def generalised_dice_loss(prediction,
ground_truth,
weight_map=None,
type_weight='Square'):
"""
Function to calculate the Generalised Dice Loss defined in
Sudre, C. et. al. (2017) Generalised Dice overlap as a deep learning
loss function for highly unbalanced segmentations. DLMIA 2017
:param prediction: the logits
:param ground_truth: the segmentation ground truth
:param weight_map:
:param type_weight: type of weighting allowed between labels (choice
between Square (square of inverse of volume),
Simple (inverse of volume) and Uniform (no weighting))
:return: the loss
"""
ground_truth = tf.to_int64(ground_truth)
n_voxels = ground_truth.shape[0].value
n_classes = prediction.shape[1].value
ids = tf.constant(np.arange(n_voxels), dtype=tf.int64)
ids = tf.stack([ids, ground_truth], axis=1)
one_hot = tf.SparseTensor(indices=ids,
values=tf.ones([n_voxels], dtype=tf.float32),
dense_shape=[n_voxels, n_classes])
if weight_map is not None:
weight_map_nclasses = tf.reshape(
tf.tile(weight_map, [n_classes]), prediction.get_shape())
ref_vol = tf.sparse_reduce_sum(
weight_map_nclasses * one_hot, reduction_axes=[0])
intersect = tf.sparse_reduce_sum(
weight_map_nclasses * one_hot * prediction, reduction_axes=[0])
seg_vol = tf.reduce_sum(
tf.multiply(weight_map_nclasses, prediction), 0)
else:
ref_vol = tf.sparse_reduce_sum(one_hot, reduction_axes=[0])
intersect = tf.sparse_reduce_sum(one_hot * prediction,
reduction_axes=[0])
seg_vol = tf.reduce_sum(prediction, 0)
if type_weight == 'Square':
weights = tf.reciprocal(tf.square(ref_vol))
elif type_weight == 'Simple':
weights = tf.reciprocal(ref_vol)
elif type_weight == 'Uniform':
weights = tf.ones_like(ref_vol)
else:
raise ValueError("The variable type_weight \"{}\""
"is not defined.".format(type_weight))
new_weights = tf.where(tf.is_inf(weights), tf.zeros_like(weights), weights)
weights = tf.where(tf.is_inf(weights), tf.ones_like(weights) *
tf.reduce_max(new_weights), weights)
generalised_dice_numerator = \
2 * tf.reduce_sum(tf.multiply(weights, intersect))
generalised_dice_denominator = \
tf.reduce_sum(tf.multiply(weights, seg_vol + ref_vol))
generalised_dice_score = \
generalised_dice_numerator / generalised_dice_denominator
return 1 - generalised_dice_score
def _map_to_tfidf(x):
"""Calculates the inverse document frequency of terms in the corpus.
Args:
x : a SparseTensor of int64 representing string indices in vocab.
Returns:
The tf*idf values
"""
# Add one to the reduced term freqnencies to avoid dividing by zero.
idf = tf.log(tf.to_double(corpus_size) / (
1.0 + tf.to_double(reduced_term_freq)))
dense_doc_sizes = tf.to_double(tf.sparse_reduce_sum(tf.SparseTensor(
indices=x.indices,
values=tf.ones_like(x.values),
dense_shape=x.dense_shape), 1))
# For every term in x, divide the idf by the doc size.
# The two gathers both result in shape <sum_doc_sizes>
idf_over_doc_size = (tf.gather(idf, x.values) /
tf.gather(dense_doc_sizes, x.indices[:, 0]))
return tf.SparseTensor(
indices=x.indices,
values=idf_over_doc_size,
dense_shape=x.dense_shape)
def dice(prediction, ground_truth, weight_map=None):
"""
Function to calculate the dice loss with the definition given in
Milletari, F., Navab, N., & Ahmadi, S. A. (2016)
V-net: Fully convolutional neural
networks for volumetric medical image segmentation. 3DV 2016
using a square in the denominator
:param prediction: the logits
:param ground_truth: the segmentation ground_truth
:param weight_map:
:return: the loss
"""
ground_truth = tf.to_int64(ground_truth)
prediction = tf.cast(prediction, tf.float32)
ids = tf.range(tf.to_int64(tf.shape(ground_truth)[0]), dtype=tf.int64)
ids = tf.stack([ids, ground_truth], axis=1)
one_hot = tf.SparseTensor(
indices=ids,
values=tf.ones_like(ground_truth, dtype=tf.float32),
dense_shape=tf.to_int64(tf.shape(prediction)))
if weight_map is not None:
n_classes = prediction.shape[1].value
weight_map_nclasses = tf.reshape(
tf.tile(weight_map, [n_classes]), prediction.get_shape())
dice_numerator = 2.0 * tf.sparse_reduce_sum(
weight_map_nclasses * one_hot * prediction, reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(weight_map_nclasses * tf.square(prediction),
reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot * weight_map_nclasses,
reduction_axes=[0])
else:
dice_numerator = 2.0 * tf.sparse_reduce_sum(
one_hot * prediction, reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(tf.square(prediction), reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot, reduction_axes=[0])
epsilon_denominator = 0.00001
dice_score = dice_numerator / (dice_denominator + epsilon_denominator)
# dice_score.set_shape([n_classes])
# minimising (1 - dice_coefficients)
return 1.0 - tf.reduce_mean(dice_score)
def dice_nosquare(prediction, ground_truth, weight_map=None):
"""
Function to calculate the classical dice loss
:param prediction: the logits
:param ground_truth: the segmentation ground_truth
:param weight_map:
:return: the loss
"""
ground_truth = tf.to_int64(ground_truth)
n_voxels = ground_truth.shape[0].value
n_classes = prediction.shape[1].value
# construct sparse matrix for ground_truth to save space
ids = tf.constant(np.arange(n_voxels), dtype=tf.int64)
ids = tf.stack([ids, ground_truth], axis=1)
one_hot = tf.SparseTensor(indices=ids,
values=tf.ones([n_voxels], dtype=tf.float32),
dense_shape=[n_voxels, n_classes])
# dice
if weight_map is not None:
weight_map_nclasses = tf.reshape(
tf.tile(weight_map, [n_classes]), prediction.get_shape())
dice_numerator = 2.0 * tf.sparse_reduce_sum(
weight_map_nclasses * one_hot * prediction, reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(prediction * weight_map_nclasses,
reduction_indices=[0]) + \
tf.sparse_reduce_sum(weight_map_nclasses * one_hot,
reduction_axes=[0])
else:
dice_numerator = 2.0 * tf.sparse_reduce_sum(one_hot * prediction,
reduction_axes=[0])
dice_denominator = tf.reduce_sum(prediction, reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot, reduction_axes=[0])
epsilon_denominator = 0.00001
dice_score = dice_numerator / (dice_denominator + epsilon_denominator)
# dice_score.set_shape([n_classes])
# minimising (1 - dice_coefficients)
return 1.0 - tf.reduce_mean(dice_score)
def _count_docs_with_term(term_frequency):
"""Computes the number of documents in a batch that contain each term.
