Rejig ELBO so that baselines don't have to output scaled costs.

This means that the baseline objective doesn't change with the batch
size, and it also makes our baseline net step sizes more likely to be
comparable to those used outside of Pyro.

pyro/infer/tracegraph_elbo.py

@@ -235,7 +235,7 @@ def _loss_and_grads_particle(self, weight, model_trace, guide_trace):
             for node in non_reparam_nodes:
                 guide_site = guide_trace.nodes[node]
                 log_pdf_key = 'batch_log_pdf' if node in guide_vec_md_nodes else 'log_pdf'
-                downstream_cost = downstream_costs[node]
+                downstream_cost = downstream_costs[node] / guide_site["scale"]  # not scaled by subsampling
                 baseline = 0.0
                 (nn_baseline, nn_baseline_input, use_decaying_avg_baseline, baseline_beta,
                     baseline_value) = _get_baseline_options(guide_site)
@@ -260,7 +260,7 @@ def _loss_and_grads_particle(self, weight, model_trace, guide_trace):
                     # accumulate baseline loss
                     baseline_loss += torch.pow(downstream_cost.detach() - baseline, 2.0).sum()
 
-                guide_log_pdf = guide_site[log_pdf_key] / guide_site["scale"]  # not scaled by subsampling
+                guide_log_pdf = guide_site[log_pdf_key]
                 if use_nn_baseline or use_decaying_avg_baseline or use_baseline_value:
                     if downstream_cost.size() != baseline.size():
                         raise ValueError("Expected baseline at site {} to be {} instead got {}".format(