Expand description
gras — neural architecture search over random topologies, run continuous step-race.
One global step clock: every live net trains and evaluates on the same
shared batch each step. Divergence over the population’s smoothed fitness
culls the worst net and births a caught-up replacement. Blueprints
(Topology) are compiled into executable Networks; you supply
loss and fitness as plain closures.
Quick start:
let config = RaceConfig::builder()
.set_pop_size(6)
.set_stop_max_steps(100)
.build();Re-exports§
pub use evolution::crossover;pub use evolution::mutation;pub use evolution::pools;pub use evolution::selection;pub use graph::network;pub use graph::node;pub use graph::topology;pub use utils::fitness;pub use utils::markdown;pub use utils::tabular_data;pub use engine::RLSpec;pub use engine::RaceConfig;pub use engine::RaceSnapshot;pub use engine::RlEngine;pub use engine::RlRaceConfig;pub use engine::RunMode;pub use engine::RunSpec;pub use engine::StopReason;pub use engine::TabularEngine;pub use engine::TabularRaceConfig;pub use engine::TabularSpec;pub use utils::fitness::Direction;pub use utils::fitness::Fitness;pub use utils::fitness::FitnessLabel;pub use graph::network::Network;pub use graph::network::NetworkOptions;pub use graph::node::Activation;pub use graph::node::CombineOp;pub use graph::node::Node;pub use graph::node::NodeKind;pub use graph::node::StandardizeOp;pub use graph::topology::FeatureSplitPolicy;pub use graph::topology::Topology;pub use graph::topology::TopologyOptions;pub use evolution::crossover::CrossoverMethod;pub use evolution::mutation::MutationMethod;pub use evolution::selection::SelectionMethod;pub use trainer::stream::BatchStream;pub use trainer::stream::PoolSplit;pub use trainer::AnyTrainer;pub use trainer::Env;pub use trainer::LossFn;pub use trainer::RlContext;pub use trainer::RlStep;pub use trainer::RlStepMeta;pub use trainer::RunData;pub use trainer::StepEnv;pub use trainer::StepReport;pub use trainer::StreamShape;pub use trainer::TabularContext;pub use trainer::TabularStep;pub use trainer::TabularTrainer;pub use trainer::menu::FITNESS_NAMES;pub use trainer::menu::Loss;pub use trainer::menu::METRIC_NAMES;pub use trainer::menu::TargetKind;pub use trainer::menu::fitness_for;pub use trainer::menu::metric_for;pub use trainer::menu::metrics_for;pub use utils::tabular_data::DataFormat;pub use utils::tabular_data::Dataset;pub use utils::tabular_data::PoolLayout;pub use utils::tabular_data::detect_pool_layout;pub use utils::tabular_data::exam_dir;pub use utils::tabular_data::load_csv_dataset;pub use utils::tabular_data::load_dataset;pub use utils::tabular_data::load_dataset_auto;pub use utils::tabular_data::load_tensor;pub use utils::tabular_data::one_hot;pub use utils::tabular_data::resolve_dataset;pub use utils::tabular_data::resolve_inputs_targets_datasets;pub use utils::tabular_data::resolve_train_test_datasets;pub use utils::tabular_data::resolve_train_val_datasets;pub use utils::tabular_data::save_csv_dataset;pub use utils::tabular_data::save_dataset;pub use utils::tabular_data::save_dataset_as;pub use utils::tabular_data::save_tensor;pub use utils::score::accuracy_score;pub use utils::score::argmax_classes;pub use utils::score::cross_entropy_onehot;pub use utils::score::cross_entropy_onehot_loss;pub use utils::score::f1_from_vecs;pub use utils::score::f1_score;pub use utils::score::l1_loss_score;pub use utils::score::label_smoothing_cross_entropy_loss;pub use utils::score::mse_loss_score;pub use utils::score::precision_from_vecs;pub use utils::score::precision_score;pub use utils::score::r2_score;pub use utils::score::rmse_score;pub use utils::race_steps::deterministic_train_step;pub use utils::race_steps::deterministic_train_step_fn;pub use utils::race_steps::eval_one_step;pub use utils::race_steps::eval_one_step_fn;pub use utils::race_steps::seed_step_randomness;pub use utils::race_steps::train_one_step;pub use utils::race_steps::train_one_step_fn;pub use utils::race_steps::train_one_step_pred_only;pub use flodl;
Modules§
- engine
- The engine — the step-race NAS loop.
- evolution
- Evolution operators — the per-generation genetic operators.
- graph
- The computational graph — nodes, topologies (blueprints), and networks.
- prelude
- One import for the common surface:
use gras::prelude::*;. The common surface: engine entry points, fitness/direction types, trainer contracts + reports, the batch stream, graph types, config enums, and the tensor/device re-exports. - spec
- JSON serialization for the graph blueprint and network diagnostics.
- state
- State serialization and persistence for the race engine.
- trainer
- Training schemes for the step-race engine.
- utils
- Utilities — data I/O, graph helpers, scoring, rendering, and the step-race training primitives.
Structs§
- Variable
- A differentiable variable wrapping a Tensor.
Enums§
- DType
- Element data type of a tensor. Maps to PyTorch’s
torch.dtype. - Device
- Device represents where a tensor’s data lives.
Functions§
- auto_
device - CUDA device 0 with the
cudafeature, otherwise CPU.