reshape(coefs_global, [nb_nodes, nb_nodes])
neigh_embs_global = tf.sparse.sparse_dense_matmul(
coefs_global, seq_fts)
neigh_embs_sum_1 = tf.matmul(
tf.add(tf.add(seq_fts, neigh_embs), neigh_embs_global), w_1)
neigh_embs_sum_2 = tf.matmul(
tf.concat(
[tf.concat([seq_fts, neigh_embs], axis=-1), neigh_embs_global],
axis=-1), w_2)
final_embs = activation(neigh_embs_sum_1) + activation(
neigh_embs_sum_2)
return final_embs
def calc_log_loss(self, Pairwise, Question, Answer, Review, TermtoTermR,
TermtoTermP, Question_I, Answer_I, Review_I):
#print 'Doing for item %d'%(i)
shape1 = tf.shape(Pairwise)
shape2 = tf.shape(Answer)
nq = shape1[0]
nr = shape1[1]
na = shape2[1]
pairwise = tf.reshape(Pairwise, [-1, self.PairwiseDim])
pairwise = tf.reshape(tf.matmul(pairwise, self.theta), [nq, nr])
termTotermR = tf.sparse_reshape(TermtoTermR, [-1, self.V])
termTotermR = tf.reshape(
tf.sparse_tensor_dense_matmul(termTotermR, self.RelvPar), [nq, nr])
QProj = tf.sparse_tensor_dense_matmul(Question_I, self.A)
RProjR = tf.sparse_tensor_dense_matmul(Review_I, self.B)
BilinearR = tf.matmul(QProj, tf.transpose(RProjR))
Relevance = tf.nn.softmax(pairwise + termTotermR + BilinearR)
termTotermP = tf.sparse_reshape(TermtoTermP, [-1, self.V])
termTotermP = tf.reshape(
tf.sparse_tensor_dense_matmul(termTotermP, self.PredPar),
[nq, na, nr])
AProj = tf.sparse_tensor_dense_matmul(
tf.sparse_reshape(Answer_I, [-1, self.V]), self.X)
RProjP = tf.sparse_tensor_dense_matmul(Review_I, self.Y)
BilinearP = tf.reshape(tf.matmul(AProj, tf.transpose(RProjP)),
[nq, na, nr])
Prediction = BilinearP + termTotermP
Prediction = tf.expand_dims(Prediction[:, 0, :], 1) - Prediction
Prediction = Prediction[:, 1:, :]
Prediction = tf.sigmoid(Prediction)
MoE = tf.reduce_sum(tf.multiply(Prediction,
tf.expand_dims(Relevance, axis=1)),
axis=2)
accuracy_count = tf.cast(tf.shape(tf.where(MoE > 0.5))[0], tf.float64)
count = nq * na
log_likelihood = tf.reduce_sum(tf.log(MoE))
R1 = tf.reduce_sum(tf.square(self.A)) + tf.reduce_sum(tf.square(
self.B))
R2 = tf.reduce_sum(tf.square(self.X)) + tf.reduce_sum(tf.square(
self.Y))
log_likelihood -= self.Lambda * (R1 + R2)
return -1 * log_likelihood, MoE, Relevance
def train_filequeue_reader(self, filename_queue):
(keys, values) = self.train_reader.read_up_to(filename_queue,
self.config.batch_size)
label_features = {
"label_indices": tf.VarLenFeature(dtype=tf.int64),
"label_values": tf.VarLenFeature(dtype=tf.int64),
"label_shape": tf.FixedLenFeature([1], dtype=tf.int64),
"seq_len": tf.FixedLenFeature([1], dtype=tf.int64)
}
audio_features = {
"audio": tf.FixedLenSequenceFeature([self.last_dim],
dtype=tf.float32)
}
audio_list = []
label_list = []
len_list = []
for i in range(self.config.batch_size):
context, sequence = tf.parse_single_sequence_example(
serialized=values[i],
context_features=label_features,
sequence_features=audio_features
)
audio = sequence['audio']
seq_len = context['seq_len']
label_values = context['label_values']
label_indices = context['label_indices']
label_shape = context['label_shape']
label_indices = tf.sparse_tensor_to_dense(
tf.sparse_reshape(label_indices, [-1, 1]))
label_values = tf.sparse_tensor_to_dense(label_values)
sparse_label = tf.SparseTensor(label_indices, label_values,
label_shape)
# then we can sparse_concat at axis 0
sparse_label = tf.sparse_reshape(sparse_label, [1, -1])
audio_list.append(audio)
label_list.append(sparse_label)
len_list.append(seq_len)
seq_lengths = tf.cast(tf.reshape(tf.stack(len_list, name='seq_lengths'),
(-1,)), tf.int32)
audio_tensor = tf.stack(audio_list, name='input_audio')
label_tensor = tf.sparse_concat(0, label_list,
expand_nonconcat_dim=True)
return audio_tensor, label_tensor, seq_lengths
def _sparse_tensor_dense_matmul(sp_a, b, **kwargs):
"""Returns (batched) matmul of a SparseTensor with a Tensor.
Args:
sp_a: `SparseTensor` representing a (batch of) matrices.
b: `Tensor` representing a (batch of) matrices, with the same batch shape of
`sp_a`. The shape must be compatible with the shape of `sp_a` and kwargs.