Args:
term_frequency: The `SparseTensor` output of _to_term_frequency.
Returns:
A `Tensor` of shape (vocab_size,) that contains the number of documents in
the batch that contain each term.
"""
count_of_doc_inter = tf.SparseTensor(
indices=term_frequency.indices,
values=tf.ones_like(term_frequency.values),
dense_shape=term_frequency.dense_shape)
out = tf.sparse_reduce_sum(count_of_doc_inter, axis=0)
return tf.expand_dims(out, 0)
def _to_term_frequency(x, vocab_size):
"""Creates a SparseTensor of term frequency for every doc/term pair.
Args:
x : a SparseTensor of int64 representing string indices in vocab.
vocab_size: An int - the count of vocab used to turn the string into int64s
including any OOV buckets.
Returns:
a SparseTensor with the count of times a term appears in a document at
indices <doc_index_in_batch>, <term_index_in_vocab>,
with size (num_docs_in_batch, vocab_size).
"""
# Construct intermediary sparse tensor with indices
# [<doc>, <term_index_in_doc>, <vocab_id>] and tf.ones values.
split_indices = tf.to_int64(
tf.split(x.indices, axis=1, num_or_size_splits=2))
expanded_values = tf.to_int64(tf.expand_dims(x.values, 1))
next_index = tf.concat(
[split_indices[0], split_indices[1], expanded_values], axis=1)
next_values = tf.ones_like(x.values)
vocab_size_as_tensor = tf.constant([vocab_size], dtype=tf.int64)
next_shape = tf.concat([x.dense_shape, vocab_size_as_tensor], 0)
next_tensor = tf.SparseTensor(indices=tf.to_int64(next_index),
values=next_values,
dense_shape=next_shape)
# Take the intermediary tensor and reduce over the term_index_in_doc
# dimension. This produces a tensor with indices [<doc_id>, <term_id>]
# and values [count_of_term_in_doc] and shape batch x vocab_size
term_count_per_doc = tf.sparse_reduce_sum_sparse(next_tensor, 1)
dense_doc_sizes = tf.to_double(
tf.sparse_reduce_sum(
tf.SparseTensor(indices=x.indices,
values=tf.ones_like(x.values),
dense_shape=x.dense_shape), 1))
gather_indices = term_count_per_doc.indices[:, 0]
gathered_doc_sizes = tf.gather(dense_doc_sizes, gather_indices)
term_frequency = (tf.to_double(term_count_per_doc.values) /
tf.to_double(gathered_doc_sizes))
return tf.SparseTensor(indices=term_count_per_doc.indices,
values=term_frequency,
dense_shape=term_count_per_doc.dense_shape)
def conv(features, adj, weights):
degree = tf.sparse_reduce_sum(adj, axis=1) + 1
degree = tf.cast(degree, tf.float32)
degree = tf.pow(degree, -0.5)
adj = sparse_tensor_diag_matmul(adj, degree, transpose=True)
adj = sparse_tensor_diag_matmul(adj, degree, transpose=False)
output = tf.sparse_tensor_dense_matmul(adj, features)
features = tf.transpose(features)
features = tf.multiply(tf.multiply(degree, features), degree)
features = tf.transpose(features)
output = output + features
return tf.matmul(output, weights)
def swtich_loss(prediction, ground_truth, for_brats, weight_map=None):
if not for_brats:
prediction = tf.cast(prediction, tf.float32)
if len(ground_truth.shape) == len(prediction.shape):
ground_truth = ground_truth[..., -1]
one_hot = labels_to_one_hot(ground_truth, tf.shape(prediction)[-1])
dice_numerator = 2.0 * tf.sparse_reduce_sum(one_hot * prediction,
reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(tf.square(prediction), reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot, reduction_axes=[0])
epsilon_denominator = 0.00001
dice_score = dice_numerator / (dice_denominator +
epsilon_denominator)
dice_score.set_shape([7])
# dice_score.set_shape([n_classes])
# minimising (1 - dice_coefficients)
return 1.0 - tf.reduce_mean(dice_score[:4])
else:
prediction = tf.cast(prediction, tf.float32)
if len(ground_truth.shape) == len(prediction.shape):
ground_truth = ground_truth[..., -1]
one_hot = labels_to_one_hot(ground_truth, tf.shape(prediction)[-1])
#prediction = tf.concat([tf.reduce_sum(prediction[:,:4],axis=1,keep_dims=True),prediction[:,1:]],axis=-1)
#one_hot = tf.concat([tf.expand_dims(one_hot[:,0],axis=-1),one_hot[:,-3:]],axis=-1)
if weight_map is not None:
n_classes = prediction.shape[1].value
weight_map_nclasses = tf.reshape(
tf.tile(weight_map, [n_classes]), prediction.get_shape())
dice_numerator = 2.0 * tf.sparse_reduce_sum(
weight_map_nclasses * one_hot * prediction,
reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(weight_map_nclasses * tf.square(prediction),
reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot * weight_map_nclasses,
reduction_axes=[0])
else:
dice_numerator = 2.0 * tf.sparse_reduce_sum(
one_hot * prediction, reduction_axes=[0])
dice_denominator = \
tf.reduce_sum(tf.square(prediction), reduction_indices=[0]) + \
tf.sparse_reduce_sum(one_hot, reduction_axes=[0])
epsilon_denominator = 0.00001
dice_score = dice_numerator / (dice_denominator +
epsilon_denominator)
# minimising (1 - dice_coefficients)
#dice_score[:4] = 0
#dice_score = tf.concat([dice_score[-3:],dice_score[:1]],axis=-1)
print(dice_score)
return 1.0 - tf.reduce_mean(dice_score)
def parse(self, tensors, dtype):
atoms = tensors['atoms']
sparse = 0.0
