Module optimization 

Model optimization techniques for efficient inference

This module provides various model optimization techniques to reduce model size, improve inference speed, and reduce memory consumption while maintaining accuracy.

Techniques

• Quantization: Reduce precision (FP32 -> INT8/FP16)
• Pruning: Remove unnecessary weights and connections
• Knowledge Distillation: Transfer knowledge from large to small models
• Model Compression: GZIP, Huffman coding, weight sharing

Example

use oxigdal_ml::optimization::quantize_model;
use oxigdal_ml::optimization::{QuantizationConfig, QuantizationType};

let config = QuantizationConfig::builder()
    .quantization_type(QuantizationType::Int8)
    .per_channel(true)
    .build();

quantize_model("model.onnx", "model_quantized.onnx", &config)?;

Re-exports

pub use distillation::DenseLayer;

pub use distillation::DistillationConfig;

pub use distillation::DistillationConfigBuilder;

pub use distillation::DistillationLoss;

pub use distillation::DistillationStats;

pub use distillation::DistillationTrainer;

pub use distillation::EarlyStopping;

pub use distillation::ForwardCache;

pub use distillation::LearningRateSchedule;

pub use distillation::MLPGradients;

pub use distillation::OptimizerType;

pub use distillation::SimpleMLP;

pub use distillation::SimpleRng;

pub use distillation::Temperature;

pub use distillation::TrainingState;

pub use distillation::cross_entropy_loss;

pub use distillation::cross_entropy_with_label;

pub use distillation::kl_divergence;

pub use distillation::kl_divergence_from_logits;

pub use distillation::log_softmax;

pub use distillation::mse_loss;

pub use distillation::soft_targets;

pub use distillation::softmax;

pub use distillation::train_student_model;

pub use pruning::FineTuneCallback;

pub use pruning::GradientInfo;

pub use pruning::ImportanceMethod;

pub use pruning::LotteryTicketState;

pub use pruning::MaskCreationMode;

pub use pruning::NoOpFineTune;

pub use pruning::PruningConfig;

pub use pruning::PruningConfigBuilder;

pub use pruning::PruningGranularity;

pub use pruning::PruningMask;

pub use pruning::PruningSchedule;

pub use pruning::PruningStats;

pub use pruning::PruningStrategy;

pub use pruning::UnstructuredPruner;

pub use pruning::WeightStatistics;

pub use pruning::WeightTensor;

pub use pruning::compute_channel_importance;

pub use pruning::compute_gradient_importance;

pub use pruning::compute_magnitude_importance;

pub use pruning::compute_taylor_importance;

pub use pruning::iterative_pruning;

pub use pruning::prune_model;

pub use pruning::prune_weights_direct;

pub use pruning::prune_weights_with_gradients;

pub use pruning::select_weights_to_prune;

pub use pruning::structured_pruning;

pub use pruning::unstructured_pruning;

pub use quantization::QuantizationConfig;

pub use quantization::QuantizationMode;

pub use quantization::QuantizationParams;

pub use quantization::QuantizationResult;

pub use quantization::QuantizationType;

pub use quantization::calibrate_quantization;

pub use quantization::dequantize_tensor;

pub use quantization::quantize_model;

pub use quantization::quantize_tensor;

Modules

distillation

Knowledge distillation for model compression

pruning

Model pruning for sparse neural networks

quantization

Model quantization for reduced precision inference

Structs

OptimizationPipeline

Combined optimization pipeline

OptimizationStats

Model optimization statistics

Enums

OptimizationProfile

Model optimization profile