**kwargs: Keyword arguments to `tf.sparse_tensor_dense_matmul`.
Returns:
product: A dense (batch of) matrix-shaped Tensor of the same batch shape and
dtype as `sp_a` and `b`. If `sp_a` or `b` is adjointed through `kwargs` then
the shape is adjusted accordingly.
"""
batch_shape = _get_shape(sp_a)[:-2]
# Reshape the SparseTensor into a rank 3 SparseTensors, with the
# batch shape flattened to a single dimension. If the batch rank is 0, then
# we add a batch dimension of rank 1.
sp_a = tf.sparse_reshape(sp_a,
tf.concat([[-1], _get_shape(sp_a)[-2:]], axis=0))
# Reshape b to stack the batch dimension along the rows.
b = tf.reshape(b, tf.concat([[-1], _get_shape(b)[-1:]], axis=0))
# Convert the SparseTensor to a matrix in block diagonal form with blocks of
# matrices [M, N]. This allow us to use tf.sparse_tensor_dense_matmul which
# only accepts rank 2 (Sparse)Tensors.
out = tf.sparse_tensor_dense_matmul(_sparse_block_diag(sp_a), b, **kwargs)
# Finally retrieve the original batch shape from the resulting rank 2 Tensor.
# Note that we avoid inferring the final shape from `sp_a` or `b` because we
# might have transposed one or both of them.
return tf.reshape(
out,
tf.concat([batch_shape, [-1], _get_shape(out)[-1:]], axis=0))
def call(self, words, **kwargs):
#e.g [batch_size,sent_number,seq_len]
context_ids = self.feature_lookup_table.lookup(words)
if isinstance(context_ids, tf.SparseTensor):
context_ids = tf.sparse_tensor_to_dense(context_ids,default_value=0,name = "sparseid2dense_C")
#[batch_size,sent_number,c_seq_len], every sentence real length
mask = tf.abs(tf.sign(context_ids))
#[batch_size, sent_number]
sen_len = tf.reduce_sum(mask, -1)
#[batch_size,sent_number,c_seq_len,1],
mask = tf.expand_dims(tf.cast(mask,tf.float32),-1)
context_embedding = tf.nn.embedding_lookup(self.word_embedding,context_ids)
context_embedding = tf.layers.dropout(context_embedding,self.dropout_rate,training=self.trainable)
if self.use_char_embedding :
chars = kwargs.get("chars")
shape = tf.shape(chars)
if isinstance(chars, tf.SparseTensor):
chars= tf.sparse_reshape(chars,[-1])
chars = tf.sparse_tensor_to_dense(chars,default_value="0")
else:
chars = tf.reshape(chars,[-1])
chars = tf.string_split(chars,delimiter=",")
chars_ids = self.char_lookup_table.lookup(chars)
chars_ids = tf.sparse_tensor_to_dense(chars_ids)
char_mask = tf.cast(tf.expand_dims(tf.sign(chars_ids),-1),tf.float32)
char_embedding = tf.nn.embedding_lookup(self.char_embeddding,chars_ids)*char_mask #context_mask cares about this
char_embedding = tf.layers.dropout(char_embedding,self.dropout_rate*0.5,training=self.trainable)
char_embedding = self.conv_chars(char_embedding,shape)
context_embedding = tf.concat([context_embedding,char_embedding],-1)
context_embedding = self.highway(context_embedding,mask = mask) if self.use_highway else context_embedding*mask
return context_embedding,sen_len,mask,context_ids
def map_function(x):
i, dense_slice = x[0], x[1]
sparse_slice = tf.sparse_reshape(tf.sparse_slice(sp_a, [i, 0, 0],
[1, sp_a.dense_shape[1], sp_a.dense_shape[2]]),
[sp_a.dense_shape[1], sp_a.dense_shape[2]])
mult_slice = tf.sparse_tensor_dense_matmul(sparse_slice, dense_slice)
return mult_slice
def testMismatchedSizesWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, -1])
with self.assertRaisesOpError("requested shape requires a multiple"):
sess.run(sp_output, {sp_input: input_val})
def testMultipleInferredDims(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, -1, -1])
with self.assertRaisesOpError("only one output shape size may be -1"):
sess.run(sp_output, {sp_input: input_val})
def labels_to_onehot(lables, class_num = 1):
'''
:param lables: shape [batchsize, depth, height, width], 4D, no channel axis
:param class_num:
:return:
'''
if isinstance(class_num, tf.Tensor):
class_num_tf = tf.to_int32(class_num)
else:
class_num_tf = tf.constant(class_num, tf.int32)
in_shape = tf.shape(lables)
out_shape = tf.concat([in_shape, tf.reshape(class_num_tf, (1,))], 0) # add a extra axis for classNum, 5D
if class_num == 1:
return tf.reshape(lables, out_shape)
else:
lables = tf.reshape(lables, (-1,)) # squeeze labels to one row x N cols vector [0,0,0,1,......]