for key, val in self.dress.items():
sparse += tf.cast(tf.equal(atoms.sparse, key), dtype) * val
sparse = tf.SparseTensor(atoms.indices, sparse, atoms.mask.shape)
energy = tf.sparse_reduce_sum(sparse, [-1])
if 'e_data' in tensors:
# We are in training
tensors['e_data'] -= energy
tensors['e_data'] *= 627.509
tensors['energy'] = tf.constant(0.0)
def encoder_attr(self, embs_cate, w_list, b_list):
embs_cate = tf.nn.l2_normalize(embs_cate, axis=1)
embs_cnxt = tf.nn.l2_normalize(self.qInfer_network_sparse(
w_list, b_list),
axis=1)
p_assign = tf.matmul(embs_cnxt, embs_cate, transpose_b=True) / self.tau
if not self.gumbel:
cates = tf.nn.softmax(p_assign, axis=1)
else:
cates_dist = RelaxedOneHotCategorical(1, p_assign)
cates_sample = cates_dist.sample()
cates_mode = tf.nn.softmax(p_assign, axis=1)
cates = (self.is_training_ph * cates_sample +
(1 - self.is_training_ph) * cates_mode)
# VAE based encoding
z_list = []
zItem_list, cateItem_list = [], []
kl = None
x_input2attr = tf.sparse_tensor_dense_matmul(self.kg_mat,
self.input_ph,
adjoint_a=True,
adjoint_b=True)
x_input2attr = tf.transpose(x_input2attr)
for k in range(self.K):
cates_k = tf.reshape(cates[:, k], (1, -1))
# q-network for user aspects
x_k = x_input2attr * cates_k
mu_k, std_k, kl_k = self.q_network(x_k, w_list, b_list)
eps = tf.random_normal(tf.shape(std_k), dtype=tf.float64)
z_k = mu_k + self.is_training_ph * eps * std_k
z_list.append(z_k)
kl = (kl_k if (kl is None) else (kl + kl_k))
# q-network for item aspects
x_k = self.kg_mat.__mul__(cates_k)
mu_k, std_k, kl_k = self.q_network_sparse(x_k, w_list, b_list)
eps = tf.random_normal(tf.shape(std_k), dtype=tf.float64)
z_k = mu_k + self.is_training_ph * eps * std_k
zItem_list.append(z_k)
cates_sum_k = tf.sparse_reduce_sum(x_k, axis=1)
cates_sum_k = tf.reshape(cates_sum_k, (1, -1))
cateItem_list.append(cates_sum_k / tf.reduce_sum(cates_sum_k))
return z_list, zItem_list, cateItem_list, kl
def make_bag_of_words(self, active_words, doc_tokens, batch_size,
num_samples, doc_len, vocab_size):
"""Turn sequences of words with extraction labels into a bag of words.
Args:
active_words: batch of binary masks indicating picked words.
doc_tokens: batch of token sequences for each document.
batch_size: number of sequences in batch.
num_samples: number of samples of masks for each sequence.
doc_len: length of sequences.
vocab_size: size of vocabulary.
Returns:
dense_pred_counts: batch of bags of words.
"""
# can't make a sparse count tensor directly since doesn't support repeated
# words -- make a new sparse tensor with extra indices to handle repeats
with tf.variable_scope("bag_of_words"):
batch_idx = shared_util.repeat_row(num_samples * doc_len,
tf.range(0, batch_size, 1))
sample_idx = tf.tile(
shared_util.repeat_row(doc_len, tf.range(0, num_samples, 1)),
[batch_size])
tok_idx = tf.tile(tf.range(0, doc_len, 1),
[batch_size * num_samples])
vocab_idx = tf.tile(
tf.reshape(doc_tokens, [batch_size, 1, doc_len]),
[1, num_samples, 1])
active_tok_indices = tf.concat(1, [
tf.reshape(batch_idx, [-1, 1]),
tf.reshape(sample_idx, [-1, 1]),
tf.reshape(tok_idx, [-1, 1]),
tf.reshape(vocab_idx, [-1, 1])
])
active_toks_sparse = tf.SparseTensor(
indices=tf.to_int64(active_tok_indices),
values=tf.reshape(active_words, [-1]),
shape=tf.to_int64(
tf.pack([batch_size, num_samples, doc_len, vocab_size])))
dense_pred_counts = tf.sparse_reduce_sum(active_toks_sparse,
reduction_axes=2)
return dense_pred_counts
def vector_to_adjacency_sym_sparse(inputs):
G, e, dense_shape = inputs
e = add_epsilon(e)
A = tf.SparseTensor(indices=G.indices, values=e, dense_shape=dense_shape)
D = tf.pow(tf.sparse_reduce_sum(A, 1), -0.5)
#row wise normalization
Drow = tf.gather(D, G.indices[:, 0])
#column wise normalization, currently disabled
#Dcol = tf.gather(D, G.indices[:, 1])
#multiply values by D
# e_ = tf.multiply(tf.multiply(Dcol, e), Drow)
e_ = tf.multiply(e, Drow)
A_ = tf.SparseTensor(indices=G.indices, values=e_, dense_shape=dense_shape)
return A_
def body(x, previous):
# batch [max_rule_len, batch_size, num_relations]
# ==> a_vector : [batch_size, num_relations, 1, 1]
a_vector = tf.nn.embedding_lookup(self.a_matrix, x)
weighted_operator_tensor = tensor_mul_batch_vector(a_vector, self.operator_tensor)
# a_vector_exp = tf.expand_dims(tf.expand_dims(a_vector, 2), 2)
# [batch_size, num_relations, num_entities, num_entities]
# weighted_operator_tensor = tf.multiply(a_vector_exp,
# tf.expand_dims(self.operator_tensor, 0))
# score with previous input
u = tf.sparse_reduce_sum(weighted_operator_tensor.__mul__(
tf.expand_dims(tf.expand_dims(previous, 1), 1)), 3)
# sum up relation weighted vectors
u_sum = tf.reduce_sum(u, axis=1)
# [batch_size,num_relations,num_entities,num_entities] * [batch_size,1,num_entities]
return [x+1, u_sum]
def ElectrostaticDampedShiftedLinear(Ds, Qs, NZP, alpha, Rc):
"""
A tensorflow linear scaling implementation of the Damped Shifted Electrostatic Force
http://aip.scitation.org.proxy.library.nd.edu/doi/pdf/10.1063/1.2206581
Batched over molecules.
