perforatedai
Perforated AI: Better accuracy, smaller models, less data - enabled by perforated learning.
Perforated AI adds neuron-specific learning signals (artificial dendrites) during training, helping models achieve higher accuracy with fewer parameters, less data, and lower deployment costs. It integrates directly into existing PyTorch workflows with minimal code changes.
Modules
perforatedai.globals_perforatedai— Configuration classes and utilities (PAIConfig): device settings, dendrite management, module conversion options, and training parameters.perforatedai.utils_perforatedai— Entry point for converting a model (perforate_model) plus helpers for saving, loading, and inspecting PAI networks.perforatedai.modules_perforatedai— The core module wrappers that add dendritic copies to layers and manage dendrite state during training.perforatedai.network_perforatedai— Network-level conversion and checkpoint loading for PAI models.perforatedai.tracker_perforatedai— Training tracker: validation-score history, learning-rate management, and deciding when to add dendrites.perforatedai.library_perforatedai— Processors and reference architectures (LSTM processors, ResNet variants) for modules that need custom input/output handling.perforatedai.blockwise_perforatedai— Blockwise optimization of converted networks.perforatedai.clean_perforatedai— Utilities for exporting a trained PAI model to a cleaned-up, scaffold-free form.
See the README for installation, examples, and key results.
1"""Perforated AI: Better accuracy, smaller models, less data - enabled by perforated learning. 2 3Perforated AI adds neuron-specific learning signals (artificial dendrites) 4during training, helping models achieve higher accuracy with fewer 5parameters, less data, and lower deployment costs. It integrates directly 6into existing PyTorch workflows with minimal code changes. 7 8Modules 9------- 10- `perforatedai.globals_perforatedai` — Configuration classes and utilities 11 (`PAIConfig`): device settings, dendrite management, module conversion 12 options, and training parameters. 13- `perforatedai.utils_perforatedai` — Entry point for converting a model 14 (`perforate_model`) plus helpers for saving, loading, and inspecting 15 PAI networks. 16- `perforatedai.modules_perforatedai` — The core module wrappers that add 17 dendritic copies to layers and manage dendrite state during training. 18- `perforatedai.network_perforatedai` — Network-level conversion and 19 checkpoint loading for PAI models. 20- `perforatedai.tracker_perforatedai` — Training tracker: validation-score 21 history, learning-rate management, and deciding when to add dendrites. 22- `perforatedai.library_perforatedai` — Processors and reference 23 architectures (LSTM processors, ResNet variants) for modules that need 24 custom input/output handling. 25- `perforatedai.blockwise_perforatedai` — Blockwise optimization of 26 converted networks. 27- `perforatedai.clean_perforatedai` — Utilities for exporting a trained 28 PAI model to a cleaned-up, scaffold-free form. 29 30See the [README](https://github.com/PerforatedAI/PerforatedAI) for 31installation, examples, and key results. 32"""