def create_feed_dict(batch, wLabels=True, dtype='train'):
X, Y, DepEdges, SrlEdges = zip(*batch)
x_paragraph, x_sentences, x_entity, x_len_paragraph, x_len_sentences, x_len_entity, x_mask_sentence, para_id_list, entity_list = get_placeholder_values(
X)
# print (x_entity)
feed_dict = {}
feed_dict[input_x_paragraph] = x_paragraph
feed_dict[input_x_sentence] = x_sentences
feed_dict[input_x_entity] = x_entity
feed_dict[input_x_len_paragraph] = x_len_paragraph
feed_dict[input_x_len_sentence] = x_len_sentences
feed_dict[input_x_len_entity] = x_len_entity
feed_dict[input_x_mask_sentence] = x_mask_sentence
if wLabels:
y_known, y_start, y_end = get_placeholder_labels(Y)
feed_dict[input_y_known] = y_known
feed_dict[input_y_start] = y_start
feed_dict[input_y_end] = y_end
de_adj_in, de_adj_out = get_adj(DepEdges, batch_size, paragraph_size,
num_deLabel)
srl_adj_in, srl_adj_out = get_adj(SrlEdges, batch_size, paragraph_size,
num_srlLabel)
for i in range(batch_size):
for lbl in range(num_deLabel):
feed_dict[de_adj_mat_in[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[de_adj_in[i][lbl].row, de_adj_in[i][lbl].col]).T,
values=de_adj_in[i][lbl].data,
dense_shape=de_adj_in[i][lbl].shape)
feed_dict[de_adj_mat_out[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[de_adj_out[i][lbl].row, de_adj_out[i][lbl].col]).T,
values=de_adj_out[i][lbl].data,
dense_shape=de_adj_out[i][lbl].shape)
for i in range(batch_size):
for lbl in range(num_srlLabel):
feed_dict[srl_adj_mat_in[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[srl_adj_in[i][lbl].row, srl_adj_in[i][lbl].col]).T,
values=srl_adj_in[i][lbl].data,
dense_shape=srl_adj_in[i][lbl].shape)
feed_dict[srl_adj_mat_out[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[srl_adj_out[i][lbl].row, srl_adj_out[i][lbl].col]).T,
values=srl_adj_out[i][lbl].data,
dense_shape=srl_adj_out[i][lbl].shape)
if dtype != 'train':
feed_dict[dropout] = 1.0
feed_dict[rec_dropout] = 1.0
return feed_dict, para_id_list, entity_list
def _convert_sparse_matrix_to_sparse_tensor(X, got_limit = False, limit = 5):
coo = X.tocoo()
indices = np.mat([coo.row, coo.col]).transpose()
if got_limit:
coo.data[coo.data > limit] = limit
return (
tf.SparseTensorValue(indices, coo.col, coo.shape),
tf.SparseTensorValue(indices, coo.data, coo.shape),
)
def testMakeOutputDictErrorSparse(self):
schema = self.toSchema({'a': tf.VarLenFeature(tf.string)})
# SparseTensor that cannot be represented as VarLenFeature.
fetches = {
'a':
tf.SparseTensorValue(indices=np.array([(0, 2), (0, 4), (0, 8)]),
values=np.array([10.0, 20.0, 30.0]),
dense_shape=(1, 20))
}
with self.assertRaisesRegexp(
ValueError, 'cannot be decoded by ListColumnRepresentation'):
impl_helper.to_instance_dicts(schema, fetches)
# SparseTensor of invalid rank.
fetches = {
'a':
tf.SparseTensorValue(indices=np.array([(0, 0, 1), (0, 0, 2),
(0, 0, 3)]),
values=np.array([10.0, 20.0, 30.0]),
dense_shape=(1, 10, 10))
}
with self.assertRaisesRegexp(
ValueError, 'cannot be decoded by ListColumnRepresentation'):
impl_helper.to_instance_dicts(schema, fetches)
# SparseTensor with indices that are out of order.
fetches = {
'a':
tf.SparseTensorValue(indices=np.array([(0, 2), (2, 4), (1, 8)]),
values=np.array([10.0, 20.0, 30.0]),
dense_shape=(3, 20))
}
with self.assertRaisesRegexp(ValueError,
'Encountered out-of-order sparse index'):
impl_helper.to_instance_dicts(schema, fetches)