Args:
Ds: Distances Enumerated by NZP (flat)
Qs: A batch of Atomic Charges. (nmol X maxatom)
NZP: a list of nonzero atom pairs NNZ X (mol, i, j).
alpha: DSF alpha parameter (~0.2)
Rc: DSF Rc parameter. (15A)
Returns
A #Mols X MaxNAtoms X MaxNAtoms matrix of LJ kernel contributions.
"""
twooversqrtpi = tf.constant(1.1283791671, dtype=tf.float64)
NZP_shape = tf.shape(NZP)
Zs_shp = tf.shape(Zs)
maxnpairs = NZP_shape[0]
nmols = Zs_shp[0]
Ii = tf.slice(NZP, [0, 0], [-1, 2])
Ij = tf.concat(
[tf.slice(NZP, [0, 0], [-1, 1]),
tf.slice(NZP, [0, 2], [-1, 1])], 1)
Qi = tf.reshape(tf.gather_nd(Qs, Ii), [maxnpairs])
Qj = tf.reshape(tf.gather_nd(Qs, Ij), [maxnpairs])
# Gather desired LJ parameters.
Qij = Qi * Qj
# This is Dan's Equation (18)
XX = alpha * Rc
ZZ = tf.erfc(XX) / Rc
YY = twooversqrtpi * alpha * tf.exp(-XX * XX) / Rc
K = Qij * (tf.erfc(alpha * Ds) / Ds - ZZ + (Ds - Rc) * (ZZ / Rc + YY))
K = tf.where(tf.is_nan(K), tf.zeros_like(K), K)
range_index = tf.reshape(
tf.range(tf.cast(maxnpairs, tf.int64), dtype=tf.int64), [maxnpairs, 1])
mol_index = tf.reshape(tf.slice(NZP, [0, 0], [-1, 1]), [maxnpairs, 1])
inds = tf.reshape(tf.stack([mol_index, range_index], axis=1),
[maxnpairs, 2])
# Now use the sparse reduce sum trick to scatter this into mols.
sp_atomoutputs = tf.SparseTensor(
inds,
tf.reshape(K, [maxnpairs]),
dense_shape=[tf.cast(nmols, tf.int64),
tf.cast(maxnpairs, tf.int64)])
return tf.sparse_reduce_sum(sp_atomoutputs, axis=1)
def _call(self, inputs):
x = inputs
# dropout
if self.sparse_inputs:
x = sparse_dropout(x, 1 - self.dropout, self.num_features_nonzero)
else:
x = tf.nn.dropout(x, 1 - self.dropout)
# sum x
if self.sparse_inputs:
sx = tf.sparse_reduce_sum(x, axis=1, keep_dims=True)
else:
sx = tf.reduce_sum(x, axis=1, keep_dims=True)
# convolve
supports = list()
for i in range(len(self.support)):
if not self.featureless:
pre_sup = dot(x,
self.vars['weights_' + str(i)],
sparse=self.sparse_inputs)
pre_sup_b = dot(x,
self.vars['weights_b_' + str(i)],
sparse=self.sparse_inputs)
else:
pre_sup = self.vars['weights_' + str(i)]
pre_sup_b = self.vars['weights_b_' + str(i)]
# support = dot(self.support[i], pre_sup, sparse=True)
# supports.append(support)
# calculate second order interaction: XW * sum(X)
# pre_sup_sec_ord = tf.multiply(pre_sup, pre_sup_b) * self.alpha
pre_sup_sec_ord = tf.multiply(pre_sup,
pre_sup_b) * self.vars['alp']
# pre_sup_all = tf.concat([pre_sup, pre_sup_sec_ord], axis=1)
pre_sup_all = tf.add_n([pre_sup, pre_sup_b, pre_sup_sec_ord])
support = dot(self.support[i], pre_sup_all, sparse=True)
supports.append(support)
output = tf.add_n(supports)
# bias
if self.bias:
output += self.vars['bias']
return self.act(output)
def soft_ncut(image, image_segment, image_weights):
"""
Args:
image: [B, H, W, C]
image_segment: [B, H, W, K]
image_weights: [B, H*W, H*W]
Returns:
Soft_Ncut: scalar
"""
batch_size = tf.shape(image)[0]
num_class = tf.shape(image_segment)[-1]
image_shape = image.get_shape()
weight_size = image_shape[1].value * image_shape[2].value
image_segment = tf.transpose(image_segment, [0, 3, 1, 2]) # [B, K, H, W]
image_segment = tf.reshape(image_segment, tf.stack([batch_size, num_class, weight_size])) # [B, K, H*W]
# Dis-association
# [B0, H*W, H*W] @ [B1, K1, H*W] contract on [[2],[2]] = [B0, H*W, B1, K1]
W_Ak = sparse_tensor_dense_tensordot(image_weights, image_segment, axes=[[2],[2]])
W_Ak = tf.transpose(W_Ak, [0,2,3,1]) # [B0, B1, K1, H*W]
W_Ak = sycronize_axes(W_Ak, [0,1], tensor_dims=4) # [B0=B1, K1, H*W]
# [B1, K1, H*W] @ [B2, K2, H*W] contract on [[2],[2]] = [B1, K1, B2, K2]
dis_assoc = tf.tensordot(W_Ak, image_segment, axes=[[2],[2]])
dis_assoc = sycronize_axes(dis_assoc, [0,2], tensor_dims=4) # [B1=B2, K1, K2]
dis_assoc = sycronize_axes(dis_assoc, [1,2], tensor_dims=3) # [K1=K2, B1=B2]
dis_assoc = tf.transpose(dis_assoc, [1,0]) # [B1=B2, K1=K2]
dis_assoc = tf.identity(dis_assoc, name="dis_assoc")
# Association
# image_segment: [B, K, H*W]
sum_W = tf.sparse_reduce_sum(image_weights,axis=2) # [B, W*H]
assoc = tf.tensordot(image_segment, sum_W, axes=[2,1]) # [B, K, B]
assoc = sycronize_axes(assoc, [0,2], tensor_dims=3) # [B0=B1, K0]
assoc = tf.identity(assoc, name="assoc")
utils.add_activation_summary(dis_assoc)
utils.add_activation_summary(assoc)
# Soft NCut
eps = 1e-6
soft_ncut = tf.cast(num_class, tf.float32) - \
tf.reduce_sum((dis_assoc + eps) / (assoc + eps), axis=1)
return soft_ncut
def _sparse_ww(config, Rs, predictions_per_sample, zijs, bias):
# Zj has shape (batch_size, 1, output_width) dense tensor
zj = tf.sparse_reduce_sum(zijs, 1, keep_dims=True)
# Stabilizer
zj_sign = tf.sign(zj)
zj_sign = tf.where(tf.equal(zj, 0), tf.ones_like(zj_sign), zj_sign)
zj += zj_sign * EPSILON
# construct bias to add to zj
fractions = zijs / zj
# Distribute the relevance according to the fractions
R_new = _sparse_distribute_relevances(Rs, zijs.dense_shape[0],
zijs.dense_shape[1],
predictions_per_sample, fractions)
return R_new
def forward(self, save_emb=False):
denom = tf.reduce_sum(self.input_ph, axis=1, keep_dims=True)
self.input_ph = self.input_ph.__div__(denom)
denom = tf.sparse_reduce_sum(self.kg_mat, axis=1, keep_dims=True)
self.kg_mat = self.kg_mat.__div__(denom)
zAct_list, cAct_list, klAct = self.encoder_item(
self.cate_item, self.w_qItem, self.b_qItem)
zAsp_list, zItem_list, cateItem_list, klAsp = self.encoder_attr(
self.cate_attr, self.w_qAttr, self.b_qAttr)
logits, loss_recon = self.decoder(zAct_list, cAct_list,
self.embs_item_self,
self.embs_item_enti, zAsp_list,
zItem_list, cateItem_list)
return tf.train.Saver(), logits, loss_recon, klAct
def _tfidf(x):
split = tf.string_split(x)
table = lookup.string_to_index_table_from_tensor(
vocab, num_oov_buckets=0,
default_value=len(vocab))
int_text = table.lookup(split)
term_count_per_doc = get_term_count_per_doc(int_text, len(vocab) + 1)
