Method backward_prop

src/models/model.cpp:1584–1664

}

void model::backward_prop(bool compute_weight_grads_only)
{
  LBANN_CALIPER_MARK_FUNCTION;

  // Layers disabled due to not propagating error signals through
  std::unordered_set<const Layer*> disabled_layers;
  auto const& arg_parser = global_argument_parser();
  bool const envvar_disable_layers =
    !arg_parser.get<bool>(LBANN_OPTION_NO_BACKPROP_DISABLE);

  do_model_backward_prop_begin_cbs();

  for (El::Int i = get_num_layers() - 1; i >= 0; --i) {

    // Perform backward prop step on current layer
    auto& l = get_layer(i);
    bool enable_layer = (!envvar_disable_layers ||
                         disabled_layers.find(&l) == disabled_layers.end());

    // Check if all children skip gradient backpropagation
    if (enable_layer && envvar_disable_layers) {
      bool all_children_skip_gradient = l.get_num_children() > 0;
      for (auto& child : l.get_child_layers()) {
        if (disabled_layers.find(child) != disabled_layers.end())
          continue;
        all_children_skip_gradient &=
          (child->get_backprop_requirements() == PROPAGATE_NOTHING);
        if (!all_children_skip_gradient)
          break;
      }

      // Start disabling layers from this point onwards
      if (all_children_skip_gradient) {
        enable_layer = false;
        disabled_layers.insert(&l);
      }
    }

    if (this->is_subgraph_parallelism_enabled()) {
      if (l.get_run_layer_in_subgraph()) {
        do_layer_backward_prop_begin_cbs(&l);
        if (enable_layer)
          l.back_prop();
        do_layer_backward_prop_end_cbs(&l);
      }
      else {
        // To Do: Fix last batch problem in sub-graph parallelism
        // experimental code to fix last batch problem in subgraph parallelism
      }
    }
    else {
      do_layer_backward_prop_begin_cbs(&l);
      if (enable_layer)
        l.back_prop();
      do_layer_backward_prop_end_cbs(&l);
    }

    // Terminate early if all gradients have been computed