gmumcl 1.0.0__tar.gz

gmumcl-1.0.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024-2026 GMUM
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

gmumcl-1.0.0/MANIFEST.in

@@ -0,0 +1,7 @@
+include LICENSE README.md pyproject.toml
+graft configs
+graft docs
+graft examples
+graft scripts
+graft tests
+global-exclude __pycache__ *.pyc .DS_Store

gmumcl-1.0.0/PKG-INFO

@@ -0,0 +1,141 @@
+Metadata-Version: 2.4
+Name: gmumcl
+Version: 1.0.0
+Summary: A generic continual-learning framework: scenarios, methods, metrics
+License: MIT License
+        
+        Copyright (c) 2024-2026 GMUM
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+        
+Project-URL: Repository, https://github.com/gmum/gmumcl
+Keywords: continual-learning,pytorch,class-incremental,benchmark
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch<2.9,>=2.0
+Requires-Dist: numpy
+Requires-Dist: PyYAML
+Requires-Dist: activationscope>=0.1.4
+Provides-Extra: datasets
+Requires-Dist: torchvision; extra == "datasets"
+Provides-Extra: viz
+Requires-Dist: matplotlib; extra == "viz"
+Provides-Extra: test
+Requires-Dist: pytest>=7; extra == "test"
+Requires-Dist: torchvision; extra == "test"
+Requires-Dist: ruff; extra == "test"
+Dynamic: license-file
+
+# gmumcl
+
+[![CI](https://github.com/gmum/gmumcl/actions/workflows/ci.yml/badge.svg)](https://github.com/gmum/gmumcl/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/gmumcl)](https://pypi.org/project/gmumcl/)
+[![Python](https://img.shields.io/pypi/pyversions/gmumcl)](https://pypi.org/project/gmumcl/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
+A small, generic continual-learning framework for PyTorch: five incremental
+scenarios, a five-call plugin API, an accuracy-matrix trainer, and reference
+implementations of the standard CL baselines.
+
+## Install
+
+`activationscope` builds a C++ extension with `-std=c++17`, so install torch 2.8
+first (newer torch headers require C++20), then build it against that torch:
+
+```bash
+pip install "torch==2.8.*" --index-url https://download.pytorch.org/whl/cpu   # or /cu128
+pip install setuptools wheel ninja numpy
+pip install activationscope --no-build-isolation
+pip install gmumcl              # or: pip install "gmumcl[datasets]"
+```
+
+From source: `git clone https://github.com/gmum/gmumcl && cd gmumcl && pip install -e ".[test]"`.
+Full notes (why the pin, why `--no-build-isolation`): `docs/usage.md`.
+
+## Quickstart
+
+```bash
+python scripts/run.py --config configs/base.yaml -O method.name=ewc -O method.lambda=10
+```
+
+```python
+from gmumcl import train
+
+metrics = train.run({
+    "data": {"name": "toy", "num_classes": 4, "classes_per_task": 2},
+    "model": {"name": "mlp", "hidden": 16},
+    "trainer": {"epochs": 1, "lr": 0.01, "device": "cpu"},
+    "method": {"name": "ewc", "lambda": 1.0},
+})
+print(metrics["accuracy_matrix"])
+```
+
+## Scenarios (`data.scenario`)
+
+| Key | Tasks | Evaluation |
+|---|---|---|
+| `class_incremental` | disjoint class groups | all classes seen so far |
+| `task_incremental` | disjoint class groups | task oracle (logits masked to the task) |
+| `domain_incremental` | all classes, shifted inputs (rotations) | all classes |
+| `instance_incremental` | random partitions of all instances | all classes |
+| `permuted` | all classes, one fixed input permutation per task | all classes |
+
+Dataset names: `toy` (CPU clusters), `mnist`, `cifar10`, `cifar100`, plus
+`rotated_mnist` (domain-incremental) and `permuted_mnist`.
+
+## Methods (`method.name`)
+
+| Name | Reference | Source |
+|---|---|---|
+| `naive` | — | plain fine-tuning |
+| `l2` | Hsu et al. 2018 | quadratic anchor |
+| `ewc` | Kirkpatrick et al. 2017 | gmumcl / InTAct-Flex |
+| `ewc_dr` | EWC-DR | `scarlet0703/EWC-DR` (adcaaeb9) |
+| `online_ewc` | Schwarz et al. 2018 | gmumcl / InTAct-Flex |
+| `si` | Zenke et al. 2017 | gmumcl |
+| `mas` | Aljundi et al. 2018 | gmumcl |
+| `lwf` | Li & Hoiem 2017 | gmumcl / `GMvandeVen/continual-learning` |
+| `kfac` | Kao et al. 2021 | `Xin-Jing/K-FAC` (09f4565c) |
+| `ncl` | Kao et al. 2021 | `tachukao/ncl` (92c2ba5e) |
+| `owm` | Zeng et al. 2019 | `GMvandeVen/continual-learning` (e6d795aa) |
+| `gpm` | Saha et al. 2021 | `sahagobinda/GPM` (1a238ec9) |
+| `intact` | Krukowski et al. 2025 | `pkrukowski1/InTAct` (83b90c55) |
+| `sharpening` | gmumcl | `OneAndZero24/local-cl` |
+| `dreaming` | gmumcl | `OneAndZero24/local-cl` (needs `model.head_type: rbf`) |
+
+## Docs
+
+- `docs/framework.md` — plugin lifecycle, trainer, metrics, adding a method
+- `docs/scenarios.md` — the five scenarios and their protocols
+- `docs/methods.md` — formulas, defaults and provenance per method
+- `docs/usage.md` — install, CLI, configs, custom components
+- `docs/results.md` — reproduced Split-MNIST table (from AdaReg) and deviations
+
+## License
+
+MIT (GMUM). See `LICENSE`.