# Add one to the reduced term freqnencies to avoid dividing by zero.
example_count_with_oov = tf.to_float(tf.concat([example_count, [0]], 0))
idf = tf.log(tf.to_float(corpus_size) / (1.0 + example_count_with_oov))
dense_doc_sizes = tf.to_float(tf.sparse_reduce_sum(tf.SparseTensor(
indices=int_text.indices,
values=tf.ones_like(int_text.values),
dense_shape=int_text.dense_shape), 1))
idf_times_term_count = tf.multiply(
tf.gather(idf, term_count_per_doc.indices[:, 1]),
tf.to_float(term_count_per_doc.values))
tfidf_weights = (
idf_times_term_count / tf.gather(dense_doc_sizes,
term_count_per_doc.indices[:, 0]))
tfidf_ids = term_count_per_doc.indices[:, 1]
indices = tf.stack([term_count_per_doc.indices[:, 0],
segment_indices(term_count_per_doc.indices[:, 0],
int_text.dense_shape[0])],
1)
dense_shape = term_count_per_doc.dense_shape
tfidf_st_weights = tf.SparseTensor(indices=indices,
values=tfidf_weights,
dense_shape=dense_shape)
tfidf_st_ids = tf.SparseTensor(indices=indices,
values=tfidf_ids,
dense_shape=dense_shape)
if part == 'ids':
return tfidf_st_ids
else:
return tfidf_st_weights
def calculate_final_dist(self, vocab_dist, attn_dist, pointer_gen, passage_words, passage_mask=None):
'''
vocab_dist: [batch_size, vsize]
attn_dist: [batch_size, passage_length]
pointer_gen: [batch_size, 1]
passage_words: [batch_size, passage_length]
passage_mask: [batch_size, passage_length]
'''
input_shape = tf.shape(vocab_dist)
batch_size = input_shape[0]
vsize = input_shape[1]
passage_length = tf.shape(passage_words)[1]
with tf.variable_scope('final_distribution'):
vocab_dist = pointer_gen * vocab_dist
attn_dist = (1.0 - pointer_gen) * attn_dist
# match attn_dist[batch_size, passage_length] to sparse one-hot representation [batch_size, passage_length, vsize]
batch_nums = tf.range(0, limit=batch_size) # shape (batch_size)
batch_nums = tf.expand_dims(batch_nums, axis=1) # shape (batch_size, 1)
batch_nums = tf.tile(batch_nums, [1, passage_length]) # shape (batch_size, passage_length)
step_nums = tf.range(0, limit=passage_length) # [passage_length]
step_nums = tf.expand_dims(step_nums, axis=0) # shape (1, passage_length)
step_nums = tf.tile(step_nums, [batch_size, 1]) # shape (batch_size, passage_length)
indices = tf.stack((batch_nums, step_nums, passage_words), axis=2) # shape (batch_size, passage_length, 3)
indices = tf.reshape(indices, [-1, 3]) #[batch_size * passage_length, 3]
indices = tf.cast(indices, tf.int64)
shape = [batch_size, passage_length, vsize]
shape = tf.cast(shape, tf.int64)
attn_dist = tf.reshape(attn_dist, shape=[-1]) # [batch_size*passage_length]
one_hot_spare_representation = tf.SparseTensor(indices=indices, values=attn_dist, dense_shape=shape) # [batch_size, passage_length, vsize]
if passage_mask is not None:
passage_mask = tf.expand_dims(passage_mask, axis=-1)
one_hot_spare_representation = one_hot_spare_representation * passage_mask
one_hot_spare_representation = tf.sparse_reduce_sum(one_hot_spare_representation, axis=1) # [batch_size, vsize]
vocab_dist = tf.add(vocab_dist, one_hot_spare_representation)
return vocab_dist # [batch_size, vsize]
def merge_prob_dist_for_one_step(self, vocab_dist, attn_dist, p_gen, node_idxs, node_mask=None):
'''
vocab_dist: [batch_size, vsize]
attn_dist: [batch_size, passage_length]
p_gen: [batch_size, 1]
node_idxs: [batch_size, passage_length]
node_mask: [batch_size, passage_length]
'''
input_shape = tf.shape(vocab_dist)
batch_size = input_shape[0]
vsize = input_shape[1]
passage_length = tf.shape(node_idxs)[1]
with tf.variable_scope('final_distribution'):
vocab_dist = p_gen * vocab_dist
attn_dist = (1.0-p_gen) * attn_dist
# match attn_dist[batch_size, passage_length] to sparse one-hot representation [batch_size, passage_length, vsize]
batch_nums = tf.range(0, limit=batch_size) # shape (batch_size)
batch_nums = tf.expand_dims(batch_nums, axis=1) # shape (batch_size, 1)
batch_nums = tf.tile(batch_nums, [1, passage_length]) # shape (batch_size, passage_length)
step_nums = tf.range(0, limit=passage_length) # [passage_length]
step_nums = tf.expand_dims(step_nums, axis=0) # shape (1, passage_length)
step_nums = tf.tile(step_nums, [batch_size, 1]) # shape (batch_size, passage_length)
indices = tf.stack((batch_nums, step_nums, node_idxs), axis=2) # shape (batch_size, passage_length, 3)
indices = tf.reshape(indices, [-1, 3]) #[batch_size * passage_length, 3]
indices = tf.cast(indices, tf.int64)
shape = [batch_size, passage_length, vsize]
shape = tf.cast(shape, tf.int64)
attn_dist = tf.reshape(attn_dist, shape=[-1]) # [batch_size*passage_length]
one_hot_spare_rep = tf.SparseTensor(indices=indices, values=attn_dist, dense_shape=shape) # [batch_size, passage_length, vsize]
if node_mask is not None:
node_mask = tf.expand_dims(node_mask, axis=-1)
one_hot_spare_rep = one_hot_spare_rep * node_mask
# 前面所做的一切都是将attention分布转换到dict范围内，方便与P_vocab相加
one_hot_spare_rep = tf.sparse_reduce_sum(one_hot_spare_rep, axis=1) # [batch_size, vsize]
vocab_dist = tf.add(vocab_dist, one_hot_spare_rep)
return vocab_dist # [batch_size, vsize]
def text_to_label(self,
batch_text,
return_dense=True,
pad_value=-1,
return_lengths=False):
"""
给定字符型文本转化为整型标签序列，只适用于英文句子
Args:
text: ascii encoded string tensor with shape [batch_size]
return_dense: whether to return dense labels
pad_value: Value used to pad labels to the same length.