dense_shape = tf.stack([tf.shape(lables)[0], class_num_tf], 0) # denshape [N cols , classNum]
lables = tf.to_int64(lables)
ids = tf.range(tf.to_int64(dense_shape[0]), dtype= tf.int64) # ids is a 1xN vector as[0,1,2,3...., N-1]
ids = tf.stack([ids, lables], axis= 1) #ids is N x clsNum mat
one_hot = tf.SparseTensor(indices= ids, values= tf.ones_like(lables, dtype= tf.float32), dense_shape = tf.to_int64(dense_shape))
one_hot = tf.sparse_reshape(one_hot, out_shape)
return tf.cast(one_hot, tf.float32)
def read_and_decode(filename_queue):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example,
features={
'rows':
tf.FixedLenFeature([], tf.int64),
'label':
tf.FixedLenFeature([], tf.int64),
'one_feature':
tf.VarLenFeature(tf.float32)
})
one_feature = features['one_feature']
# The code below does not work, you just cannot cast a SparseTensor
# one_feature = tf.cast(features['one_feature'], tf.float32)
"""
for CNN feature: n*4096
"""
# one_feature = tf.sparse_reshape(one_feature, [-1, 4096])
"""
for gist feature: n*512
"""
one_feature = tf.sparse_reshape(one_feature, [-1, 512])
label = tf.cast(features['label'], tf.int32)
rows = tf.cast(features['rows'], tf.int32)
return one_feature, label, rows
def _get_sequence_dense_tensor(self,
inputs,
weight_collections=None,
trainable=None):
del weight_collections, trainable
if self.signature != 'sequence':
raise ValueError(
'Column {} could not be used as sequence feature column. '
'Use sequence_text_embedding_column instead'.format(self.name))
sparse_keys = inputs.get(self)
sparse_keys.shape.with_rank_at_least(2)
sparse_keys.shape.with_rank_at_most(3)
batch_size, max_length = sparse_keys.dense_shape[
0], sparse_keys.dense_shape[1]
sparse_keys = tf.sparse_reshape(sparse_keys,
shape=tf.stack(
[batch_size, max_length, -1]))
dense_tensor = self._hub_module(sparse_keys, signature=self.signature)
sequence_length = feature_column._sequence_length_from_sparse_tensor(
inputs.get(self))
return feature_column._SequenceDenseColumn.TensorSequenceLengthPair(
dense_tensor=dense_tensor, sequence_length=sequence_length)
def tensors_to_item(self, keys_to_tensors):
"""Maps the given dictionary of tensors to a concatenated list of bboxes.
Args:
keys_to_tensors: a mapping of TF-Example keys to parsed tensors.
Returns:
[time, num_boxes, 4] tensor of bounding box coordinates, in order
[y_min, x_min, y_max, x_max]. Whether the tensor is a SparseTensor
or a dense Tensor is determined by the return_dense parameter. Empty
positions in the sparse tensor are filled with -1.0 values.
"""
sides = []
for key in self._full_keys:
value = keys_to_tensors[key]
expanded_dims = tf.concat(
[tf.to_int64(tf.shape(value)),
tf.constant([1], dtype=tf.int64)], 0)
side = tf.sparse_reshape(value, expanded_dims)
sides.append(side)
bounding_boxes = tf.sparse_concat(2, sides)
if self._return_dense:
bounding_boxes = tf.sparse_tensor_to_dense(
bounding_boxes, default_value=self._default_value)
return bounding_boxes
def _parse_function(self, index_list, sp_value, label):
inputs = tf.SparseTensor(indices=index_list,
values=sp_value,
dense_shape=self.sp_shape)
inputs = tf.sparse_reshape(inputs, [-1])
return inputs, label, index_list[:, -1]
def tensors_to_item(self, keys_to_tensors):
"""Maps the given dictionary of tensors to a concatenated list of bboxes.
Args:
keys_to_tensors: a mapping of TF-Example keys to parsed tensors.
Returns:
[time, num_boxes, 4] tensor of bounding box coordinates, in order
[y_min, x_min, y_max, x_max]. Whether the tensor is a SparseTensor
or a dense Tensor is determined by the return_dense parameter. Empty
positions in the sparse tensor are filled with -1.0 values.
"""
sides = []
for key in self._full_keys:
value = keys_to_tensors[key]
expanded_dims = tf.concat([
tf.to_int64(tf.shape(value)),
tf.constant([1], dtype=tf.int64)
], 0)
side = tf.sparse_reshape(value, expanded_dims)
sides.append(side)
bounding_boxes = tf.sparse_concat(2, sides)
if self._return_dense:
bounding_boxes = tf.sparse_tensor_to_dense(
bounding_boxes, default_value=self._default_value)
return bounding_boxes
def _do_test(self, expected_result, config=None):
# Make sure that expected_result is an np array
if not type(expected_result).__module__ == np.__name__:
expected_result = np.array(expected_result)
with tf.Graph().as_default():
inputs, indices, _ = dense_to_sparse(self.input)
sparse_tensor_reordered = tf.sparse_reorder(inputs)
sparse_tensor_reshaped = tf.sparse_reshape(sparse_tensor_reordered, self.input.shape)
W = tf.constant(self.W1, dtype=tf.float32)
b = tf.constant(self.b1, dtype=tf.float32)
# Sparse layer
logits = tf.sparse_tensor_dense_matmul(sparse_tensor_reshaped, W) + b
# Dense layer
logits = logits @ tf.constant(self.W2, tf.float32) + tf.constant(self.b2, tf.float32)
explanation = lrp.lrp(inputs, logits, config)
with tf.Session() as s:
expl = s.run(explanation)
self.assertTrue(np.all(np.equal(indices, expl.indices)),
"expected indices did not equal actual indices")