# SparseTensors with different batch dimension sizes.
schema = self.toSchema({
'a': tf.VarLenFeature(tf.string),
'b': tf.VarLenFeature(tf.string)
})
fetches = {
'a':
tf.SparseTensorValue(indices=np.array([(0, 0)]),
values=np.array([10.0]),
dense_shape=(1, 20)),
'b':
tf.SparseTensorValue(indices=np.array([(0, 0)]),
values=np.array([10.0]),
dense_shape=(2, 20))
}
with self.assertRaisesRegexp(
ValueError,
r'Inconsistent batch sizes: "\w" had batch dimension \d, "\w" had batch'
r' dimension \d'):
impl_helper.to_instance_dicts(schema, fetches)
def create_feed_dict(self, batch, wLabels=True, dtype='train'):
X, Y, et_idx, ETEdges, DepEdges = zip(*batch)
x_pad, x_len, et_pad, et_mask, seq_len, max_et = self.pad_dynamic(
X, et_idx)
feed_dict = {}
feed_dict[self.input_x] = np.array(x_pad)
feed_dict[self.x_len] = np.array(x_len)
if wLabels: feed_dict[self.input_y] = np.array(Y)
feed_dict[self.et_idx] = np.array(et_pad)
feed_dict[self.et_mask] = np.array(et_mask)
feed_dict[self.seq_len] = seq_len
feed_dict[self.max_et] = max_et
et_adj_in, et_adj_out = self.get_adj(
ETEdges, self.p.batch_size, max_et + 1,
self.num_etLabel) # max_et + 1(DCT)
de_adj_in, de_adj_out = self.get_adj(DepEdges, self.p.batch_size,
seq_len, self.num_deLabel)
for i in range(self.p.batch_size):
for lbl in range(self.num_etLabel):
feed_dict[self.et_adj_mat_in[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[et_adj_in[i][lbl].row, et_adj_in[i][lbl].col]).T,
values=et_adj_in[i][lbl].data,
dense_shape=et_adj_in[i][lbl].shape)
feed_dict[self.et_adj_mat_out[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[et_adj_out[i][lbl].row, et_adj_out[i][lbl].col]).T,
values=et_adj_out[i][lbl].data,
dense_shape=et_adj_out[i][lbl].shape)
for lbl in range(self.num_deLabel):
feed_dict[self.de_adj_mat_in[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[de_adj_in[i][lbl].row, de_adj_in[i][lbl].col]).T,
values=de_adj_in[i][lbl].data,
dense_shape=de_adj_in[i][lbl].shape)
feed_dict[self.de_adj_mat_out[i][lbl]] = tf.SparseTensorValue(
indices=np.array(
[de_adj_out[i][lbl].row, de_adj_out[i][lbl].col]).T,
values=de_adj_out[i][lbl].data,
dense_shape=de_adj_out[i][lbl].shape)
if dtype != 'train':
feed_dict[self.dropout] = 1.0
feed_dict[self.rec_dropout] = 1.0
return feed_dict
def create_feed_dict(self, batch, wLabels=True, dtype='train'):
"""
Creates a feed dictionary for the batch
Parameters
----------
batch: contains a batch of bags
wLabels: Whether batch contains labels or not
split: Indicates the split of the data - train/valid/test
Returns
-------
feed_dict Feed dictionary to be fed during sess.run
"""
X, Y, et_idx, ETEdges, DepEdges = zip(*batch)
x_pad, x_len, et_pad, et_mask, seq_len, max_et = self.pad_dynamic(
X, et_idx)
feed_dict = {}
feed_dict[self.input_x] = np.array(x_pad)
feed_dict[self.x_len] = np.array(x_len)
if wLabels: feed_dict[self.input_y] = np.array(Y)