return_lengths: if True, also return text lengths
Returns:
labels: sparse or dense tensor of labels
"""
"""
# 英文句子按空格分词，例如source = ["hello world", "a b c"], delimiter='',
# 返回tf.SparseTensor对象， st.indices = [0, 0; 0, 1; 1, 0; 1, 1; 1, 2]，
# st.shape = [2, 3] st.values = ['hello', 'world', 'a', 'b', 'c']
# The first column of the indices corresponds to the row in source and the second column
# corresponds to the index of the split component in this row.
"""
chars = tf.string_split(batch_text, delimiter='')
labels_sp = tf.SparseTensor(
chars.indices, self._char_to_label_table.lookup(chars.values),
chars.dense_shape)
if return_dense:
labels = tf.sparse_tensor_to_dense(labels_sp,
default_value=pad_value)
else:
labels = labels_sp
if return_lengths:
text_lengths = tf.sparse_reduce_sum(tf.SparseTensor(
chars.indices, tf.fill([tf.shape(chars.indices)[0]], 1),
chars.dense_shape),
axis=1)
text_lengths.set_shape([None])
return labels, text_lengths
else:
return labels
def count_nonzero_wrapper(X, optype):
"""Wrapper for handling sparse and dense versions of `tf.count_nonzero`.
Parameters
----------
X : tf.Tensor (N, K)
optype : str, {'dense', 'sparse'}
Returns
-------
tf.Tensor (1,K)
"""
with tf.name_scope('count_nonzero_wrapper') as scope:
if optype == 'dense':
return tf.count_nonzero(X, axis=0, keep_dims=True)
elif optype == 'sparse':
indicator_X = tf.SparseTensor(X.indices, tf.ones_like(X.values), X.dense_shape)
return tf.sparse_reduce_sum(indicator_X, axis=0, keep_dims=True)
else:
raise NameError('Unknown input type in count_nonzero_wrapper')
def count_nonzero_wrapper(X, optype):
"""Wrapper for handling sparse and dense versions of `tf.count_nonzero`.
Parameters
----------
X : tf.Tensor (N, K)
optype : str, {'dense', 'sparse'}
Returns
-------
tf.Tensor (1,K)
"""
with tf.name_scope('count_nonzero_wrapper') as scope:
if optype == 'dense':
return tf.count_nonzero(X, axis=0, keep_dims=True)
elif optype == 'sparse':
indicator_X = tf.SparseTensor(X.indices, tf.ones_like(X.values), X.dense_shape)
return tf.sparse_reduce_sum(indicator_X, axis=0, keep_dims=True)
else:
raise NameError('Unknown input type in count_nonzero_wrapper')
def LJKernelsLinear(Ds, Zs, Ee, Re, NZP):
"""
Batched over molecules.
Args:
Ds: Distances Enumerated by NZP (flat)
Zs: A batch of Atomic Numbers. (nmol X maxatom X 1)
Ee: a matrix of LJ well depths.
Re: a matrix of Bond minima.
NZP: a list of nonzero atom pairs NNZ X (mol, i, j).
Returns
A #Mols vector of LJ energies.
"""
NZP_shape = tf.shape(NZP)
Zs_shp = tf.shape(Zs)
maxnpairs = NZP_shape[0]
nmols = Zs_shp[0]
Ii = tf.slice(NZP, [0, 0], [-1, 2])
Ij = tf.concat(
[tf.slice(NZP, [0, 0], [-1, 1]),
tf.slice(NZP, [0, 2], [-1, 1])], 1)
Zi = tf.reshape(tf.gather_nd(Zs, Ii), [maxnpairs])
Zj = tf.reshape(tf.gather_nd(Zs, Ij), [maxnpairs])
# Gather desired LJ parameters.
Zij = tf.stack([Zi, Zj], axis=1)
Eeij = tf.reshape(tf.gather_nd(Ee, Zij), [maxnpairs])
Reij = tf.reshape(tf.gather_nd(Re, Zij), [maxnpairs])
R = Reij / tf.reshape(Ds, [maxnpairs])
K = Eeij * (tf.pow(R, 12.0) - 2.0 * tf.pow(R, 6.0))
K = tf.where(tf.is_nan(K), tf.zeros_like(K), K)
range_index = tf.reshape(
tf.range(tf.cast(maxnpairs, tf.int64), dtype=tf.int64), [maxnpairs, 1])
mol_index = tf.reshape(tf.slice(NZP, [0, 0], [-1, 1]), [maxnpairs, 1])
inds = tf.reshape(tf.stack([mol_index, range_index], axis=1),
[maxnpairs, 2])
# Now use the sparse reduce sum trick to scatter this into mols.
sp_atomoutputs = tf.SparseTensor(
inds,
tf.reshape(K, [maxnpairs]),
dense_shape=[tf.cast(nmols, tf.int64),
tf.cast(maxnpairs, tf.int64)])
return tf.sparse_reduce_sum(sp_atomoutputs, axis=1)
def text_to_labels(self,
text,
return_dense=True,
pad_value=-1,
return_lengths=False):
"""Convert text strings to label sequences.
