cortenforge-training 0.6.0

Burn-based training and eval for LinearClassifier/MultiboxModel models in the CortenForge stack.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

//! Burn-based training and evaluation for CortenForge detection models.
//!
//! This crate provides:
//! - Dataset loading and collation (`collate`, `collate_from_burn_batch`).
//! - Training loop utilities (`run_train`, `TrainArgs`).
//! - Model checkpoint loading/saving helpers.
//!
//! Supports both `LinearClassifier` and `MultiboxModel` from the `models` crate.
//!
//! ## Backend Selection
//! - `backend-wgpu`: Uses WGPU for GPU-accelerated training.
//! - Default: Falls back to NdArray CPU backend.
//!
//! ## Stability
//!
//! Training APIs are **experimental** and may change as the training pipeline evolves.
//! Core model types (`TinyDet`, `BigDet`) are stable, but training utilities and loss functions
//! are subject to refinement.

#![recursion_limit = "256"]

pub mod dataset;
pub mod util;

pub use dataset::{collate, collate_from_burn_batch, CollatedBatch, DatasetPathConfig, RunSample};
pub use models::{LinearClassifier, LinearClassifierConfig, MultiboxModel, MultiboxModelConfig};
pub use util::{run_train, TrainArgs};
/// Backend alias for training/eval (NdArray by default; WGPU if enabled).
#[cfg(feature = "backend-wgpu")]
pub type TrainBackend = burn_wgpu::Wgpu<f32>;
#[cfg(not(feature = "backend-wgpu"))]
pub type TrainBackend = burn_ndarray::NdArray<f32>;