feed_dict[self.et_idx] = np.array(et_pad)
feed_dict[self.et_mask] = np.array(et_mask)
feed_dict[self.seq_len] = seq_len
feed_dict[self.max_et] = max_et
et_adj = self.get_adj(ETEdges, self.p.batch_size, max_et + 1,
self.num_etLabel) # max_et + 1(DCT)
de_adj = self.get_adj(DepEdges, self.p.batch_size, seq_len,
self.num_deLabel)
for i in range(self.p.batch_size):
for lbl in range(self.num_etLabel):
feed_dict[self.et_adj_mat[i][lbl]] = tf.SparseTensorValue(
indices=np.array([et_adj[i][lbl].row,
et_adj[i][lbl].col]).T,
values=et_adj[i][lbl].data,
dense_shape=et_adj[i][lbl].shape)
for lbl in range(self.num_deLabel):
feed_dict[self.de_adj_mat[i][lbl]] = tf.SparseTensorValue(
indices=np.array([de_adj[i][lbl].row,
de_adj[i][lbl].col]).T,
values=de_adj[i][lbl].data,
dense_shape=de_adj[i][lbl].shape)
if dtype != 'train':
feed_dict[self.dropout] = 1.0
feed_dict[self.rec_dropout] = 1.0
return feed_dict
def test_evaluate(self):
table = np.array([[1, 2, 1, 0, 0, 0], [0, 1, 2, 1, 0, 0],
[0, 0, 1, 2, 1, 0], [0, 0, 0, 1, 2, 1]])
md = np.array([[1, 2, 3, 4]]).T
md_holdout = np.array([[1.5, 2.5]]).T
table_holdout = np.array([[1, 2, 1, 0, 0, 0], [0, 1, 2, 1, 0, 0]],
dtype=np.float32)
N, D = table.shape
M, D = table_holdout.shape
p = md.shape[1]
table = coo_matrix(table)
table_holdout = coo_matrix(table_holdout)
opts = Options(batch_size=5,
num_neg_samples=3,
learning_rate=1e-1,
clipping_size=10,
beta_mean=0,
beta_scale=1,
gamma_mean=0,
gamma_scale=1)
for _ in range(10):
with tf.Graph().as_default(), tf.Session() as sess:
y_data = tf.SparseTensorValue(indices=np.array(
[table.row, table.col]).T,
values=table.data,
dense_shape=(N, D))
y_holdout = tf.SparseTensorValue(
indices=np.array([table_holdout.row, table_holdout.col]).T,
values=table_holdout.data,
dense_shape=table_holdout.shape)
G_data = tf.constant(md, dtype=tf.float32)
G_holdout = tf.constant(md_holdout, dtype=tf.float32)
model = PoissonRegression(opts, sess)
model.N = N
model.M = N
model.D = D
model.p = p
model.num_nonzero = table.nnz
batch = model.sample(y_data)
log_loss = model.loss(G_data, y_data, batch)
train = model.optimize(log_loss)
mad = model.evaluate(G_holdout, y_holdout)
tf.global_variables_initializer().run()
train_, mad_, loss_, beta, gamma = sess.run(
[train, mad, log_loss, model.qbeta, model.qgamma])
self.assertIsNotNone(beta)
self.assertIsNotNone(gamma)
# Look at mean absolute error
self.assertFalse(np.isnan(mad_))
def __init__(self,
sess,
n,
filename,
jump_prob=0.05,
drop_tol=1e-8,
verbose=False):
"""
Computes PPR using LU decomposition.
Args:
sess (Session): tensorflow session.
n (int): Number of nodes.
filename (str): A csv file denoting the graph.
jump_prob (float): Jumping probability of PPR.
drop_tol (float): Drops entries with absolute value lower than this value when computing inverse of LU.
verbose (bool): Prints step messages if True.
