def ancestral_sample()

in ais.py [0:0]

• 22 lines of code
• 6 McCabe index (conditional complexity)

def ancestral_sample(e_func, weights, batch_size=128, prop_dist=10, temp=1, hmc_step=10):
    if FLAGS.dataset == "2d":
        x = tf.placeholder(tf.float32, shape=(None, 2))
    elif FLAGS.dataset == "gauss":
        x = tf.placeholder(tf.float32, shape=(None, FLAGS.gauss_dim))
    elif FLAGS.dataset == "mnist":
        x = tf.placeholder(tf.float32, shape=(None, 28, 28))
    else:
        x = tf.placeholder(tf.float32, shape=(None, 32, 32, 3))

    x_init = x

    alpha_prev = tf.placeholder(tf.float32, shape=())
    alpha_new = tf.placeholder(tf.float32, shape=())
    approx_lr = tf.placeholder(tf.float32, shape=())

    chain_weights = tf.zeros(batch_size)
    # for i in range(1, prop_dist+1):
    #     print("processing loop {}".format(i))
    #     alpha_prev = (i-1) / prop_dist
    #     alpha_new = i / prop_dist

    prob_log_old_neg = bridge_prob_neg_log(alpha_prev, x, e_func, weights, temp)
    prob_log_new_neg = bridge_prob_neg_log(alpha_new, x, e_func, weights, temp)

    chain_weights = -prob_log_new_neg + prob_log_old_neg
    # chain_weights = tf.Print(chain_weights, [chain_weights])

    # Sample new x using HMC
    def unorm_prob(x):
        return bridge_prob_neg_log(alpha_new, x, e_func, weights, temp)

    for j in range(1):
        x = hmc(x, approx_lr, hmc_step, unorm_prob)

    return chain_weights, alpha_prev, alpha_new, x, x_init, approx_lr