gmumcl-1.0.0/README.md

@@ -0,0 +1,88 @@
+# gmumcl
+
+[![CI](https://github.com/gmum/gmumcl/actions/workflows/ci.yml/badge.svg)](https://github.com/gmum/gmumcl/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/gmumcl)](https://pypi.org/project/gmumcl/)
+[![Python](https://img.shields.io/pypi/pyversions/gmumcl)](https://pypi.org/project/gmumcl/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
+A small, generic continual-learning framework for PyTorch: five incremental
+scenarios, a five-call plugin API, an accuracy-matrix trainer, and reference
+implementations of the standard CL baselines.
+
+## Install
+
+`activationscope` builds a C++ extension with `-std=c++17`, so install torch 2.8
+first (newer torch headers require C++20), then build it against that torch:
+
+```bash
+pip install "torch==2.8.*" --index-url https://download.pytorch.org/whl/cpu   # or /cu128
+pip install setuptools wheel ninja numpy
+pip install activationscope --no-build-isolation
+pip install gmumcl              # or: pip install "gmumcl[datasets]"
+```
+
+From source: `git clone https://github.com/gmum/gmumcl && cd gmumcl && pip install -e ".[test]"`.
+Full notes (why the pin, why `--no-build-isolation`): `docs/usage.md`.
+
+## Quickstart
+
+```bash
+python scripts/run.py --config configs/base.yaml -O method.name=ewc -O method.lambda=10
+```
+
+```python
+from gmumcl import train
+
+metrics = train.run({
+    "data": {"name": "toy", "num_classes": 4, "classes_per_task": 2},
+    "model": {"name": "mlp", "hidden": 16},
+    "trainer": {"epochs": 1, "lr": 0.01, "device": "cpu"},
+    "method": {"name": "ewc", "lambda": 1.0},
+})
+print(metrics["accuracy_matrix"])
+```
+
+## Scenarios (`data.scenario`)
+
+| Key | Tasks | Evaluation |
+|---|---|---|
+| `class_incremental` | disjoint class groups | all classes seen so far |
+| `task_incremental` | disjoint class groups | task oracle (logits masked to the task) |
+| `domain_incremental` | all classes, shifted inputs (rotations) | all classes |
+| `instance_incremental` | random partitions of all instances | all classes |
+| `permuted` | all classes, one fixed input permutation per task | all classes |
+
+Dataset names: `toy` (CPU clusters), `mnist`, `cifar10`, `cifar100`, plus
+`rotated_mnist` (domain-incremental) and `permuted_mnist`.
+
+## Methods (`method.name`)
+
+| Name | Reference | Source |
+|---|---|---|
+| `naive` | — | plain fine-tuning |
+| `l2` | Hsu et al. 2018 | quadratic anchor |
+| `ewc` | Kirkpatrick et al. 2017 | gmumcl / InTAct-Flex |
+| `ewc_dr` | EWC-DR | `scarlet0703/EWC-DR` (adcaaeb9) |
+| `online_ewc` | Schwarz et al. 2018 | gmumcl / InTAct-Flex |
+| `si` | Zenke et al. 2017 | gmumcl |
+| `mas` | Aljundi et al. 2018 | gmumcl |
+| `lwf` | Li & Hoiem 2017 | gmumcl / `GMvandeVen/continual-learning` |
+| `kfac` | Kao et al. 2021 | `Xin-Jing/K-FAC` (09f4565c) |
+| `ncl` | Kao et al. 2021 | `tachukao/ncl` (92c2ba5e) |
+| `owm` | Zeng et al. 2019 | `GMvandeVen/continual-learning` (e6d795aa) |
+| `gpm` | Saha et al. 2021 | `sahagobinda/GPM` (1a238ec9) |
+| `intact` | Krukowski et al. 2025 | `pkrukowski1/InTAct` (83b90c55) |
+| `sharpening` | gmumcl | `OneAndZero24/local-cl` |
+| `dreaming` | gmumcl | `OneAndZero24/local-cl` (needs `model.head_type: rbf`) |
+
+## Docs
+
+- `docs/framework.md` — plugin lifecycle, trainer, metrics, adding a method
+- `docs/scenarios.md` — the five scenarios and their protocols
+- `docs/methods.md` — formulas, defaults and provenance per method
+- `docs/usage.md` — install, CLI, configs, custom components
+- `docs/results.md` — reproduced Split-MNIST table (from AdaReg) and deviations
+
+## License
+
+MIT (GMUM). See `LICENSE`.

