def test_complex_complex_mm():
"""
Complex mtx x complex mtx matrix multiply
:return:
"""
c = ComplexTensor(torch.zeros(4, 3)) + 1
cc = c.mm(c.t())
cc = cc.view(-1).data.numpy()
# do the same in numpy
np_c = np.ones((2, 3)).astype(np.complex64)
np_cc = np.matmul(np_c, np_c.T)
# compare
np_cc = np_cc.flatten()
np_cc = list(np_cc.real) + list(np_cc.imag)
assert np.array_equal(np_cc, cc)
def test_grad():
"""
Grad calculated first with tensorflow
:return:
"""
c = ComplexTensor([[1, 3, 5], [7, 9, 11], [2, 4, 6], [8, 10, 12]])
c.requires_grad = True
# simulate some ops
out = c + 4
out = out.mm(c.t())
# calc grad
out = out.sum()
out.backward()
# d_out/dc
g = c.grad.view(-1).data.numpy()
# solution (as provided by running same ops in tensorflow)
"""
tf_c2 = tf.constant([[1+2j, 3+4j, 5+6j], [7+8j,9+10j,11+12j]], dtype=tf.complex64)
with tf.GradientTape() as t:
t.watch(tf_c2)
tf_out = tf_c2 + 4
tf_out = tf.matmul(tf_out, tf.transpose(tf_c2, perm=[1,0]))
tf_y = tf.reduce_sum(tf_out)
dy_dc2 = t.gradient(tf_y, tf_c2)
# solution
print(dy_dc2)
"""
#
sol = np.asarray([24, 32, 40, 24, 32, 40, -20, -28, -36, -20, -28, -36])
assert np.array_equal(g, sol)