def test_inference_results(self):
np.random.seed(123)
tf.set_random_seed(123)
num_dims = 20
rg = region_graph.RegionGraph(range(num_dims))
for _ in range(0, 10):
rg.random_split(2, 3)
args = RAT_SPN.SpnArgs()
args.normalized_sums = True
spn = RAT_SPN.RatSpn(10, region_graph=rg, name="obj-spn", args=args)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
dummy_input = np.random.normal(0.0, 1.2, [10, num_dims])
input_ph = tf.placeholder(tf.float32, [10, num_dims])
output_tensor = spn.forward(input_ph)
tf_output = sess.run(output_tensor, feed_dict={input_ph: dummy_input})
output_nodes = spn.get_simple_spn(sess)
simple_output = []
for node in output_nodes:
simple_output.append(Inference.likelihood(node, dummy_input))
simple_output = np.stack(simple_output)
deviation = simple_output / np.exp(tf_output)
rel_error = np.abs(deviation - 1.0)
# print(rel_error)
self.assertTrue(np.all(rel_error < 1e-2))
num_correct_batch = np.sum(max_idx == label_batch)
num_correct += num_correct_batch
acc = num_correct / (batch_size * batches_per_epoch)
print(i, acc, cur_loss)
def softmax(x, axis=0):
"""Compute softmax values for each sets of scores in x."""
e_x = np.exp(x - np.max(x, axis=axis, keepdims=True))
return e_x / e_x.sum(axis=axis, keepdims=True)
if __name__ == "__main__":
rg = region_graph.RegionGraph(range(28 * 28))
# rg = region_graph.RegionGraph(range(3 * 3))
for _ in range(0, 2):
rg.random_split(2, 1)
args = RAT_SPN.SpnArgs()
args.normalized_sums = True
args.num_sums = 2
args.num_gauss = 2
spn = RAT_SPN.RatSpn(10, region_graph=rg, name="obj-spn", args=args)
print("num_params", spn.num_params())
sess = tf.Session()
sess.run(tf.global_variables_initializer())
(train_im, train_labels), _ = load_mnist()