def adjacent_swaps(num_replica, batch_shape=(), step_count=None, seed=None): """Make random shuffle using only one time swaps.""" del step_count # Unused for this function. with tf.name_scope(name or 'adjacent_swaps'): parity_seed, proposal_seed = samplers.split_seed(seed) # u selects parity. E.g., # u==False ==> [1, 0, 3, 2, 4] even parity swaps # u==True ==> [0, 2, 1, 4, 3] odd parity swaps # If there are only 2 replicas, then the "True" swaps are null # swaps...which would contradict the user provided `prob_swap`. # So special case num_replica==2, forcing u==False in this case. u_shape = ps.concat( (ps.ones(1, dtype=tf.int32), ps.cast(batch_shape, tf.int32)), axis=0) u = samplers.uniform(u_shape, seed=parity_seed) < 0.5 u = tf.where(num_replica > 2, u, False) x = bu.left_justified_expand_dims_to(ps.range(num_replica, dtype=tf.int64), rank=ps.size(u_shape)) y = tf.where(tf.equal(x % 2, tf.cast(u, dtype=tf.int64)), x + 1, x - 1) y = tf.clip_by_value(y, 0, num_replica - 1) # TODO(b/142689785): Consider using tf.cond and returning an empty list # then in REMC consider using a tf.cond for short-circuiting. return tf.where( samplers.uniform(batch_shape, seed=proposal_seed) < prob_swap, y, x)
def even_odd_swaps(num_replica, batch_shape=(), step_count=None, seed=None): """Make deterministic even_odd one time swaps.""" if step_count is None: raise ValueError('`step_count` must be supplied. Found `None`.') del seed # Unused for this function. with tf.name_scope(name or 'even_odd_swaps'): # Period is 1 / frequency, and we want period = Inf if frequency = 0. # safe_swap_period is the correct swap period in case swap_frequency > 0. # If swap_frequency == 0, safe_swap_period is set to 1 (to avoid integer # div by zero below). We will hard-set this case to "null swap." swap_freq = tf.convert_to_tensor(swap_frequency, name='swap_frequency') safe_swap_period = tf.cast( tf.where(swap_freq > 0, tf.math.ceil(tf.math.reciprocal_no_nan(swap_freq)), 1), # Although period = 1 / frequency may have roundoff error, and result # in a period different than what the user intended, the # user will end up with a single integer period, and thus well defined # deterministic swaps. tf.int32, ) # u selects parity. E.g., # u==False ==> [1, 0, 3, 2, 4] even parity swaps # u==True ==> [0, 2, 1, 4, 3] odd parity swaps # If there are 2 replicas, then the "True" swaps are null # swaps...which would contradict the user provided `swap_frequency`. # So special case num_replica==2, forcing u==False in this case. u_shape = ps.concat( (ps.ones(1, dtype=tf.int32), ps.cast(batch_shape, tf.int32)), axis=0) u = tf.fill(u_shape, tf.cast((step_count // safe_swap_period) % 2, tf.bool)) u = tf.where(num_replica > 2, u, False) x = bu.left_justified_expand_dims_to(tf.range(num_replica, dtype=tf.int64), rank=ps.size(u_shape)) y = tf.where(tf.equal(x % 2, tf.cast(u, dtype=tf.int64)), x + 1, x - 1) y = tf.clip_by_value(y, 0, num_replica - 1) # TODO(b/142689785): Consider using tf.cond and returning an empty list # then in REMC consider using a tf.cond for short-circuiting. return tf.where( (tf.cast(step_count % safe_swap_period, tf.bool) | tf.math.equal(swap_freq, 0)), x, # Don't swap y, # Swap )
def one_step(self, current_state, previous_kernel_results, seed=None): """Takes one step of the TransitionKernel. Args: current_state: `Tensor` or Python `list` of `Tensor`s representing the current state(s) of the Markov chain(s). previous_kernel_results: A (possibly nested) `tuple`, `namedtuple` or `list` of `Tensor`s representing internal calculations made within the previous call to this function (or as returned by `bootstrap_results`). seed: Optional, a seed for reproducible sampling. Returns: next_state: `Tensor` or Python `list` of `Tensor`s representing the next state(s) of the Markov chain(s). kernel_results: A (possibly nested) `tuple`, `namedtuple` or `list` of `Tensor`s representing internal calculations made within this function. This inculdes replica states. """ # The code below propagates one step states of shape # [n_replica] + batch_shape + event_shape. # # The step is done in three parts: # 1) Call one_step to transition states via a tempered version of # self.target_log_prob_fn (see _replica_target_log_prob). # 2) Permute values in states # 3) Update