Impossible to add multi-dimension constant tensor via tensorflow2ncnn

[Corea]

This network cannot be translated to ncnn via tensorflow2ncnn.

shape = (1, 2, 2, 2)

images = tf.placeholder(tf.float32, shape=shape, name="input/image")
const = tf.constant([1, 2, 3, 4, 5, 6, 7, 8], dtype=tf.float32, shape=shape)
layer = tf.add(images, const)
output = tf.identity(layer, name="output/score")

[Corea]

Maybe need to check nullity before use refcount?
https://github.com/Tencent/ncnn/blob/master/src/net.cpp#L555
https://github.com/Tencent/ncnn/blob/master/src/net.cpp#L607

[Corea]

And these two networks are translated to same bin/proto.

shape = (1, 1, 1, 8)

images = tf.placeholder(tf.float32, shape=shape, name="input/image")
const = tf.constant([1, 2, 3, 4, 5, 6, 7, 8], dtype=tf.float32, shape=shape)
layer = tf.add(const, images)
output = tf.identity(layer, name="output/score")

shape = (1, 2, 2, 2)

images = tf.placeholder(tf.float32, shape=shape, name="input/image")
const = tf.constant([1, 2, 3, 4, 5, 6, 7, 8], dtype=tf.float32, shape=shape)
layer = tf.add(images, const)
output = tf.identity(layer, name="output/score")

[nihui]

TODO
convert constant to memorydata layer