Chapter 1 | Joint Log Probability defined differently | TFB version

[zachlefevre]

The first joint log probability function is defined as:

def joint_log_prob(count_data, lambda_1, lambda_2, tau):
    tfd = tfp.distributions

    alpha = np.array(1. / count_data.mean(), np.float32)
    rv_lambda_1 = tfd.Exponential(rate=alpha)
    rv_lambda_2 = tfd.Exponential(rate=alpha)

    rv_tau = tfd.Uniform()

    lambda_ = tf.gather(
         [lambda_1, lambda_2],
         indices=tf.to_int32(tau * count_data.size <= np.arange(count_data.size)))
    rv_observation = tfd.Poisson(rate=lambda_)

    return (
         rv_lambda_1.log_prob(lambda_1)
         + rv_lambda_2.log_prob(lambda_2)
         + rv_tau.log_prob(tau)
         + tf.reduce_sum(rv_observation.log_prob(count_data))
    )

Which is defined again below as:

def joint_log_prob(count_data, lambda_1, lambda_2, tau):
    tfd = tfp.distributions

    alpha = (1. / tf.reduce_mean(count_data))
    rv_lambda_1 = tfd.Exponential(rate=alpha)
    rv_lambda_2 = tfd.Exponential(rate=alpha)

    rv_tau = tfd.Uniform()

    lambda_ = tf.gather(
         [lambda_1, lambda_2],
         indices=tf.to_int32(tau * tf.to_float(tf.size(count_data)) <= tf.to_float(tf.range(tf.size(count_data)))))
    rv_observation = tfd.Poisson(rate=lambda_)

    return (
         rv_lambda_1.log_prob(lambda_1)
         + rv_lambda_2.log_prob(lambda_2)
         + rv_tau.log_prob(tau)
         + tf.reduce_sum(rv_observation.log_prob(count_data))
    )

The first will not run, as the Tensor object does not have a mean method

[CamDavidsonPilon]

cc @matthew-mcateer want to take a look?

[Pindar777]

Hi there, the second version is for TF, the first includes numpy-functions. I suggest to get rid of the first variant and just state this code in order to demonstrate the "lambda-function":

lambda_ = tf.gather( [lambda_1, lambda_2], indices=tf.to_int32(tau * tf.to_float(tf.size(count_data)) <= tf.to_float(tf.range(tf.size(count_data))))) rv_observation = tfd.Poisson(rate=lambda_)

[matthew-mcateer]

THanks for pointing this out, @Pindar777 I'll take a look