state-dependent values, such as log_probs. # # We chose to swap states, rather than temperatures, because... # (i) If swapping temperatures, you *still* have to swap log_probs to # determine acceptance, as well as states (for kernel results). # So it's just as difficult to swap temperatures. # (ii) If swapping temperatures, you have to take care to swap any user- # supplied temperature related things (like step size). # A-priori, we don't know what else will need to be swapped! # (iii)In both cases, the kernel results need to be updated in a non-trivial # manner....so we either special-case, or use bootstrap. with tf.name_scope(mcmc_util.make_name(self.name, 'remc', 'one_step')): # Force a read in case the `inverse_temperatures` is a `tf.Variable`. inverse_temperatures = tf.convert_to_tensor( previous_kernel_results.inverse_temperatures, name='inverse_temperatures') target_log_prob_for_inner_kernel = _make_replica_target_log_prob_fn( target_log_prob_fn=self.target_log_prob_fn, inverse_temperatures=inverse_temperatures, untempered_log_prob_fn=self.untempered_log_prob_fn, tempered_log_prob_fn=self.tempered_log_prob_fn, ) # TODO(b/159636942): Clean up the helpful error msg after 2020-11-10. try: inner_kernel = self.make_kernel_fn( # pylint: disable=not-callable target_log_prob_for_inner_kernel) except TypeError as e: if 'argument' not in str(e): raise raise TypeError( '`ReplicaExchangeMC`s `make_kernel_fn` no longer receives a `seed` ' 'argument. `TransitionKernel` instances now receive seeds via ' '`one_step`.') seed = samplers.sanitize_seed(seed) # Retain for diagnostics. inner_seed, swap_seed, logu_seed = samplers.split_seed(seed, n=3) # Step the inner TransitionKernel. [ pre_swap_replica_states, pre_swap_replica_results, ] = inner_kernel.one_step( previous_kernel_results.post_swap_replica_states, previous_kernel_results.post_swap_replica_results, seed=inner_seed) pre_swap_replica_target_log_prob = _get_field( # These are tempered log probs (have been divided by temperature). pre_swap_replica_results, 'target_log_prob') dtype = pre_swap_replica_target_log_prob.dtype replica_and_batch_shape = ps.shape( pre_swap_replica_target_log_prob) batch_shape = replica_and_batch_shape[1:] replica_and_batch_rank = ps.rank(pre_swap_replica_target_log_prob) num_replica = ps.size0(inverse_temperatures) inverse_temperatures = bu.left_justified_broadcast_to( inverse_temperatures, replica_and_batch_shape) # Now that each replica has done one_step, it is time to consider swaps. # swap.shape = [n_replica], and is a "once only" permutation, meaning it # is achievable by a sequence of pairwise permutations, where each element # is moved at most once. # E.g. if swaps = [1, 0, 2], we will consider swapping temperatures 0 and # 1, keeping 2 fixed. This exact same swap is considered for *every* # batch member. Of course some batch members may accept and some reject. try: swaps = tf.cast( self.swap_proposal_fn( # pylint: disable=not-callable num_replica, batch_shape=batch_shape, seed=swap_seed, step_count=previous_kernel_results.step_count), dtype=tf.int32) except TypeError as e: if 'step_count' not in str(e): raise warnings.warn( 'The `swap_proposal_fn` given to ReplicaExchangeMC did not accept ' 'the `step_count` argument. Falling back to omitting the ' 'argument. This fallback will be removed after 24-Oct-2020.' ) swaps = tf.cast( self.swap_proposal_fn( # pylint: disable=not-callable num_replica, batch_shape=batch_shape, seed=swap_seed), dtype=tf.int32) null_swaps = bu.left_justified_expand_dims_like( tf.range(num_replica, dtype=swaps.dtype), swaps) swaps = _maybe_embed_swaps_validation(swaps, null_swaps, self.validate_args) # Un-temper the log probs for use in the swap acceptance ratio. if self.tempered_log_prob_fn is None: # Efficient way of re-evaluating target_log_prob_fn on the # pre_swap_replica_states. untempered_negative_energy_ignoring_ulp = ( # Since untempered_log_prob_fn is None, we may assume # inverse_temperatures > 0 (else the target is improper). pre_swap_replica_target_log_prob / inverse_temperatures) else: # The untempered_log_prob_fn does not factor into the acceptance ratio. # Proof: Suppose the tempered target is # p_k(x) = f(x)^{beta_k} g(x), # So f(x) is tempered, and g(x) is not. Then, the acceptance ratio