在tensorflow中添加新的op-形状函数

Question

I'm trying to add a new operation in Tensorflow where I have two inputs, namely a 3D tensor and a constant, which outputs a 4D tensor. The 4D tensor is obtained by replicating the 3D tensor a number of times defined by the constant. The shape function is implemented in the following way:

.SetShapeFn([](::tensorflow::shape_inference::InferenceContext* c)
{
    ::tensorflow::shape_inference::ShapeHandle output;
    ::tensorflow::shape_inference::ShapeHandle out1 = c->Vector(::tensorflow::shape_inference::DimensionOrConstant(5));
    TF_RETURN_IF_ERROR(c->Concatenate(c->input(0),out1,&output));
    c->set_output(0,output);
    return Status::OK();
})
.Doc(R"doc(
     Replicating the 3D input tensor in a 4D tensor.
)doc");

I would like that the size of the fourth dimension (defined by out1 in the code) is set to the second input (namely the constant value). How to do it?

Answer 1

Perhaps MakeShapeFromShapeTensor is what you're looking for? Something like:

.SetShapeFn([](::tensorflow::shape_inference::InferenceContext* c)
{
    ::tensorflow::shape_inference::ShapeHandle n;
    TF_RETURN_IF_ERROR(c->MakeShapeFromShapeTensor(1, &n));
    ::tensorflow::shape_inference::ShapeHandle out;
    TF_RETURN_IF_ERROR(c->Concatenate(n, c->input(0), &out));
    c->set_output(0, out);
    return Status::OK();
})

That said, you probably know this, but just to be sure: Element-wise arithmetic operations in TensorFlow support broadcasting , so at least in those case you shouldn't need this custom op.

For other cases, you could also combine tf.tile , tf.shape , tf.concat and tf.reshape to achieve the same effect. For example, the following creates a matrix by repeating a vector:

import tensorflow as tf
oneD = tf.constant([1,2])
n = tf.constant([5])
twoD = tf.reshape(tf.tile(oneD, n), tf.concat([n, tf.shape(oneD)], 0))

with tf.Session() as sess:
  print oneD.eval()
  print twoD.eval()