self.assertTrue(np.allclose(expl.values, expected_result.reshape((-1)), rtol=1.e-3, atol=1.e-3),
"expected indices did not equal actual indices")
def testFeedMultipleInferredDims(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, -1, -1])
with self.assertRaisesOpError("only one output shape size may be -1"):
sess.run(sp_output, {sp_input: input_val})
def compute_per_batch_eval_metrics(metrics, predictions, used_labels, labels, indices,
used_indices, ori_batch_idx, step_idx, config, context):
new_batch_idx = tf.gather(tf.range(tf.shape(used_indices)[0]), ori_batch_idx)
new_indices = tf.concat([
tf.expand_dims(new_batch_idx, 1),
tf.expand_dims(step_idx, 1),
tf.slice(indices,
begin=[0, 2],
size=[tf.shape(indices)[0], 1])],
axis=1)
label_shape = tf.shape(labels)
eval_labels = tf.sparse_reshape(
tf.SparseTensor(
tf.cast(new_indices, tf.int64),
tf.cast(used_labels, tf.int64),
tf.cast(
[tf.shape(used_indices)[0], label_shape[1], label_shape[2]],
tf.int64)
),
[tf.shape(used_indices)[0], label_shape[1] * label_shape[2]])
ori_predictions = tf.gather(predictions, ori_batch_idx)
updates = compute_recommendation_metrics(
metrics, eval_labels, predictions)
updates += compute_regression_metrics(
metrics, ori_predictions, used_labels, indices, config, context)
updates += compute_ranking_metrics(
metrics, used_indices, new_batch_idx, used_labels, predictions, config)
return updates
def testFeedDenseReshapeSemantics(self):
with self.test_session(use_gpu=False) as sess:
# Compute a random rank-5 initial shape and new shape, randomly sparsify
# it, and check that the output of SparseReshape has the same semantics
# as a dense reshape.
factors = np.array([2] * 4 + [3] * 4 +
[5] * 4) # 810k total elements
orig_rank = np.random.randint(2, 7)
orig_map = np.random.randint(orig_rank, size=factors.shape)
orig_shape = [
np.prod(factors[orig_map == d]) for d in range(orig_rank)
]
new_rank = np.random.randint(2, 7)
new_map = np.random.randint(new_rank, size=factors.shape)
new_shape = [
np.prod(factors[new_map == d]) for d in range(new_rank)
]
orig_dense = np.random.uniform(size=orig_shape)
orig_indices = np.transpose(np.nonzero(orig_dense < 0.5))
orig_values = orig_dense[orig_dense < 0.5]
new_dense = np.reshape(orig_dense, new_shape)
new_indices = np.transpose(np.nonzero(new_dense < 0.5))
new_values = new_dense[new_dense < 0.5]
sp_input = self._SparseTensorPlaceholder()
input_val = tf.SparseTensorValue(orig_indices, orig_values,
orig_shape)
sp_output = tf.sparse_reshape(sp_input, new_shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, new_indices)
self.assertAllEqual(output_val.values, new_values)
self.assertAllEqual(output_val.shape, new_shape)
def testFeedMismatchedSizesWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, -1])
with self.assertRaisesOpError("requested shape requires a multiple"):
sess.run(sp_output, {sp_input: input_val})
def sp_attn_head(seq,
out_sz,
adj_mat,
adj_all_mat,
adj_neig_mat,
N_target_mat,
activation,
nb_nodes,
in_drop=0.0,
coef_drop=0.0,
residual=False):
with tf.name_scope('sp_attn'):
if coef_drop != 0.0:
adj_mat = tf.SparseTensor(indices=adj_mat.indices,
values=tf.nn.dropout(
adj_mat.values, 1.0 - coef_drop),
dense_shape=adj_mat.dense_shape)
adj_neig_mat = tf.SparseTensor(
indices=adj_neig_mat.indices,
values=tf.nn.dropout(adj_neig_mat.values, 1.0 - coef_drop),
dense_shape=adj_neig_mat.dense_shape)
if in_drop != 0.0:
seq = tf.nn.dropout(seq, 1.0 - in_drop)
seq_fts = tf.layers.conv1d(seq, out_sz, 1, use_bias=False)
# simplest self-attention possible
f_1 = tf.layers.conv1d(seq_fts, 1, 1)
f_2 = tf.layers.conv1d(seq_fts, 1, 1)
f_1 = tf.reshape(f_1, (nb_nodes, 1))
f_2 = tf.reshape(f_2, (nb_nodes, 1))
f_1 = adj_mat * f_1
f_2 = adj_mat * tf.transpose(f_2, [1, 0])
logits = tf.sparse_add(f_1, f_2)
lrelu = tf.SparseTensor(indices=logits.indices,
values=tf.nn.leaky_relu(logits.values),
dense_shape=logits.dense_shape)
coefs = tf.sparse_softmax(lrelu)
if in_drop != 0.0:
seq_fts = tf.nn.dropout(seq_fts, 1.0 - in_drop)
coefs = tf.sparse_reshape(coefs, [nb_nodes, nb_nodes])
seq_fts = tf.squeeze(seq_fts) ###HW
out_bi = BILinear_pooling(adj_neig_mat, seq_fts)
out_bi = dot(N_target_mat, out_bi, True)
out_gat = tf.sparse_tensor_dense_matmul(coefs, seq_fts)
vals = (1 - FLAGS.alpha) * out_gat + FLAGS.alpha * out_bi
vals = tf.expand_dims(vals, axis=0)
vals.set_shape([1, nb_nodes, out_sz])
ret = tf.contrib.layers.bias_add(vals)
return activation(ret) # activation
def _call(self, inputs):
seq_fts = tf.layers.conv1d(inputs, self.out_sz, 1, use_bias=False)
# simplest self-attention possible
f_1_t = tf.layers.conv1d(seq_fts, 1, 1)
f_2_t = tf.layers.conv1d(seq_fts, 1, 1)
f_1 = tf.reshape(f_1_t, (self.nb_nodes, 1))
f_2 = tf.reshape(f_2_t, (self.nb_nodes, 1))
f_1 = self.bias_mat * f_1
f_2 = self.bias_mat * tf.transpose(f_2, [1, 0])
logits = tf.sparse_add(f_1, f_2)
lrelu = tf.SparseTensor(indices=logits.indices,
values=tf.nn.leaky_relu(logits.values),
dense_shape=logits.dense_shape)
coefs = tf.sparse_softmax(lrelu)
