Member El::load (ArchiveT &archive, ::El::Matrix< T, D > &mat)

Perhaps it's better to throw an exception for these archives?

Namespace lbann

Remove this file.

Rename this file to file.hpp

Member lbann::adam< TensorDataType >::get_current_beta1 () const noexcept

This probably shouldn't be exposed.

Member lbann::adam< TensorDataType >::get_current_beta2 () const noexcept

This probably shouldn't be exposed.

Member lbann::adam< TensorDataType >::set_current_beta1 (TensorDataType current_beta1)

This probably shouldn't be exposed.

Member lbann::adam< TensorDataType >::set_current_beta2 (TensorDataType current_beta2)

This probably shouldn't be exposed.

Member lbann::Al::nccl_backend

MPI-CUDA backend

Member lbann::base_convolution_layer< TensorDataType, Device >::base_convolution_layer (int num_data_dims, int output_channels, std::vector< int > conv_dims, std::vector< int > pads, std::vector< int > strides, std::vector< int > dilations, int groups, bool has_bias)

Remove num_data_dims from arg list

Member lbann::callback::checkpoint::get_distributed_checkpoint_rootdir ()

BVE FIMME this looks wrong I think that the order should be reversed

Class lbann::callback::set_weights_value

Support weights with arbitrary data types. Currently only floats are supported.

Member lbann::callback::summary::summary (const std::shared_ptr< lbann_summary > &summarizer, int batch_interval=1, int mat_interval=25)

Document mat_interval parameter.

Member lbann::callback_base::m_batch_interval

Make callback data private

Member lbann::callback_type

Fix the callback types to properly track execution phases

Class lbann::categorical_random_layer< TensorDataType, T_layout, Dev >

Remove.

Class lbann::channelwise_mean_layer< TensorDataType, Layout, Device >

Replace with more general reduction layer.

Member lbann::check_if_dir_exists (const std::string &dirname)

Deprecated. Use lbann::file::directory_exists instead.

Member lbann::check_if_file_exists (const std::string &filename)

Deprecated. Use lbann::file::file_exists instead.

Member lbann::closeread (int fd, const char *filename)

Deprecated.

Member lbann::closewrite (int fd, const char *filename)

Deprecated.

Member lbann::create_dir (const std::string output_dir)

Deprecated. Use lbann::file::make_directory instead.

Member lbann::crop_layer< TensorDataType, T_layout, Dev >::setup_data (size_t max_mini_batch_size) override

Setup the input tensor with this data distribution

Member lbann::cross_entropy_layer< TensorDataType, T_layout, Dev >::fp_compute () override

Consider reduce rather than allreduce

Member lbann::cufft::cuFFTWrapper< InputTypeT >::setup_forward (InputMatType &in, std::vector< int > const &full_dims)

Assert this is ok for R2C cases!!!

Member lbann::data_coordinator::epoch_complete (execution_mode mode)=0

BVE FIXME this should probably be a property of the execution mode

Member lbann::data_type_optimizer< TensorDataType >::m_learning_rate

Consider moving this to the derived classes.

Class lbann::discrete_random_layer< TensorDataType, T_layout, Dev >

Remove.

Member lbann::EGrid

Remove

Member lbann::ExecutionContext::get_type () const =0

Absorb completely into get_state_string().

Member lbann::exists (const char *filename)

Deprecated.

Member lbann::fftw::FFTWWrapper< InputTypeT >::setup_forward (InputMatType &in, std::vector< int > const &full_dims)

Assert this is ok for R2C cases!!!

Member lbann::file::join_path (PathNameType &&... paths)

Assert every

Member lbann::fully_connected_layer< TensorDataType, T_layout, Dev >::fully_connected_layer (int output_size, bool transpose=false, WeightsType *weight=nullptr, bool has_bias=true)

Accept a vector for output_size

Class lbann::gather_layer< TensorDataType, Layout, Device >

Support higher-dimensional data

Member lbann::gather_layer< TensorDataType, Layout, Device >::setup_dims () override

Support gathering from/into higher-order tensors

Member lbann::generic_data_reader::get_num_iterations_per_epoch () const

BVE FIXME merge this with alternate approach

Member lbann::generic_data_reader::set_num_iterations_per_epoch (int num_iterations_per_epoch)

BVE FIXME merge this with alternate approach

Class lbann::gru_layer< TensorDataType, Layout, Device >

Support bidirectional RNNs

Member lbann::hdf5_data_reader::m_useme_node_map_ptrs

- cleanup node map pointers

Member lbann::input_layer< TensorDataType, T_layout, Dev >::input_layer (lbann_comm *comm, std::string const data_field="")

make the map and vector references

Member lbann::kfac::KFACExecutionContext::get_type () const override

Absorb completely into get_state_string().

Member lbann::KFAC::on_forward_prop_end (ExeContextType &context, model &model)

Break up into more manageable pieces

Member lbann::l1_norm_layer< TensorDataType, T_layout, Dev >::fp_compute () override

Consider reduce rather than allreduce

Member lbann::l2_norm2_layer< TensorDataType, T_layout, Dev >::fp_compute () override

Consider reduce rather than allreduce

Member lbann::l2_weight_regularization::differentiate () override

Come up with a better function name in the base class.