Args:
text: ascii encoded string tensor with shape [batch_size]
dense: whether to return dense labels
pad_value: Value used to pad labels to the same length.
return_lengths: if True, also return text lengths
Returns:
labels: sparse or dense tensor of labels
"""
batch_size = tf.shape(text)[0]
chars = tf.string_split(text, delimiter='')
labels_sp = tf.SparseTensor(
chars.indices,
self._char_to_label_table.lookup(chars.values),
chars.dense_shape
)
if return_dense:
labels = tf.sparse_tensor_to_dense(labels_sp, default_value=pad_value)
else:
labels = labels_sp
if return_lengths:
text_lengths = tf.sparse_reduce_sum(
tf.SparseTensor(
chars.indices,
tf.fill([tf.shape(chars.indices)[0]], 1),
chars.dense_shape
),
axis=1
)
text_lengths.set_shape([None])
return labels, text_lengths
else:
return labels
def test_t4():
print("----------------------")
sp1 = tf.SparseTensor(indices=[[0, 0], [0, 1], [1, 0]],
values=[12, -1, 0],
dense_shape=(3, 3)) #
#sp1.values = sp1.values+1
sp1_dense = tf.sparse_tensor_to_dense(sp1, default_value=-1)
sess = tf.Session()
print("sp1:", sess.run(sp1))
sp1_sign = tf.sign(sp1)
sp2 = tf.SparseTensor(indices=sp1.indices,
values=sp1.values + 2,
dense_shape=sp1.dense_shape) #
#sp1_sign = tf.sign(tf.sparse_add(sp1)))
#print("sp1_sign:",sess.run(sp1_sign))
print("sp2:", sess.run(sp2))
sp2_sign = tf.sign(sp2)
print("sp2_sign:", sess.run(sp2_sign))
print("sp2_sign_sum:", sess.run(tf.sparse_reduce_sum(sp2_sign, axis=1)))
print("sp2_sign_sum_fun:",
sess.run(count_sparse_nonzero(sp1, axis=1, add_default=2)))
def _post_setup(self):
self.E_predict = _tf.reduce_sum([
_tf.sparse_tensor_dense_matmul(self.atom_maps[t],
self.ANNs[t].output)
for t in self.atom_types
],
axis=[0, 2],
name="E_prediction")
# Tensorflow operation to initialize the variables of the atomic networks
#self.init_vars = [a.init_vars for a in self.ANNs.itervalues()]
self.num_atoms = _tf.reduce_sum([
_tf.sparse_reduce_sum(self.atom_maps[t], axis=1)
for t in self.atom_types
],
axis=0,
name="NumberOfAtoms")
# Tensorflow operation that calculates the sum squared error per atom.
# Note that the whole error per atom is squared.
with _tf.name_scope("RMSE"):
self.rmse_weights = _tf.placeholder(shape=(None, ),
dtype=precision,
name="weights")
self.rmse = self.error_scaling * _tf.sqrt(
_tf.reduce_mean(
(self.target - self.E_predict)**2 * self.rmse_weights))
#self.rmse = self.error_scaling*_tf.sqrt(
# _tf.losses.mean_squared_error(self.target,
# self.E_predict, weights = 1.0/self.num_atoms**2))
self.rmse_summ = _tf.summary.scalar("RMSE",
self.rmse,
family="performance")
self.variables = _tf.get_collection(
_tf.GraphKeys.MODEL_VARIABLES,
scope=_tf.get_default_graph().get_name_scope())
self.saver = _tf.train.Saver(self.variables,
max_to_keep=None,
save_relative_paths=True)
def weighted_margin_rank_batch(self, tf_prediction_serial, tf_interactions, tf_sample_predictions, tf_n_items,
tf_n_sampled_items):
positive_interaction_mask = tf.greater(tf_interactions.values, 0.0)
positive_interaction_indices = tf.boolean_mask(tf_interactions.indices,
positive_interaction_mask)
positive_interaction_values = tf.boolean_mask(tf_interactions.values,
positive_interaction_mask)
positive_interactions = tf.SparseTensor(indices=positive_interaction_indices,
values=positive_interaction_values,
dense_shape=tf_interactions.dense_shape)
listening_sum_per_item = tf.sparse_reduce_sum(positive_interactions, axis=0)
gathered_sums = tf.gather(params=listening_sum_per_item,
indices=tf.transpose(positive_interaction_indices)[1])
# [ n_positive_interactions ]
positive_predictions = tf.boolean_mask(tf_prediction_serial,
positive_interaction_mask)
n_items = tf.cast(tf_n_items, dtype=tf.float32)
n_sampled_items = tf.cast(tf_n_sampled_items, dtype=tf.float32)
# [ n_positive_interactions, n_sampled_items ]
mapped_predictions_sample_per_interaction = tf.gather(params=tf_sample_predictions,
indices=tf.transpose(positive_interaction_indices)[0])
# [ n_positive_interactions, n_sampled_items ]
summation_term = tf.maximum(1.0
- tf.expand_dims(positive_predictions, axis=1)
+ mapped_predictions_sample_per_interaction,
0.0)
# [ n_positive_interactions ]
sampled_margin_rank = ((n_items / n_sampled_items)
* tf.reduce_sum(summation_term, axis=1)
* positive_interaction_values / gathered_sums)
loss = tf.log(sampled_margin_rank + 1.0)
return loss
def parse(self, tensors, dtype):
nodes = tensors['nodes'][self.order]
sparse = nodes.sparse
for i, n_out in enumerate(self.n_nodes):
sparse = tf.layers.dense(sparse,
n_out,
activation=self.act,
name='{}-{}'.format(self.name, i))
sparse = tf.layers.dense(sparse,
1,
use_bias=False,
activation=None,
name='{}-en'.format(self.name))
sparse = tf.squeeze(sparse, -1)
sparse = tf.SparseTensor(nodes.indices, sparse, nodes.mask.shape)
sparse = tf.sparse_reduce_sum(sparse,
[-i - 1 for i in range(self.order + 1)])
tensors['energy'] += sparse
def _encode_proto(values_dict, message_type):
"""A wrapper around encode_proto_op.encode_proto."""