# As tf.sparse_tensor_dense_matmul expects its arguments to have rank-2,
# here we make an assumption that our input is of batch size 1, and reshape appropriately.
# The method will fail in all other cases!
coefs = tf.sparse_reshape(coefs, [self.nb_nodes, self.nb_nodes])
seq_fts = tf.squeeze(seq_fts)
vals = tf.sparse_tensor_dense_matmul(coefs, seq_fts)
vals = tf.expand_dims(vals, axis=0)
vals.set_shape([1, self.nb_nodes, self.out_sz])
ret = self.act(tf.contrib.layers.bias_add(vals))
return ret # activation
def __call__(self, u_inputs, v_inputs, u_size, v_size):
x = v_inputs
adj_mat = self.adj_mat
# simplest self-attention possible
f_1 = tf.layers.conv1d(u_inputs, 1, 1)
f_2 = tf.layers.conv1d(v_inputs, 1, 1)
f_1 = tf.reshape(f_1, (u_size, 1))
f_2 = tf.reshape(f_2, (v_size, 1))
seq_fts = tf.layers.conv1d(x, self.output_dim, 1, use_bias=False)
f_1 = adj_mat * f_1
f_2 = adj_mat * tf.transpose(f_2, [1, 0])
logits = tf.sparse_add(f_1, f_2)
lrelu = tf.SparseTensor(indices=logits.indices,
values=tf.nn.leaky_relu(logits.values),
dense_shape=logits.dense_shape)
coefs = tf.sparse_softmax(lrelu)
coefs = tf.sparse_reshape(coefs, [u_size, v_size])
seq_fts = tf.squeeze(seq_fts)
vals = tf.sparse_tensor_dense_matmul(coefs, seq_fts)
print('--------vals.shape------', vals.shape)
# vals = tf.expand_dims(vals, axis=0)
# vals.set_shape([1, nb_nodes, out_sz])
ret = tf.contrib.layers.bias_add(vals)
return self.act(ret) # activation
def testFeedDenseReshapeSemantics(self):
with self.test_session(use_gpu=False) as sess:
# Compute a random rank-5 initial shape and new shape, randomly sparsify
# it, and check that the output of SparseReshape has the same semantics
# as a dense reshape.
factors = np.array([2] * 4 + [3] * 4 + [5] * 4) # 810k total elements
orig_rank = np.random.randint(2, 7)
orig_map = np.random.randint(orig_rank, size=factors.shape)
orig_shape = [np.prod(factors[orig_map == d]) for d in range(orig_rank)]
new_rank = np.random.randint(2, 7)
new_map = np.random.randint(new_rank, size=factors.shape)
new_shape = [np.prod(factors[new_map == d]) for d in range(new_rank)]
orig_dense = np.random.uniform(size=orig_shape)
orig_indices = np.transpose(np.nonzero(orig_dense < 0.5))
orig_values = orig_dense[orig_dense < 0.5]
new_dense = np.reshape(orig_dense, new_shape)
new_indices = np.transpose(np.nonzero(new_dense < 0.5))
new_values = new_dense[new_dense < 0.5]
sp_input = self._SparseTensorPlaceholder()
input_val = tf.SparseTensorValue(orig_indices, orig_values, orig_shape)
sp_output = tf.sparse_reshape(sp_input, new_shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, new_indices)
self.assertAllEqual(output_val.values, new_values)
self.assertAllEqual(output_val.shape, new_shape)
def read_and_decode(filename_queue):
"""
decode tfrecords
:param filename_queue: the filename
:return: one_feature, label, rows(image number of one scene spot, that is, rows of the data of a .npy file)
"""
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(serialized_example,
features={
'rows':
tf.FixedLenFeature([], tf.int64),
'label':
tf.FixedLenFeature([], tf.int64),
'one_feature':
tf.VarLenFeature(tf.float32)
})
one_feature = features['one_feature']
# The code below does not work, you just cannot cast a SparseTensor
# one_feature = tf.cast(features['one_feature'], tf.float32)
"""
for CNN feature: n*2048
"""
# one_feature = tf.sparse_reshape(one_feature, [-1, 2048])
"""
for gist feature: n*512
"""
one_feature = tf.sparse_reshape(one_feature, [-1, FLAGS.feature_col])
label = tf.cast(features['label'], tf.int32)
rows = tf.cast(features['rows'], tf.int32)
return one_feature, label, rows
def _sparse_or_dense_matmul_onehot(sparse_or_dense_matrix, col_index, size):
"""Returns a (dense) column of a Tensor or SparseTensor.
