CODEWITHSUNDEEP
python-2.7tensorflow
lurscher
lurscher

Reputation: 26983

Why `tf.constant_initializer` does not take a constant tensor?

It seems pretty silly to me that Tensorflow has decided to make a constant_initializer op that takes only scalar values. It would make a lot of sense to be able to initialize variables with constant tensors:

tf.get_variable('some_var', shape = [4,3], initializer=tf.constant_initializer(tf.constant([[0,0,0], [0,0,1],[0,1,0],[1,0,0]])))

Is using placeholders and feed_dict the only way to initialize tensor variables to custom values? This forces one to have declarations and data initializations in separated places, which is a hassle

Upvotes: 2

Views: 4983

Answers (2)

junqin zhang
junqin zhang

Reputation: 19

Remove the "tf.constant", do like below, works in TF 1.13

tf.get_variable('some_var', shape = [4,3], initializer=tf.constant_initializer([[0,0,0], [0,0,1],[0,1,0],[1,0,0]]))

Upvotes: 1

mrry
mrry

Reputation: 126184

The tf.constant_initializer() function might not accept a tf.Tensor as an argument, but tf.get_variable() does accept a tf.Tensor as its initializer argument. This means you can write:

v = tf.get_variable('some_var', initializer=tf.constant([[0, 0, 0],
                                                         [0, 0, 1],
                                                         [0, 1, 0],
                                                         [1, 0, 0]]))

...which requires even fewer characters!

The reason tf.constant_initializer() doesn't take an arbitrary tensor is that it is designed to initialize variables of many different shapes with the same constant value for each element. For example, a statement like:

v = tf.get_variable('some_var', shape=[15, 37], initializer=tf.constant_initializer(
    tf.constant([[0, 0, 0],
                 [0, 0, 1],
                 [0, 1, 0],
                 [1, 0, 0]])))

...wouldn't make much sense. Arguably we could make tf.constant_initializer() accept a scalar tf.Tensor, and then it would have semantics similar to tf.fill(), but we haven't had any demand for that yet. Feel free to raise a GitHub issue if it would be useful though!

Upvotes: 5

Related Questions