for # a 1 <--> 2 swap is... # (p_1(x_2) p_2(x_1)) / (p_1(x_1) p_2(x_2)) # which depends only on f(x), since terms involving g(x) cancel. untempered_negative_energy_ignoring_ulp = self.tempered_log_prob_fn( *pre_swap_replica_states) # Since `swaps` is its own inverse permutation we automatically know the # swap counterpart: range(num_replica). We use this idea to compute the # acceptance in a vectorized manner at the cost of wasting roughly half # our computation. Although we could use `unique` to solve this problem, # we expect the cost of `unique` to be higher than the dozens of wasted # arithmetic calculations. Worse, it'd mean we need dynamic sized Tensors # (eg, using `tf.where(bool)`) and so we wouldn't be able to XLA compile. # Note: diffs would normally be "proposed - current" however energy is # flipped since `energy == -log_prob`. # Note: The untempered_log_prob_fn (if provided) is not included in # untempered_pre_swap_replica_target_log_prob, and hence does not factor # into energy_diff. Why? Because, it cancels out in the acceptance ratio. energy_diff = (untempered_negative_energy_ignoring_ulp - mcmc_util.index_remapping_gather( untempered_negative_energy_ignoring_ulp, swaps, name='gather_swap_tlp')) swapped_inverse_temperatures = mcmc_util.index_remapping_gather( inverse_temperatures, swaps, name='gather_swap_temps') inverse_temp_diff = swapped_inverse_temperatures - inverse_temperatures # If i and j are swapping, log_accept_ratio[] i and j are equal. log_accept_ratio = (energy_diff * bu.left_justified_expand_dims_to( inverse_temp_diff, replica_and_batch_rank)) log_accept_ratio = tf.where(tf.math.is_finite(log_accept_ratio), log_accept_ratio, tf.constant(-np.inf, dtype=dtype)) # Produce log[Uniform] draws that are identical at swapped indices. log_uniform = tf.math.log( samplers.uniform(shape=replica_and_batch_shape, dtype=dtype, seed=logu_seed)) anchor_swaps = tf.minimum(swaps, null_swaps) log_uniform = mcmc_util.index_remapping_gather( log_uniform, anchor_swaps) is_swap_accepted_mask = tf.less(log_uniform, log_accept_ratio, name='is_swap_accepted_mask') def _swap_tensor(x): return mcmc_util.choose( is_swap_accepted_mask, mcmc_util.index_remapping_gather(x, swaps), x) post_swap_replica_states = [ _swap_tensor(s) for s in pre_swap_replica_states ] expanded_null_swaps = bu.left_justified_broadcast_to( null_swaps, replica_and_batch_shape) is_swap_proposed = _compute_swap_notmatrix( # Broadcast both so they have shape [num_replica] + batch_shape. # This (i) makes them have same shape as is_swap_accepted, and # (ii) keeps shape consistent if someday swaps has a batch shape. expanded_null_swaps, bu.left_justified_broadcast_to(swaps, replica_and_batch_shape)) # To get is_swap_accepted in ordered position, we use # _compute_swap_notmatrix on current and next replica positions. post_swap_replica_position = _swap_tensor(expanded_null_swaps) is_swap_accepted = _compute_swap_notmatrix( post_swap_replica_position, expanded_null_swaps) if self._state_includes_replicas: post_swap_states = post_swap_replica_states else: post_swap_states = [s[0] for s in post_swap_replica_states] post_swap_replica_results = _set_swapped_fields_to_nan( _swap_log_prob_and_maybe_grads(pre_swap_replica_results, post_swap_replica_states, inner_kernel)) if mcmc_util.is_list_like(current_state): # We *always* canonicalize the states in the kernel results. states = post_swap_states else: states = post_swap_states[0] post_swap_kernel_results = ReplicaExchangeMCKernelResults( post_swap_replica_states=post_swap_replica_states, pre_swap_replica_results=pre_swap_replica_results, post_swap_replica_results=post_swap_replica_results, is_swap_proposed=is_swap_proposed, is_swap_accepted=is_swap_accepted, is_swap_proposed_adjacent=_sub_diag(is_swap_proposed), is_swap_accepted_adjacent=_sub_diag(is_swap_accepted), # Store the original pkr.inverse_temperatures in case its a # `tf.Variable`. inverse_temperatures=previous_kernel_results. inverse_temperatures, swaps=swaps, step_count=previous_kernel_results.step_count + 1, seed=seed, potential_energy=-untempered_negative_energy_ignoring_ulp, ) return states, post_swap_kernel_results