Args:
sparse_or_dense_matrix: matrix-shaped, `float` `Tensor` or `SparseTensor`.
col_index: scalar, `int` `Tensor` representing the index of the desired
column.
size: scalar, `int` `Tensor` representing the number of rows in
`sparse_or_dense_matrix`. Used only in the sparse case, so that the
caller can give side information about the shape of
`sparse_or_dense_matrix`.
Returns:
column: vector-shaped, `float` `Tensor` with the same dtype as
`sparse_or_dense_matrix`, representing the `col_index`th column of
`sparse_or_dense_matrix`.
"""
if isinstance(sparse_or_dense_matrix,
(tf.SparseTensor, tf.SparseTensorValue)):
# TODO(b/111924846): Implement better (ideally in a way that allows us to
# eliminate the `size` arg, if possible).
return tf.sparse_tensor_to_dense(
tf.sparse_reshape(
tf.sparse_slice(sparse_or_dense_matrix,
tf.cast([0, col_index], tf.int64),
tf.cast([size, 1], tf.int64)), [size]))
else:
return tf.gather(sparse_or_dense_matrix, col_index, axis=-1)
def sp_attn_head(seq,
out_sz,
adj_mat,
activation,
nb_nodes,
in_drop=0.0,
coef_drop=0.0,
residual=False):
with tf.name_scope('sp_attn'):
if in_drop != 0.0:
seq = tf.nn.dropout(seq, 1.0 - in_drop)
seq_fts = tf.layers.conv1d(seq, out_sz, 1, use_bias=False)
# simplest self-attention possible
f_1 = tf.layers.conv1d(seq_fts, 1, 1)
f_2 = tf.layers.conv1d(seq_fts, 1, 1)
f_1 = tf.reshape(f_1, (nb_nodes, 1))
f_2 = tf.reshape(f_2, (nb_nodes, 1))
f_1 = adj_mat * f_1
f_2 = adj_mat * tf.transpose(f_2, [1, 0])
logits = tf.sparse_add(f_1, f_2)
lrelu = tf.SparseTensor(indices=logits.indices,
values=tf.nn.leaky_relu(logits.values),
dense_shape=logits.dense_shape)
coefs = tf.sparse_softmax(lrelu)
if coef_drop != 0.0:
coefs = tf.SparseTensor(indices=coefs.indices,
values=tf.nn.dropout(
coefs.values, 1.0 - coef_drop),
dense_shape=coefs.dense_shape)
if in_drop != 0.0:
seq_fts = tf.nn.dropout(seq_fts, 1.0 - in_drop)
# As tf.sparse_tensor_dense_matmul expects its arguments to have rank-2,
# here we make an assumption that our input is of batch size 1, and reshape appropriately.
# The method will fail in all other cases!
coefs = tf.sparse_reshape(coefs, [nb_nodes, nb_nodes])
seq_fts = tf.squeeze(seq_fts)
vals = tf.sparse_tensor_dense_matmul(coefs, seq_fts)
vals = tf.expand_dims(vals, axis=0)
vals.set_shape([1, nb_nodes, out_sz])
ret = tf.contrib.layers.bias_add(vals)
# residual connection
if residual:
if seq.shape[-1] != ret.shape[-1]:
ret = ret + conv1d(seq, ret.shape[-1], 1) # activation
else:
seq_fts = ret + seq
if activation == None: ## for the final layer
return ret
else:
return activation(ret) # activation
def xletter_feature_extractor(text,model_prefix,input_mode, op_dict=None,xletter_cnt=None,win_size=None,dim_xletter_emb=None):
with tf.variable_scope("xletter_layer", reuse=tf.AUTO_REUSE):
if input_mode=='mstf':
xletter_emb = tf.get_variable(name='xletter_emb_' + model_prefix, shape = [xletter_cnt * win_size, dim_xletter_emb])
indices, ids, values, offsets = mstf.dssm_xletter(input=text, win_size=win_size, dict_handle=op_dict)
offsets_to_dense = tf.segment_sum(tf.ones_like(offsets), offsets)
batch_id = tf.cumsum(offsets_to_dense[:-1])
index_tensor = tf.concat([tf.expand_dims(batch_id,axis=-1), tf.expand_dims(indices,axis=-1)],axis=-1)
value_tensor = ids
dense_shape = tf.concat([tf.shape(offsets),tf.expand_dims(tf.reduce_max(indices) + 1,axis=-1)],axis=0)
text_tensor = tf.SparseTensor(indices=tf.cast(index_tensor,tf.int64), values = value_tensor, dense_shape=tf.cast(dense_shape,tf.int64))
#conv
text_tensor = tf.sparse_reshape(text_tensor,[-1])
text_tensor,text_mask = tf.sparse_fill_empty_rows(text_tensor,0)
text_vecs = tf.nn.embedding_lookup_sparse(xletter_emb,text_tensor,None,combiner='sum')
text_vecs = tf.where(~text_mask, text_vecs, tf.zeros_like(text_vecs))
text_vecs = tf.reshape(text_vecs,[-1,tf.reduce_max(indices) + 1,dim_xletter_emb])
step_mask = ~tf.equal(tf.reduce_sum(text_vecs,axis=2),0)
sequence_length = tf.cast(tf.count_nonzero(step_mask,axis=1),tf.int32)
elif input_mode=='pyfunc':
query_split = tf.string_split(text,';')
term_split = tf.string_split(query_split.values,',')
xletter_tensor_indices = tf.transpose(tf.stack([tf.gather(query_split.indices[:,0],term_split.indices[:,0]),tf.gather(query_split.indices[:,1],term_split.indices[:,0])]))
xletter_tensor = tf.SparseTensor(indices = xletter_tensor_indices, values = tf.string_to_number(term_split.values,out_type=tf.int32), dense_shape = query_split.dense_shape)
xletter_emb = tf.get_variable(name='xletter_emb_' + model_prefix, shape = [xletter_cnt * win_size, dim_xletter_emb])
xletter_tensor_reshape = tf.sparse_reshape(xletter_tensor,[-1])
xletter_tensor,text_mask = tf.sparse_fill_empty_rows(xletter_tensor_reshape,0)
