Expand description
Neural network modules for ToRSh
This crate provides PyTorch-compatible neural network layers and modules, built on top of scirs2-neural for optimized implementations.
§Modular Architecture
The neural network core system is organized into specialized modules for improved maintainability:
- core: Core Module trait system and essential interfaces
- parameter: Comprehensive parameter management and initialization
- hooks: Hook system infrastructure for module callbacks
- base: ModuleBase helper for module implementations
- composition: Module composition patterns (sequential, parallel, etc.)
- construction: Module construction and configuration patterns
- diagnostics: Module and parameter diagnostics and health checking
- utils: Module utilities and helper functions
All components maintain full backward compatibility through comprehensive re-exports.
Re-exports§
pub use core::Module;pub use parameter::LayerType;pub use parameter::Parameter;pub use parameter::ParameterCollection;pub use parameter::ParameterDiagnostics;pub use parameter::ParameterStats;pub use hooks::HookCallback;pub use hooks::HookHandle;pub use hooks::HookRegistry;pub use hooks::HookType;pub use base::ModuleBase;pub use composition::ComposedModule;pub use composition::ConditionalModule;pub use composition::ModuleBuilder;pub use composition::ModuleComposition;pub use composition::ParallelModule;pub use composition::ResidualModule;pub use construction::ModuleConfig;pub use construction::ModuleConstruct;pub use diagnostics::ModuleDiagnostics;pub use diagnostics::ModuleInfo;pub use utils::ModuleApply;pub use utils::ModuleParameterStats;
Modules§
- base
- Module base infrastructure for neural network implementations
- compile_
time - Compile-time optimization features for zero-cost abstractions
- composition
- Module composition system for neural network modules
- construction
- Module construction and configuration system
- container
- Container modules for organizing neural network layers and parameters
- conversion
- Model conversion utilities
- core
- Core module trait system for neural network modules
- cuda_
kernels - Custom CUDA kernels integration for torsh-nn via scirs2
- diagnostics
- Module diagnostics and analysis system
- export
- Model export functionality
- functional
- Functional interface for neural network operations Enhanced with SciRS2-Neural integration for optimized performance
- gradcheck
- Gradient checking utilities for neural network layers
- hardware_
opts - Hardware-Specific Layer Optimizations
- hooks
- Hook system infrastructure for neural network modules
- init
- Parameter initialization functions
- layers
- Neural network layer modules
- lazy
- Lazy initialization utilities for neural network modules
- mixed_
precision - Mixed Precision Training Support
- model_
zoo - Model zoo with pre-built architectures and pretrained weights
- modules
- Common neural network modules
- numerical_
stability - Numerical Stability Testing and Validation
- optimization
- Performance optimization utilities for neural networks
- parameter
- Parameter management system for neural network modules
- parameter_
updates - Optimized parameter update strategies for better training performance
- prelude
- Prelude module for convenient imports
- pruning
- Neural network pruning utilities
- quantization
- Quantization support for model compression and deployment optimization
- research
- Research neural network layers and components
- scirs2_
neural_ integration - Comprehensive scirs2-neural integration for advanced neural network capabilities
- sparse
- Sparse Neural Network Support
- summary
- Model summary utilities for analyzing neural network architectures
- utils
- Module utilities and helper functions
- visualization
- Network architecture visualization tools
Macros§
- func_
error - Macro for consistent error wrapping
- impl_
module_ constructor - Macro to implement standardized constructors
- lazy_
module - Macro for creating lazy modules with factory functions
- validate_
inputs - Macro for consistent function validation patterns
Structs§
- Sparse
Matrix - Sparse Matrix placeholder for compatibility