def checkAllIntegrals(self, prior_scale, likelihood_scale, inverse_temperatures, iid_chain_ndims): prior_scale = tf.convert_to_tensor(prior_scale, name='prior_scale') likelihood_scale = tf.convert_to_tensor(likelihood_scale, name='likelihood_scale') # Create (normalized) prior and likelihood. Their product of course is not # normalized. In particular, there is a number `normalizing_const` such that # posterior(z) = prior.prob(x) * likelihood.prob(x) / normalizing_const # is normalized. prior = tfd.Normal(0., prior_scale) likelihood = tfd.Normal(0., likelihood_scale) posterior = tfd.Normal(0., (prior_scale**-2 + likelihood_scale**-2)**(-0.5)) # Get a good step size, custom for every replica/batch member. bcast_inv_temperatures = bu.left_justified_expand_dims_to( inverse_temperatures, # Broadcast over replicas. 1 + # Broadcast over chains. iid_chain_ndims + # Broadcast over batch dims. tf.rank(likelihood_scale)) tempered_posteriors = tfd.Normal( 0., # One tempered posterior for every inverse_temperature. (prior_scale**-2 + bcast_inv_temperatures * likelihood_scale**-2 )**(-0.5)) step_size = 0.71234 * tempered_posteriors.stddev() num_leapfrog_steps = tf.cast( tf.math.ceil(1.567 / tf.reduce_min(step_size)), tf.int32) def make_kernel_fn(target_log_prob_fn): return tfp.mcmc.HamiltonianMonteCarlo( target_log_prob_fn=target_log_prob_fn, step_size=step_size, num_leapfrog_steps=num_leapfrog_steps, ) remc = tfp.mcmc.ReplicaExchangeMC( target_log_prob_fn=None, untempered_log_prob_fn=prior.log_prob, tempered_log_prob_fn=likelihood.log_prob, inverse_temperatures=inverse_temperatures, state_includes_replicas=False, make_kernel_fn=make_kernel_fn, swap_proposal_fn=tfp.mcmc.even_odd_swap_proposal_fn(1.), ) def trace_fn(state, results): # pylint: disable=unused-argument return { 'replica_log_accept_ratio': results.post_swap_replica_results.log_accept_ratio, 'is_swap_accepted_adjacent': results.is_swap_accepted_adjacent, 'is_swap_proposed_adjacent': results.is_swap_proposed_adjacent, 'potential_energy': results.potential_energy, } if tf.executing_eagerly(): num_results = 100 else: num_results = 1000 num_burnin_steps = num_results // 10 n_samples_per_chain = 2 initial_sample_shape = [n_samples_per_chain] * iid_chain_ndims unused_replica_states, trace = self.evaluate( tfp.mcmc.sample_chain( num_results=num_results, # Start at one of the modes, in order to make mode jumping necessary # if we want to pass test. current_state=prior.sample(initial_sample_shape, seed=test_util.test_seed()), kernel=remc, num_burnin_steps=num_burnin_steps, trace_fn=trace_fn, seed=test_util.test_seed())) # Tolerance depends on samples * replicas * number of (iid) chains. # ess.shape = [n_replica, ...] # We will sum over batch dims, then take min over replica. ess = tfp.mcmc.effective_sample_size(trace['potential_energy']) if iid_chain_ndims: ess = tf.reduce_sum(ess, axis=tf.range(1, 1 + iid_chain_ndims)) min_ess = self.evaluate(tf.reduce_min(ess)) n_combined_results = min_ess * inverse_temperatures.shape[0] # Make sure sampling worked well enough, for every replica/chain. conditional_swap_prob = ( np.sum(trace['is_swap_accepted_adjacent'], axis=0) / np.sum(trace['is_swap_proposed_adjacent'], axis=0)) self.assertAllGreater(conditional_swap_prob, 0.5) replica_mean_accept_prob = np.mean(np.exp( np.minimum(0, trace['replica_log_accept_ratio'])), axis=0) self.assertAllGreater(replica_mean_accept_prob, 0.5) integrals = self.evaluate( tfp.experimental.mcmc.remc_thermodynamic_integrals( inverse_temperatures, trace['potential_energy'], iid_chain_ndims=iid_chain_ndims, )) self.assertAllEqual(posterior.batch_shape, integrals.log_normalizing_constant_ratio.shape) actual_log_normalizing_const = self.evaluate( # Use arbitrary point, 0, to find the constant. prior.log_prob(0.) + likelihood.log_prob(0.) - posterior.log_prob(0.)) self.assertAllClose(integrals.log_normalizing_constant_ratio, actual_log_normalizing_const, rtol=10 / np.sqrt(n_combined_results)) self.assertAllEqual(posterior.batch_shape, integrals.cross_entropy_difference.shape) def cross_entropy(dist): z = dist.sample(50000, seed=test_util.test_seed()) return tf.reduce_mean(likelihood.log_prob(z), axis=0) iid_cross_entropy_difference = self.evaluate( cross_entropy(posterior) - cross_entropy(prior)) self.assertAllClose(integrals.cross_entropy_difference, iid_cross_entropy_difference, rtol=30 / np.sqrt(n_combined_results))