xletter_vecs = tf.nn.embedding_lookup_sparse(xletter_emb, xletter_tensor, None, combiner='sum')
xletter_vecs = tf.where(~text_mask, xletter_vecs, tf.zeros_like(xletter_vecs))
text_vecs = tf.reshape(xletter_vecs, shape=tf.stack([-1,tf.reduce_max(query_split.indices[:,1])+1,dim_xletter_emb]))
step_mask = ~tf.equal(tf.reduce_sum(text_vecs,axis=2),0)
sequence_length = tf.cast(tf.count_nonzero(step_mask,axis=1),tf.int32)
elif input_mode=='pyfunc_batch':
indices, values, dense_shape = tf.py_func(op_dict.batch_xletter_extractor,[text],[tf.int64,tf.int32,tf.int64])
xletter_tensor = tf.SparseTensor(indices = indices, values = values, dense_shape = dense_shape)
xletter_emb = tf.get_variable(name='xletter_emb_' + model_prefix, shape = [xletter_cnt * win_size, dim_xletter_emb])
xletter_tensor_reshape = tf.sparse_reshape(xletter_tensor,[-1])
xletter_tensor,text_mask = tf.sparse_fill_empty_rows(xletter_tensor_reshape,0)
xletter_vecs = tf.nn.embedding_lookup_sparse(xletter_emb, xletter_tensor, None, combiner='sum')
xletter_vecs = tf.where(~text_mask, xletter_vecs, tf.zeros_like(xletter_vecs))
text_vecs = tf.reshape(xletter_vecs, shape=tf.stack([-1,dense_shape[1],dim_xletter_emb]))
step_mask = ~tf.equal(tf.reduce_sum(text_vecs,axis=2),0)
sequence_length = tf.cast(tf.count_nonzero(step_mask,axis=1),tf.int32)
else:
NotImplementedError
return text_vecs, step_mask, sequence_length
def testFeedMismatchedSizes(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, 7])
with self.assertRaisesOpError(
"Input to reshape is a tensor with 30 dense values"):
sess.run(sp_output, {sp_input: input_val})
def dot_product(params, rankings, x):
two_dims = LatentFactorPortfolioGraph.two_dim_shape(x, rankings)
x_two_dim = tf.sparse_reshape(x, shape=two_dims)
dot_prod = tf.sparse_tensor_dense_matmul(x_two_dim, params)
three_dims = LatentFactorPortfolioGraph.three_dim_shape(
params, rankings)
dot_prod_three_dim = tf.reshape(dot_prod, shape=three_dims)
return dot_prod_three_dim
def _attn_r_n_m_h(self):
h, r, n = self.heads, self.relations, self._nodes
attn_h_n_rm = self._attn_h_n_rm
attn_h_n_r_m = tf.sparse_reshape(attn_h_n_rm, [h, n, r, n])
attn_r_n_m_h = tf.sparse_transpose(attn_h_n_r_m, [2, 1, 3, 0])
return attn_r_n_m_h
def preprocessing_fn(inputs):
return {
'dense_out':
mappers.scale_to_0_1(inputs['dense_1']),
'sparse_out':
api.map(lambda x: tf.sparse_reshape(x, (1, 10)),
inputs['sparse'])
}
def testFeedMismatchedSizes(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [4, 7])
with self.assertRaisesOpError(
"Input to reshape is a tensor with 30 dense values"):
sess.run(sp_output, {sp_input: input_val})
def testSameShape(self):
with self.test_session(use_gpu=False) as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(input_val, [5, 6])
output_val = sess.run(sp_output)
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.shape, input_val.shape)
def testFeedSameShapeWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [-1, 6])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.shape, input_val.shape)
def create_word_vectors_from_post(self, raw_post, mxlen):
# vocab has only lowercase words
word2index = self.index
if self.do_lowercase:
raw_post = self.lowercase(raw_post)
word_tokens = tf.string_split(tf.reshape(raw_post, [-1]))
word_indices = word2index.lookup(word_tokens)
# Reshape them out to the proper length
reshaped_words = tf.sparse_reshape(word_indices, shape=[-1])
return self.reshape_indices(reshaped_words, [mxlen])
def _create_word_vectors_from_post_mixed_case(self, nraw_post, mxlen):
# vocab has only lowercase words
word_tokens = tf.string_split(tf.reshape(nraw_post, [-1]))
word_indices = self.word2index.lookup(word_tokens)
# Reshape them out to the proper length
reshaped_words = tf.sparse_reshape(word_indices, shape=[-1])
x = self._reshape_indices(reshaped_words, [mxlen])
return x
def testUpRank(self):
with self.test_session(use_gpu=False) as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(input_val, [2, 3, 5])
output_val = sess.run(sp_output)
self.assertAllEqual(output_val.indices, np.array([
[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0], [1, 1, 0], [1, 1, 1]
]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.shape, [2, 3, 5])
def testFeedUpRankWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [2, -1, 5])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, np.array([
[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0], [1, 1, 0], [1, 1, 1]
]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [2, 3, 5])
def testFeedNewShapeSameRankWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = tf.sparse_reshape(sp_input, [3, -1])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, np.array([
[0, 0], [0, 6], [0, 9], [1, 0], [2, 0], [2, 1]
]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.shape, [3, 10])