"""
self.alias = 'ludc'
self.verbose = verbose
self.pp("initializing")
self.sess = sess
self.n = n
self.c = jump_prob
d = 1 - self.c
t = drop_tol
exact = False
if t is None:
t = np.power(n, -0.5)
elif t == 0:
exact = True
self.pp("reading")
self.node2index, H = read_matrix(filename, d=-d, add_identity=True)
self.pp("sorting H")
self.perm = degree_reverse_rank_perm(H)
H = reorder_matrix(H, self.perm).tocsc()
self.pp("computing LU decomposition")
if exact:
self.LU = splu(H)
else:
self.LU = spilu(H, drop_tol=t)
Linv = inv(self.LU.L).tocoo()
Uinv = inv(self.LU.U).tocoo()
self.pp("tf init")
with tf.variable_scope('ppr_lu_decomposition_tf'):
t_Linv = tf.SparseTensorValue(list(zip(Linv.row, Linv.col)),
Linv.data,
dense_shape=self.LU.L.shape)
t_Uinv = tf.SparseTensorValue(list(zip(Uinv.row, Uinv.col)),
Uinv.data,
dense_shape=self.LU.U.shape)
self.t_q = tf.placeholder(tf.float64, shape=[self.n, 1])
self.t_r = _sdmm(t_Uinv, _sdmm(t_Linv, self.c * self.t_q))
def Train(self, inputs, ArcNode, target, step, nodegraph=0.0, mask=None):
''' train methods: has to receive the inputs, arch-node matrix conversion, target,
and optionally nodegraph indicator '''
# Creating a SparseTEnsor with the feeded ArcNode Matrix
arcnode_ = tf.SparseTensorValue(indices=ArcNode.indices,
values=ArcNode.values,
dense_shape=ArcNode.dense_shape)
if self.graph_based:
# lizx changed:
indices1, indices2 = nodegraph.nonzero()
nodegraph = tf.SparseTensorValue(
indices=np.stack([indices1, indices2], axis=1),
values=nodegraph[indices1, indices2],
dense_shape=nodegraph.shape)
if self.mask_flag:
fd = {
self.NodeGraph: nodegraph,
self.comp_inp: inputs,
self.state: np.zeros((ArcNode.dense_shape[0], self.state_dim)),
self.state_old: np.ones(
(ArcNode.dense_shape[0], self.state_dim)),
self.ArcNode: arcnode_,
self.y: target,
self.mask: mask
}
else:
fd = {
self.NodeGraph: nodegraph,
self.comp_inp: inputs,
self.state: np.zeros((ArcNode.dense_shape[0], self.state_dim)),
self.state_old: np.ones(
(ArcNode.dense_shape[0], self.state_dim)),
self.ArcNode: arcnode_,
self.y: target
}
if self.tensorboard:
_, loss, loop, merge_all, merge_tr = self.session.run([
self.train_op, self.loss, self.loss_op, self.merged_all,
self.merged_train
],
feed_dict=fd)
if step % 100 == 0:
self.writer.add_summary(merge_all, step)
self.writer.add_summary(merge_tr, step)
else:
_, loss, loop = self.session.run(
[self.train_op, self.loss, self.loss_op], feed_dict=fd)
return loss, loop[1]
def _SparseTensorValue_3x50(self, indices_dtype, values_dtype):
# NOTE: This input is intentionally not sorted to validate the
# already_sorted flag below.
ind = np.array([[0, 0], [1, 0], [1, 2], [2, 0], [2, 1], [1, 1]])
# NB: these are not sorted
indices = np.array([0, 13, 10, 33, 32, 14])
values = np.array([-3, 4, 1, 9, 5, 1])
shape = np.array([3, 3])
indices = tf.SparseTensorValue(np.array(ind, np.int64),
np.array(indices, indices_dtype),
np.array(shape, np.int64))
values = tf.SparseTensorValue(np.array(ind, np.int64),
np.array(values, values_dtype),
np.array(shape, np.int64))
return indices, values
def evaluate_simfunc(self, W_sparse_vals):
if self.eval_get_data is None:
self.eval_ids_i = tf.placeholder(tf.int32, shape=[None])
self.eval_ids_j = tf.placeholder(tf.int32, shape=[None])
self.eval_X_descr = tf.placeholder(tf.float32)