field_names = []
sizes = []
values = []
for field_name, value in sorted(values_dict.items(), key=lambda x: x[0]):
if isinstance(value, tf.SparseTensor):
size = tf.sparse_reduce_sum(tf.SparseTensor(
value.indices, tf.ones_like(value.values, dtype=tf.int32),
value.dense_shape),
axis=1)
value = tf.sparse_tensor_to_dense(
value, _DEFAULT_VALUE_BY_DTYPE[value.dtype])
else:
value = tf.reshape(value, [tf.shape(value)[0], -1])
size = tf.fill((tf.shape(value)[0], ), tf.shape(value)[1])
field_names.append(field_name)
values.append(value)
sizes.append(size)
sizes = tf.stack(sizes, axis=1)
return encode_proto_op.encode_proto(sizes, values, field_names,
message_type)
def attention3(self, inputs, support, la, num1, num2):
# input: N * E
attention_size = 64
s_m = tf.sparse_tensor_dense_matmul(support, inputs) # M*E support:[M*N]
sm_num = tf.sparse_reduce_sum(support, axis=1) # M*1
sm_mean = tf.divide(1, sm_num)
s_m = tf.multiply(sm_mean, tf.transpose(s_m)) # M * e*M
s_m = tf.transpose(s_m) # M*E
# TODO: W * sm
inputs_t = tf.transpose(inputs) # M*N
w_sm = tf.tensordot(s_m, self.vars['w_omega_%s' % (la)], axes=1) # [M*E] * [E*A] = [M*A]
# TODO: modifiy
support_3d = tf.sparse_reshape(support, [num2, 1, num1]) # M*N
support_3d = tf.sparse_to_dense(support_3d.indices, [num2, 1, num1], support_3d.values,
validate_indices=False, default_value=0.0)
s_j_all = tf.transpose(tf.multiply(support_3d, inputs_t), [0,2,1]) # M*1*N * [E*N] = [M*E*N]
w_sj_all = tf.tensordot(s_j_all, self.vars['w_omega_%s' % (la)], axes=1) # [M*N*E] * [E*A]
# M*A - M*N*A = M*N*A M*N
w_sm = tf.reshape(w_sm, [num2, 1, attention_size]) # M*1*A
euclidean_set = tf.sqrt(tf.reduce_sum(tf.square(w_sm - w_sj_all), 2)) # [M*1*A] * [M*N*A]
eucl = tf.tanh(euclidean_set) # M*N
print("euclidean computing success!")
print(eucl.shape)
vu = tf.multiply(self.vars['u_omega_%s' % (la)], eucl)
# vu = tf.matmul(self.vars['u_omega_%s' % (la)], eucl)
#vu = tf.tensordot(eucl,self.vars['u_omega_%s' % (la)],axes=1)
print("vu computing success!")
alphas = tf.nn.softmax(vu, name='alphas_%s' % (la)) # (B,T) shape
print("alphas computing success!")
return alphas
def q_network_sparse(self, x, w_list, b_list):
# The q-network contains two layers, an embedding layer and a mapping layer
mu_q, std_q, kl = None, None, None
denom = tf.sparse_reduce_sum(tf.square(x), axis=1, keep_dims=True)
repr = x.__div__(denom)
for i, (w, b) in enumerate(zip(w_list, b_list)):
if i != len(w_list) - 1:
repr = tf.sparse_tensor_dense_matmul(repr,
w,
adjoint_a=False,
adjoint_b=False)
repr = tf.nn.tanh(repr)
else:
repr = tf.matmul(repr, w, a_is_sparse=(i == 0)) + b
mu_q = repr[:, :self.dim]
mu_q = tf.nn.l2_normalize(mu_q, axis=1)
lnvarq_sub_lnvar0 = -repr[:, self.dim:]
std_q = tf.exp(0.5 * lnvarq_sub_lnvar0) * self.std
kl = tf.reduce_mean(
tf.reduce_sum(
0.5 *
(-lnvarq_sub_lnvar0 + tf.exp(lnvarq_sub_lnvar0) - 1.),
axis=1))
return mu_q, std_q, kl
def _sparse_distribute_bias(config, zijs, bias):
# Return if bias or the bias strategy is none or throw error if it is all
# since all will kill the memory (remember we are in sparse land here ;) )
if bias is None or config.bias_strategy == BIAS_STRATEGY.IGNORE:
return zijs
if config.bias_strategy == BIAS_STRATEGY.NONE:
return zijs
elif config.bias_strategy == BIAS_STRATEGY.ALL:
raise NotImplementedError(
"BIAS_STRATEGY.ALL is not implemented for sparse matmul")
# Dense Shape (batch_size, input_width, output_width)
indicators = tf.SparseTensor(
zijs.indices,
tf.where(tf.equal(zijs.values, 0), tf.zeros_like(zijs.values),
tf.ones_like(zijs.values)), zijs.dense_shape)
# Count all the indicators in each column
# Shape: (batch_size, 1, output_width)
counts = tf.sparse_reduce_sum(indicators, axis=1, keep_dims=True)
# Hack to avoid dividing by zero (doesn't matter for the final result.
counts = tf.where(tf.equal(counts, 0), tf.ones_like(counts), counts)
# Divide the bias by broadcasting it to every sample in the batch
bias_divided = bias / counts
# Scale the indicators by the bias
# Dense shape: (batch_size, input_width, output_width)
bias_to_add = indicators * bias_divided
# Create new zijs with the divided bias
zijs_new = tf.SparseTensor(zijs.indices, zijs.values + bias_to_add.values,
zijs.dense_shape)
return zijs_new
def _sparse_epsilon(config, Rs, predictions_per_sample, zijs, bias):
# Zj has shape (batch_size, 1, output_width) dense tensor
zj = tf.sparse_reduce_sum(zijs, 1, keep_dims=True)
# Prepare sparse tensor with duplicated bias for addition with zj
if bias is not None and config.bias_strategy != BIAS_STRATEGY.IGNORE:
zj = zj + bias
zj_sign = tf.sign(zj)
zj_sign = tf.where(tf.equal(zj, 0), tf.ones_like(zj_sign), zj_sign)
zj += zj_sign * config.epsilon
# Distribute bias according to config
zijs = _sparse_distribute_bias(config, zijs, bias)
# construct bias to add to zj
fractions = zijs / zj
# Distribute the relevance according to the fractions
R_new = _sparse_distribute_relevances(Rs, zijs.dense_shape[0],
zijs.dense_shape[1],
predictions_per_sample, fractions)
return R_new
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