SMC2

[aleslamitz]

Added SMC2 from https://arxiv.org/abs/1101.1528

[jburnim]

This is looking great, thanks!

I'm still reviewing, but I left some initial comments.

[jburnim]

I don't think this added code should be necessary. The lines right below this should already handle any necessary expand_dims for particles_dim.

[aleslamitz]

I think it is necessary due to the dimensions mismatch

[jburnim]

With these two lines included, ParticleFilterTestFloat32.test_batch_of_filters_particles_dim_1 fails when I run it. If I delete these two lines, then that tests passes.

[jburnim]

With these two lines included, ParticleFilterTestFloat32.test_batch_of_filters_particles_dim_1 fails when I run it. If I delete these two lines, then that tests passes.

[jburnim]

I think this should be:

inner_initial_state_prior=lambda _, params: mvn_diag.MultivariateNormalDiag(
    loc=tf.broadcast_to([0., 0.], params.shape + [2]),
    scale_diag=tf.broadcast_to([0.01, 0.01], params.shape + [2])

Because:

• params.shape is just the batch shape (i.e., [num_outer_particles])
• Which will give inner_inital_state_prior(0, params) the batch_shape of [num_outer_particles] and the correct event_shape of [2]. (And the SMC^2 code is then responsible for drawing num_inner_particles samples from this distribution and re-arranging the dimensions to get initial particles of shape [num_outer_particles, num_inner_particles, 2].)

[jburnim]

params is has shape [num_outer_particles] and previous_state has shape [num_outer_particles, num_inner_particles, 2], so here it should be sufficient to just do:

previous_state + params[..., tf.newaxis, tf.newaxis] + 1

[jburnim]

It should work to just return cond, instead of using tf.cond(cond, ...).

[jburnim]

I would hope this assertion would pass if we changed this line to observed_positions[..., tf.newaxis, tf.newaxis].

But if that didn't work, we could also do: tf.broadcast_to(observation_positions, [num_timesteps] + batch_shape).