self.eval_get_data = self._sim_func(
X1=tf.gather(self.eval_X_descr, self.eval_ids_i),
X2=tf.gather(self.eval_X_descr, self.eval_ids_j))
ids_i = np.reshape(np.asarray(W_sparse_vals.indices[:, 0]), (-1, ))
ids_j = np.reshape(np.asarray(W_sparse_vals.indices[:, 1]), (-1, ))
data = np.zeros(ids_i.shape)
i = 0
while i < ids_i.shape[0]:
nxt_i = min(i + self.X.shape[0], ids_i.shape[0])
data[i:nxt_i] = self.sess.run(self.eval_get_data,
feed_dict={
self.W.indices:
W_sparse_vals.indices,
self.W.values:
W_sparse_vals.values,
self.eval_ids_i: ids_i[i:nxt_i],
self.eval_ids_j: ids_j[i:nxt_i],
self.eval_X_descr: self.X_descr
})
i = nxt_i
updated_W = tf.SparseTensorValue(W_sparse_vals.indices, data,
W_sparse_vals.dense_shape)
return updated_W
def testCopyTensorsProducesEquivalentTensors(self):
tensors = {
'dense': tf.placeholder(tf.int64, (None, ), name='my_dense_input'),
'sparse': tf.sparse_placeholder(tf.int64, name='my_sparse_input')
}
copied_tensors = impl_helper.copy_tensors(tensors)
with tf.Session() as session:
dense_value = [1, 2]
sparse_value = tf.SparseTensorValue(indices=[[0, 0], [0, 2],
[1, 1]],
values=[3, 4, 5],
dense_shape=[2, 3])
sample_tensors = session.run(copied_tensors,
feed_dict={
tensors['dense']: dense_value,
tensors['sparse']: sparse_value
})
self.assertAllEqual(sample_tensors['dense'], dense_value)
self.assertAllEqual(sample_tensors['sparse'].indices,
sparse_value.indices)
self.assertAllEqual(sample_tensors['sparse'].values,
sparse_value.values)
self.assertAllEqual(sample_tensors['sparse'].dense_shape,
sparse_value.dense_shape)
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
sentences = cast(Iterable[List[str]],
dataset.maybe_get_series(self.data_id))
if sentences is None and train:
raise ValueError("When training, you must feed "
"reference sentences")
if sentences is not None:
vectors, paddings = self.vocabulary.sentences_to_tensor(
list(sentences), train_mode=train, max_len=self.max_length)
# sentences_to_tensor returns time-major tensors, targets need to
# be batch-major
vectors = vectors.T
paddings = paddings.T
# Need to convert the data to a sparse representation
bool_mask = (paddings > 0.5)
indices = np.stack(np.where(bool_mask), axis=1)
values = vectors[bool_mask]
fd[self.train_targets] = tf.SparseTensorValue(
indices=indices, values=values, dense_shape=vectors.shape)
return fd
def testMaterializeFeaturesWithExcludes(self):
example1 = self._makeExample(age=3.0,
language='english',
label=1.0,
slice_key='first_slice')
features = {
'f': {
encoding.NODE_SUFFIX: np.array([1])
},
's': {
encoding.NODE_SUFFIX:
tf.SparseTensorValue(indices=[[0, 5], [1, 2], [3, 6]],
values=[100., 200., 300.],
dense_shape=[4, 10])
}
}
predictions = {'p': {encoding.NODE_SUFFIX: np.array([2])}}
labels = {'l': {encoding.NODE_SUFFIX: np.array([3])}}
extracts = {
constants.INPUT_KEY:
example1.SerializeToString(),
constants.FEATURES_PREDICTIONS_LABELS_KEY:
types.FeaturesPredictionsLabels(input_ref=0,
features=features,
predictions=predictions,
labels=labels)
}
result = feature_extractor._MaterializeFeatures(extracts,
excludes=['s'])
self.assertFalse('features__s' in result)
def as_tf_sparse(a):
"""Convert a to tf.SparseTensorValue
Args:
a: input array
Returns:
SparseTensorValue: converted object.
