gistfile1.txt

class _SequentialAutoregressive(tf.contrib.distributions.Distribution):

    def __init__(self, f_process, base_dist_cls, initial_dist,
                 n_time_steps=None,
                 dtype=tf.float32,
                 name='sequential_auto_regressive'):
        self.f_process = f_process
        self.base_dist_cls = base_dist_cls
        self.initial_dist = initial_dist
        self.n_time_steps = n_time_steps

        super(_SequentialAutoregressive, self).__init__(
            dtype=dtype,
            parameters={},
            is_continuous=True,
            is_reparameterized=False,
            validate_args=False,
            allow_nan_stats=False,
            name=name
        )

    def _log_prob(self, value):
        log_pdf0 = self.initial_dist.log_pdf(value[0, :, :])
        _, stats = self.f_process(
            value[:-1, :, :],
            value.get_shape()[0].value - 1, None)
        rv = self.base_dist_cls(*stats)
        log_pdfs = rv.log_pdf(value[1:, :, :])

        return tf.concat(0, [tf.expand_dims(log_pdf0, 0), log_pdfs])

    def _sample_n(self, n, seed=None):
        obs = [tf.reshape(self.initial_dist.sample_n(n), (n, -1))]
        state_tm1 = None
        or i in range(self.n_time_steps - 1):
            states, stats = self.f_process(obs[-1:], 1, state_tm1)
            state_tm1 = states[-1]
            rv = self.base_dist_cls(*stats)
            obs.append(tf.reshape(rv.sample(1)[0], (1, -1)))
        return tf.pack(obs)

    def _sample(self, sample_shape=(), seed=None):
        raise NotImplemented
        
    def _get_batch_shape(self):
        return tf.TensorShape([self.n_time_steps])

    def _get_event_shape(self):
        return tf.TensorShape([None, None])