Skip to content

CHANGELOG.md

Latest commit

 

History

History
68 lines (50 loc) · 2.63 KB

CHANGELOG.md

File metadata and controls

68 lines (50 loc) · 2.63 KB

Changelog

All notable changes to this project are documented here.

The format follows Keep a Changelog. This project does not use semantic versioning — changes are tracked by milestone.

[Unreleased]

Planned

  • FSM control unit for MAC orchestration
  • SRAM/BRAM on-chip memory module
  • Systolic array (2D MAC grid)
  • AXI4 / AXI-Stream bus wrapper

[2026-03-16] — Multi-dataset, model factory, training hardening

Added

  • Multi-dataset support: CIFAR-10, Tiny-ImageNet (auto-download), ImageNet (manual)
  • Model factory get_model() dispatching SimpleCNN / ResNet-18 / ResNet-50 per dataset
  • Dataset-aware hyperparameter presets (DATASET_PRESETS in config.py)
  • Configurable optimizer: Adam and AdamW with weight decay
  • Hardware package restructuring: units/, bus/, testbenches/
  • cleaner.sh for artifact cleanup with --force flag
  • data_purge.py for GPU VRAM and system RAM release
  • initializer.sh multi-package-manager support (uv / poetry / conda / pip)
  • --dataset and --manager flags in initializer.sh
  • Community files: CODE_OF_CONDUCT.md, CONTRIBUTING.md, SECURITY.md
  • GitHub PR template and issue templates

Changed

  • SimpleCNN scaled from 16/32 to 32/64 filters (+5% val acc on CIFAR-10)
  • Trainer class refactored with Metrics dataclass (from Jérémy Alcime's PR #4)
  • Checkpoint now stores optimizer_state_dict, scheduler_state_dict, epoch, best_acc
  • Checkpoint filename is now dataset-dependent (cifar10.pth, tiny_imagenet.pth)
  • All print() replaced by logging module in train.py, dataset.py, testbench.py
  • train.py now uses apply_dataset_preset() for automatic hyperparameter coordination

Fixed

  • Testbench always uses CIFAR-10 for pixel extraction regardless of active dataset
  • avg_loss() now derives window from metrics.total instead of batch_idx
  • Duplicate metrics.loss accumulation removed from train_one_epoch()
  • Missing f-strings in logger calls corrected
  • AMP autocast and GradScaler wired correctly in training loop

[2026-03-14] — Initial release

Added

  • SimpleCNN CNN architecture (PyTorch) with hardware-aware design
  • CIFAR-10 training pipeline with CosineAnnealingLR and model checkpointing
  • MACUnit cycle-accurate MAC hardware unit in Amaranth HDL
  • Verilog RTL export from Amaranth (mac.v)
  • HW/SW co-simulation testbench with int8 quantization and assertion
  • VCD waveform generation for GTKWave inspection
  • Pydantic-based unified configuration (src/config.py)
  • uv project management with pyproject.toml
  • README.md with full setup, usage, and GTKWave guide