Examples:
>>> # numpy input
>>> a = np.random.random((3,4))
>>> a2 = as_tf_sparse(a)
>>> import tensorflow as tf
>>> isinstance(a2, tf.SparseTensorValue)
True
>>> np.testing.assert_array_equal(a, as_numpy_array(a2))
>>> a2 is as_tf_sparse(a2)
True
"""
_config.assert_has_package('tensorflow')
import tensorflow as tf
if isinstance(a, tf.SparseTensorValue):
return a
a = as_scipy_coo(a)
indices = np.asarray(np.mat([a.row, a.col]).transpose())
return tf.SparseTensorValue(indices, a.data, a.shape)
def sp2tf(sp_t, shape=None):
t = sparse_to_tuple(sp_t)
if shape is not None:
t[2] = shape
tensor = tf.SparseTensorValue(t[0], t[1].astype(np.float32), t[2])
return tensor
def load_training_labels(random_images_path):
"""
load labels
:param random_images_path:
:return:
"""
labels = []
max_size = 0
for p in random_images_path:
code_label = []
_, image_name = os.path.split(p)
l = str(image_name).split('-')[0].replace('#', '')
label_length = len(l)
if max_size < label_length:
max_size = label_length
for c in l:
code = char_mapping[c]
code_label.append(code)
labels.append(code_label)
indices = []
values = []
size = [len(labels), max_size]
for x in range(len(labels)):
for y in range(len(labels[x])):
indices.append([x, y])
values.append(labels[x][y])
ten = tf.SparseTensorValue(indices, values, size)
return ten
def feed_dict_predict(sentence, doc_positive_spt, on_training=True):
"""
input: data_sets is a dict and the value type is numpy
describe: to match the text classification the data_sets's content is the doc in df
"""
#该地方插入函数，把query_iin，doc_positive_in,doc_negative_in转化成one_hot，再转化成coo_matrix
query_in = query_input_list[0]
doc_positive_in = doc_positive_input_list[0]
doc_negative_in = doc_negative_input_list[0]
query = train_data_set.get_one_hot_from_sentence(sentence)
query = coo_matrix(query)
query = tf.SparseTensorValue(
np.transpose([
np.array(query.row, dtype=np.int64),
np.array(query.col, dtype=np.int64)
]), np.array(query.data, dtype=np.float),
np.array(query.shape, dtype=np.int64))
return {
query_in: query,
doc_positive_in: doc_positive_spt,
on_train: on_training
}
def shape_test_stack(self, feat):
np.random.seed(abs(myhash(feat.key)) % (2**31 - 1))
phs = [
feat.get_placeholder_and_feature(batch=False)[1]
for i in range(self.NTESTS)
]
vals_dense = [
np.random.binomial(1, 0.1, size=feat.shape)
for i in range(self.NTESTS)
]
vals_sparse = []
for A in vals_dense:
idx, vals = basefeat.np_dense_to_sparse(A)
vals_sparse.append(
tf.SparseTensorValue(np.stack(idx, -1), vals, feat.shape))
X = feat.stack(phs)
X_ = self.sess.run(X, feed_dict=dict(zip(phs, vals_sparse)))
total_len = sum([len(v.values) for v in vals_sparse])
run_len = len(X_.values)
msg = 'Differing number of values: {} vs {}'.format(run_len, total_len)
self.assertEqual(run_len, total_len, msg)
for i, v in enumerate(vals_sparse):
for vval, vinds in zip(v.values, v.indices):
value_found = False
for xval, xinds in zip(X_.values, X_.indices):
if xinds[0] == i and np.allclose(xinds[1:], vinds):
msg = 'Values {} and {} not equal with inds {}'.format(
vval, xval, xinds)
self.assertEqual(xval, vval)
value_found = True
self.assertTrue(value_found,
msg='Values {}, indices {} not found'.format(
vval, vinds))
def sparse_to_tensor(value):
"""Convert a scipy sparse matrix to a tensorflow SparseTensorValue."""
row = np.reshape(value.row, (-1, 1))
col = np.reshape(value.col, (-1, 1))
indices = np.concatenate((row, col), axis=1)
return tf.SparseTensorValue(indices, value.data, value.shape)
def _merge_models(self, m1, m2):
sparse_merges = ['context', 'features']
dense_merges = ['seq_len', 'label', 'forloeb']
ignores = ['test', 'doc_ids', 'text_len']
new_model = {}
for key in sparse_merges:
new_shape = np.concatenate(
[[m1[key].dense_shape[0] + m2[key].dense_shape[0]],
m1[key].dense_shape[1:]],
axis=0)
new_indices = m1[key].indices
to_append = m2[key].indices
to_append[:, 0] += m1[key].dense_shape[0]
new_indices = np.concatenate([new_indices, to_append], axis=0)
new_values = np.concatenate([m1[key].values, m2[key].values],
axis=0)
new_model[key] = tf.SparseTensorValue(new_indices, new_values,
new_shape)
for key in dense_merges:
new_model[key] = np.concatenate([m1[key], m2[key]], axis=0)
for key in ignores:
new_model[key] = None
return new_model
def batch_to_feed_dict(batch, is_train, add_noise=False):
'''
Create the dictionnary that is fed into the Session.run(..) calls.
