TensorFlow documentation提到sp_a (indices, values)期望的SparseTensor格式为：[0, 1]: a
[1, 0]: b
[1, 4]: c
[2, 2]: d
：
embedding_lookup_sparse sp_ids sp_weights期望的SparseTensor格式：[0, 0]: 1 [0, 0]: a
[1, 0]: 0 [1, 0]: b
[1, 1]: 4 [1, 1]: c
[2, 0]: 2 [2, 0]: d
sp_a如何在TensorFlow中将sp_ids转换为sp_weights和def clean(condition):
if type == "object" then
if condition
then empty
else
with_entries(
if (.value|type) == "object" and (.value|condition)
then empty
else .value |= clean(condition)
end
)
end
elif type == "array" then
map(
if type == "object" and condition
then empty
else clean(condition)
end
)
else .
end
;
clean(
has("field") and (.field == "status")
)
秒？如果不可能，我怎么能在numpy中做到？
答案 0 :(得分：0)
我忽略了Tensorflow API中是否存在此转换的函数，但这是我将sp_a转换为sp_ids和sp_weights的方式：
import tensorflow as tf
indices = tf.constant([[0, 1],
[1, 0],
[1, 4],
[2, 2]], dtype=tf.int64)
values = tf.constant([1, 2, 3, 4]) # a, b, c, d
dense_shape = tf.constant([3, 5], dtype=tf.int64)
sp_a = tf.SparseTensor(indices=indices,
values=values,
dense_shape=dense_shape)
# transform sp_a into sp_ids and sp_weights
# Get sp_ids values
sp_ids_values = tf.slice(sp_a.indices,
begin=[0, 1],
size=[-1, 1])
sp_ids_values = tf.squeeze(sp_ids_values)
# Get the indices for sp_ids and sp_weights
d1 = tf.slice(sp_a.indices,
begin=[0, 0],
size=[-1, 1])
d2 = tf.expand_dims(scan_accum(tf.squeeze(d1)),
axis=1)
indices_ = tf.concat([d1, d2],
axis=1)
# Build sp_ids and sp_weights
sp_ids = tf.SparseTensor(indices=indices_,
values=sp_ids_values,
dense_shape=sp_a.dense_shape)
sp_weights = tf.SparseTensor(indices=indices_,
values=sp_a.values,
dense_shape=sp_a.dense_shape)
with tf.Session() as sess:
print(sess.run(sp_ids))
print(sess.run(sp_weights))
我定义scan_accum here。