def testConvOutputKroneckerFactorInit(self): with tf.Graph().as_default(): tf.set_random_seed(200) tensor = tf.ones((2, 3, 4, 5), name='a/b/c') factor = ff.ConvOutputKroneckerFactor(((tensor, ), )) factor.instantiate_cov_variables() self.assertEqual([5, 5], factor.cov.get_shape().as_list())
def testMakeCovarianceUpdateOp(self): with tf.Graph().as_default(), self.test_session() as sess: tf.set_random_seed(200) tensor = np.arange(1, 17).reshape(2, 2, 2, 2).astype(np.float32) factor = ff.ConvOutputKroneckerFactor(((tf.constant(tensor), ), )) factor.instantiate_cov_variables() sess.run(tf.global_variables_initializer()) new_cov = sess.run(factor.make_covariance_update_op(.5)) self.assertAllClose([[43, 46.5], [46.5, 51.5]], new_cov)
def testConvOutputKroneckerFactorInitFloat64(self): with tf.Graph().as_default(): dtype = dtypes.float64_ref tf.set_random_seed(200) tensor = tf.ones((2, 3, 4, 5), dtype=dtype, name='a/b/c') factor = ff.ConvOutputKroneckerFactor(((tensor, ), )) factor.instantiate_cov_variables() cov = factor.cov self.assertEqual(cov.dtype, dtype) self.assertEqual([5, 5], cov.get_shape().as_list())
def test3DConvolution(self): with tf.Graph().as_default(): batch_size = 1 width = 3 out_channels = width**3 factor = ff.ConvOutputKroneckerFactor(outputs_grads=([ tf.random_uniform( (batch_size, width, width, width, out_channels), seed=0) ], )) factor.instantiate_cov_variables() with self.test_session() as sess: sess.run(tf.global_variables_initializer()) sess.run(factor.make_covariance_update_op(0.0)) cov = sess.run(factor.cov) # Cov should be rank 3^3, as each spatial position donates a rank-1 # update. self.assertMatrixRank(width**3, cov)