public void tf_partial_shape_test()
{
// Note: Keras/TensorFlow represent unknown dimensions
// as None, whereas TensorFlowSharp represents as -1:
/*
* import keras
*
* from keras.models import Sequential
* from keras.layers import Dense
* from keras import backend as K
* import numpy as np
*
* a = K.placeholder(shape = (None, 2))
* b = K.variable(np.matrix([[ 1, 2, 3], [4, 5, 6]]))
* ab = K.dot(a, b)
*
* shape_a = K.int_shape(a)
* shape_b = K.int_shape(b)
* shape_ab = K.int_shape(ab)
*
* print(shape_a)
* print(shape_b)
* print(shape_ab)
*
* >>> Using TensorFlow backend.
* (None, 2)
* (2, 3)
* (None, 3)
*/
using (var K = new TensorFlowBackend())
{
Tensor a = K.placeholder(shape: new int?[] { null, 2 });
Tensor b = K.variable(array: new float[, ] {
{ 1, 2, 3 },
{ 4, 5, 6 }
});
var ab = K.dot(a, b);
int?[] shape_a = K.int_shape(a);
int?[] shape_b = K.int_shape(b);
int?[] shape_ab = K.int_shape(ab);
long[] tf_shape_a = K.In(a).TF_Shape;
long[] tf_shape_b = K.In(b).TF_Shape;
long[] tf_shape_ab = K.In(ab).TF_Shape;
AssertEx.AreEqual(new int?[] { null, 2 }, shape_a);
AssertEx.AreEqual(new int?[] { 2, 3 }, shape_b);
AssertEx.AreEqual(new int?[] { null, 3 }, shape_ab);
AssertEx.AreEqual(new long[] { -1, 2 }, tf_shape_a);
AssertEx.AreEqual(new long[] { 2, 3 }, tf_shape_b);
AssertEx.AreEqual(new long[] { -1, 3 }, tf_shape_ab);
}
}