:param batch:
:return:
'''
features_np, phns_np, seq_len_np, wav_np = batch
n_time = max([len(i) for i in wav_np])
wav_np = np.stack([pad_vector(w, n_time) for w in wav_np], axis=0)
n_batch, n_time, n_features = features_np.shape
relevance_mask_np = [(np.arange(n_time) < seq_len_np[i]) / seq_len_np[i]
for i in range(n_batch)]
relevance_mask_np = np.array(relevance_mask_np)
if FLAGS.n_repeat > 1:
# Extend sequences with the repeat in time
features_np = np.repeat(features_np, FLAGS.n_repeat, axis=1)
n_batch, n_time, n_features = features_np.shape
phns_labels = tf.SparseTensorValue(phns_np['indices'], phns_np['values'],
[n_batch, n_time])
return {
features: features_np,
phns: phns_labels,
seq_len: seq_len_np,
weighted_relevant_mask: relevance_mask_np,
keep_prob: FLAGS.drop_out_probability if is_train else 1.,
batch_size: n_batch,
noise_gate: 1. if is_train else 0.,
audio: wav_np,
gd_noise: FLAGS.gd_noise if add_noise else 0.
}
def _run_graph(self, sess, qq, hh, tt, mdb, to_fetch):
feed = {}
if not self.query_is_language:
feed[self.queries] = [[q] * (self.num_step - 1) + [self.num_query]
for q in qq]
else:
feed[self.queries] = [[q] * (self.num_step - 1) +
[[self.num_vocab] * self.num_word]
for q in qq]
feed[self.heads] = hh
feed[self.tails] = tt
# In[232]: data.matrix_db[0][0][:7]
# Out[232]:
# [[0, 0],
# [2675, 2698],
# [2268, 2274],
# [2713, 1240],
# [1978, 2013],
# [2417, 1881],
# [2068, 701]]
#
# In[233]: data.matrix_db[0][1][:7]
# Out[233]: [0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
#
# In[234]: data.matrix_db[0][2]
# Out[234]: [3007, 3007]
for r in xrange(self.num_operator / 2):
feed[self.database[r]] = tf.SparseTensorValue(
*mdb[r]
) ## AJAY NOTE: constructs a sparse tensor from indices, values and dense shape for every relation ( by 2 is for symmetric relations)
fetches = to_fetch
graph_output = sess.run(fetches, feed)
return graph_output
def _evaluate(self, sess, source='test'):
kg = self.kg
hp = self.hparams
triples_eval = []
idx_tail_pred = []
for batch in kg.get_eval_batch(hp.batch_size, source=source):
head, query, tail = kg.get_triples_from_batch(batch)
kg_inputs = kg.get_kg(batch)
feed = {
self.head_input: head,
self.tail_input: tail,
self.query_sequence: [[q] * (hp.n_steps - 1) + kg.END_RELATION for q in query]
}
for r in range(kg.n_relations):
feed[self.kg_inputs[r]] = tf.SparseTensorValue(*kg_inputs[r])
loss, predication = sess.run([self.total_loss, self.prediction], feed_dict=feed)
idx_pred = np.argsort(-predication)
idx_tail_pred.append(idx_pred)
triples_eval.append(np.stack([head, query, tail], axis=-1))
idx_tail_pred = np.concatenate(idx_tail_pred, 0).tolist()
triples_eval = np.concatenate(triples_eval, 0).tolist()
_, eval_tail, _ = kg.calc_metrics(triples_eval, idx_tail_pred)
return eval_tail
def convert_sparse_matrix_to_sparse_tensor(X):
'''
code borrowed from https://stackoverflow.com/questions/40896157/scipy-sparse-csr-matrix-to-tensorflow-sparsetensor-mini-batch-gradient-descent
'''
coo = X.tocoo()
indices = np.mat([coo.row, coo.col]).transpose()
return tf.SparseTensorValue(indices, coo.data, coo.shape)
def convert2SparseTensorValue(list_labels):
#
# list_labels: batch_major
#
#
#print(list_labels)
#
num_samples = len(list_labels)
num_maxlen = max(map(lambda x: len(x), list_labels))
#
indices = []
values = []
shape = [num_samples, num_maxlen]
#
for idx in range(num_samples):
#
item = list_labels[idx]
#
values.extend(item)
indices.extend([[idx, posi] for posi in range(len(item))])
#
#
return tf.SparseTensorValue(indices=indices,
values=values,
dense_shape=shape)
def Apply(self):