def testFeedDownRankWithInferredDim(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_2x3x4()
sp_output = tf.sparse_reshape(sp_input, [6, -1])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, np.array([
[0, 1], [1, 0], [1, 2], [3, 3], [4, 1], [4, 3], [5, 2]
]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.shape, [6, 4])
def testFeedPartialShapes(self):
with self.test_session(use_gpu=False):
# Incorporate new rank into shape information if known
sp_input = self._SparseTensorPlaceholder()
sp_output = tf.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(), [None, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [3])
# Incorporate known shape information about input indices in output
# indices
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
sp_output = tf.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(), [5, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [3])
# Even if new_shape has no shape information, we know the ranks of
# output indices and shape
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
new_shape = tf.placeholder(tf.int64)
sp_output = tf.sparse_reshape(sp_input, new_shape)
self.assertListEqual(sp_output.indices.get_shape().as_list(), [5, None])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [None])
def sparse_message_pass(node_states,
adjacency_matrices,
num_edge_types,
hidden_size,
use_bias=True,
average_aggregation=False,
name="sparse_ggnn"):
"""One message-passing step for a GNN with a sparse adjacency matrix.
Implements equation 2 (the message passing step) in
[Li et al. 2015](https://arxiv.org/abs/1511.05493).
N = The number of nodes in each batch.
H = The size of the hidden states.
T = The number of edge types.
Args:
node_states: Initial states of each node in the graph. Shape is [N, H].
adjacency_matrices: Adjacency matrix of directed edges for each edge
type. Shape is [N, N, T] (sparse tensor).
num_edge_types: The number of edge types. T.
hidden_size: The size of the hidden state. H.
use_bias: Whether to use bias in the hidden layer.
average_aggregation: How to aggregate the incoming node messages. If
average_aggregation is true, the messages are averaged. If it is false,
they are summed.
name: (optional) The scope within which tf variables should be created.
Returns:
The result of one step of Gated Graph Neural Network (GGNN) message passing.
Shape: [N, H]
"""
n = tf.shape(node_states)[0]
t = num_edge_types
incoming_edges_per_type = tf.sparse_reduce_sum(adjacency_matrices, axis=1)
# Convert the adjacency matrix into shape [T, N, N] - one [N, N] adjacency
# matrix for each edge type. Since sparse tensor multiplication only supports
# two-dimensional tensors, we actually convert the adjacency matrix into a
# [T * N, N] tensor.
adjacency_matrices = tf.sparse_transpose(adjacency_matrices, [2, 0, 1])
adjacency_matrices = tf.sparse_reshape(adjacency_matrices, [t * n, n])
# Multiply the adjacency matrix by the node states, producing a [T * N, H]
# tensor. For each (edge type, node) pair, this tensor stores the sum of
# the hidden states of the node's neighbors over incoming edges of that type.
messages = tf.sparse_tensor_dense_matmul(adjacency_matrices, node_states)
# Rearrange this tensor to have shape [N, T * H]. The incoming states of each
# nodes neighbors are summed by edge type and then concatenated together into
# a single T * H vector.
messages = tf.reshape(messages, [t, n, hidden_size])
messages = tf.transpose(messages, [1, 0, 2])
messages = tf.reshape(messages, [n, t * hidden_size])
# Run each of those T * H vectors through a linear layer that produces
# a vector of size H. This process is equivalent to running each H-sized
# vector through a separate linear layer for each edge type and then adding
# the results together.
#
# Note that, earlier on, we added together all of the states of neighbors
# that were connected by edges of the same edge type. Since addition and
# multiplying by a linear layer are commutative, this process was equivalent
# to running each incoming edge through a linear layer separately and then
# adding everything at the end.
with tf.variable_scope(name, default_name="sparse_ggnn"):
final_node_states = common_layers.dense(
messages, hidden_size, use_bias=False)
# Multiply the bias by for each edge type by the number of incoming nodes
# of that edge type.
if use_bias:
bias = tf.get_variable("bias", initializer=tf.zeros([t, hidden_size]))
final_node_states += tf.matmul(incoming_edges_per_type, bias)
if average_aggregation:
incoming_edges = tf.reduce_sum(incoming_edges_per_type, -1, keepdims=True)
incoming_edges = tf.tile(incoming_edges, [1, hidden_size])
final_node_states /= incoming_edges + 1e-7
return final_node_states