Member lbann::Layer::m_grid_tag

tym: Clean up and document

Member lbann::Layer::subgraph_communication_method

Remove

Member lbann::ltfb::LTFBExecutionContext::get_type () const override

Absorb completely into get_state_string().

Member lbann::ltfb::RandomPairwiseExchange::ExchangeStrategy::get_partner_model (model const &m, El::Int partner_trainer, size_t step)=0

The step parameter is only used by CheckpointFile; we should consider alternatives that don't clutter the API.

Class lbann::ltfb::SendRecvWeights

More general approach to exchange optimizer state. Currently only SGD and Adam are supported.

Member lbann::makedir (const char *dirname)

Deprecated.

Class lbann::matmul_layer< TensorDataType, Layout, Device >

Support >3 dimensions, matvecs, and dot products

Member lbann::mean_absolute_error_layer< TensorDataType, T_layout, Dev >::fp_compute () override

Consider reduce rather than allreduce

Member lbann::mean_squared_error_layer< TensorDataType, T_layout, Dev >::fp_compute () override

Consider reduce rather than allreduce

Member lbann::openread (const char *filename)

Deprecated.

Member lbann::openwrite (const char *filename)

Deprecated.

Member lbann::optimizer::gradient_manager_type

Refactor this out. It's a hack.

Member lbann::OwningLayerPtr

Replace with std::unique_ptr<Layer> when C++ and Cereal support std::observer_ptr .

Member lbann::OwningWeightsPtr

Replace with std::unique_ptr<weights> when C++ and Cereal support std::observer_ptr .

Member lbann::parse_path (const std::string &path, std::string &dir, std::string &basename)

Deprecated. Use lbann::file::extract_parent_directory and lbann::file::extract_base_name instead.

Member lbann::protobuf::assign_to_repeated (google::protobuf::RepeatedField< T > &field, ContainerT const &values)

Change to Assign if Protobuf v3.16.0 and greater is required.

Class lbann::reduction_layer< TensorDataType, Layout, Device >

Reduction over specified dimensions.

Member lbann::rowwise_weights_norms_layer< TensorDataType, Layout, Device >::fp_compute () override

Synchronize

Class lbann::scatter_layer< TensorDataType, Layout, Device >

Support higher-dimensional data

Class lbann::sgd< TensorDataType >

Dedicated optimizers for momentum or Nesterov SGD.

Member lbann::sort_layer< TensorDataType, T_layout, Dev >::m_indices

Switch to distributed integer matrix once it's supported in Hydrogen.

Member lbann::tessellate_layer< TensorDataType, Layout, Device >::bp_compute () override

Support >3 dimensions

Member lbann::tessellate_layer< TensorDataType, Layout, Device >::fp_compute () override

Support >3 dimensions

Member lbann::tessellate_layer< TensorDataType, Layout, Device >::setup_dims () override

Support tessellation with >3 dimensions

Member lbann::to_string (El::Device const &d)

This should move to hydrogen

Class lbann::top_k_categorical_accuracy_layer< TensorDataType, T_layout, Dev >

Gracefully handle case where label is not a one-hot vector.

Member lbann::trainer::trainer (lbann_comm *comm, std::unique_ptr< data_coordinator > dc, size_t mini_batch_size, std::unique_ptr< TrainingAlgorithm > alg=nullptr)

I don't know why mini_batch_size is here.

Class lbann::TrainingAlgorithm

One component that we need to address yet is the issue of logically encapsulating multiple models, as either inputs or outputs to a training algorithm. Specifically, consider the LTFB "meta-learning" method. Rather than producing the single best model, a user might be interested in the K best models. In this case, tournament-based evolution will begin with a single model (per trainer) but could output several models. Similarly, one might begin with an arbitrary collection of models that are evolved until a single best model emerges. This draws in other issues to be addressed elsewhere in LBANN such as "How do we export models?" Currently, this is done by writing to files on disk via callbacks. However, one might imagine "in-core" interaction between training and inference, perhaps in an online learning scenario, in which repeatedly writing to and reading from disk is not sufficient.

Class lbann::unpooling_layer< TensorDataType, T_layout, Dev >

GPU support.

Member lbann::utils::details::do_tensor_copy_between_grids (const BaseDistMat &src, El::AbstractDistMatrix< TDT > &tgt)

Do this more systematically and support all matrix classes

Member lbann::utils::from_string (std::string const &str)

chars, shorts, unsigned. Bool needs some work.

Member lbann::ViewingLayerPtr

Replace with std::observer_ptr<Layer> when supported by C++ and Cereal.

Member lbann::ViewingWeightsPtr

Replace with std::observer_ptr<Weights> when supported by C++ and Cereal.

Member lbann::weights::set_values (El::BaseDistMatrix const &values)

(trb 05/28/2020): Should this check the DistData of the input against the expected DistData for the weights object?