audio = self.feed_dict[self.model.get_data_layer().input_tensors["source_tensors"][0]]
audio_length = self.feed_dict[self.model.get_data_layer().input_tensors["source_tensors"][1]]
x_id = self.feed_dict[self.model.get_data_layer().input_tensors["source_ids"][0]]
internal_request = predict_pb2.PredictRequest()
internal_request.model_spec.name = 'deepspeech2'
internal_request.model_spec.signature_name = 'predict_output'
internal_request.inputs['audio'].CopyFrom(
tf.contrib.util.make_tensor_proto(audio, shape=list(audio.shape)))
internal_request.inputs['audio_length'].CopyFrom(
tf.contrib.util.make_tensor_proto(audio_length, shape=list(audio_length.shape)))
internal_request.inputs['x_id'].CopyFrom(
tf.contrib.util.make_tensor_proto(x_id, shape=list(x_id.shape)))
internal_result = self.istub.Predict(internal_request, 10.0) # 5 seconds
self.inputs = Deepspeech2.model.get_data_layer().input_tensors
indices_decoded_sequence = tensor_util.MakeNdarray(
internal_result.outputs['indices_decoded_sequence'])
values_decoded_sequence = tensor_util.MakeNdarray(
internal_result.outputs['values_decoded_sequence'])
dense_shape_decoded_sequence = tensor_util.MakeNdarray(
internal_result.outputs['dense_shape_decoded_sequence'])
outputs = tf.SparseTensorValue(indices=indices_decoded_sequence,
values=values_decoded_sequence,
dense_shape=dense_shape_decoded_sequence)
self.outputs = [outputs]
results = Deepspeech2.model.infer(self.inputs, self.outputs)
self.final_result = results[0][0]
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {} # type: FeedDict
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
fd[self.train_mode] = train
if sentences is not None:
vectors, paddings = self.vocabulary.sentences_to_tensor(
list(sentences), train_mode=train)
# sentences_to_tensor returns time-major tensors, targets need to
# be batch-major
vectors = vectors.T
paddings = paddings.T
# Need to convert the data to a sparse representation
bool_mask = (paddings > 0.5)
indices = np.stack(np.where(bool_mask), axis=1)
values = vectors[bool_mask]
fd[self.train_targets] = tf.SparseTensorValue(
indices=indices, values=values, dense_shape=vectors.shape)
return fd
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 feed_step(batch, label, dtype="train"):
indices = np.mat([batch.tocoo().row,
batch.tocoo().col]).transpose()
values = batch.tocoo().data
shape = batch.tocoo().shape
feed_dict = {
model.input_x: tf.SparseTensorValue(indices, values, shape),
model.input_y: np.reshape(label, [-1, 1])
}
if dtype == "train":
train_operation = [
model.train_step, model.global_step, model.loss,
model.auc_score, model.summary_op
]
_, step, loss_val, auc_val, merged = sess.run(
train_operation, feed_dict=feed_dict)
writer.add_summary(merged, step)
return step, loss_val, auc_val
elif dtype == "evaluate":
loss_val, auc_val = sess.run([model.loss, model.auc_score],
feed_dict=feed_dict)
return loss_val, auc_val
elif dtype == "test":
logit_val = sess.run([model.logit], feed_dict=feed_dict)[0]
print("accuracy:",
accuracy_score(label, [round(x) for x in logit_val]))
print("precision:",
precision_score(label, [round(x) for x in logit_val]))
print("reall:",
recall_score(label, [round(x) for x in logit_val]))
print("f1:", f1_score(label, [round(x) for x in logit_val]))
print("auc:", roc_auc_score(label, logit_val))
def _SparseTensorValue_5x6(self):
ind = np.array([[0, 0], [1, 0], [1, 3], [1, 4], [3, 2],
[3, 3]]).astype(np.int64)
val = np.array([0, 10, 13, 14, 32, 33]).astype(np.float64)
shape = np.array([5, 6]).astype(np.int64)
return tf.SparseTensorValue(ind, val, shape)