gmumcl-1.0.0/configs/base.yaml

@@ -0,0 +1,32 @@
+# CPU quickstart: tiny Gaussian clusters, class-incremental, EWC.
+# Run: python scripts/run.py --config configs/base.yaml
+seed: 0
+
+data:
+  name: toy              # toy | mnist | cifar10 | cifar100 | permuted_mnist | rotated_mnist
+  scenario: class_incremental   # class_incremental | task_incremental | domain_incremental | instance_incremental | permuted
+  dimension: 8
+  num_classes: 6
+  classes_per_task: 2
+  samples_per_class: 150
+  num_tasks: 5           # used by instance_incremental / domain_incremental / permuted
+  calibration_fraction: 0.1
+  min_calibration: 50
+  batch_size: 32
+
+model:
+  name: mlp              # mlp | small_cnn | resnet18 | vit
+  hidden: 64
+  layers: 2
+  head_type: linear      # linear | rbf (rbf is required by Dreaming)
+  growing_head: true     # default: true for class/task-incremental, false otherwise
+
+trainer:
+  epochs: 1
+  lr: 0.01
+  device: cpu
+  verbose: false         # per-task/epoch progress and timing prints
+
+method:
+  name: ewc              # naive | l2 | ewc | ewc_dr | online_ewc | si | mas | lwf |
+  lambda: 0.1            # kfac | ncl | owm | gpm | intact | sharpening | dreaming

gmumcl-1.0.0/configs/baselines/dreaming.yaml

@@ -0,0 +1,35 @@
+# Dreaming on Split MNIST class-incremental.
+# The Hsu-adapted protocol is reused so the row is comparable; the model uses
+# the RBF head the method replays, and the generation budget follows the
+# original gmumcl defaults. See docs/results.md for the protocol, provenance
+# and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/dreaming.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+  head_type: rbf
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: dreaming
+  alpha: 1.0                  # feature-distillation weight
+  samples_per_class: 32
+  steps: 100                  # per-class input optimisation steps
+  in_shape: [1, 28, 28]

gmumcl-1.0.0/configs/baselines/ewc.yaml

@@ -0,0 +1,30 @@
+# EWC, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/ewc.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: ewc
+  mode: ewc
+  lambda: 600.0               # Hsu et al. grid-search optimum

gmumcl-1.0.0/configs/baselines/ewc_dr.yaml

@@ -0,0 +1,32 @@
+# EWC-DR, port of `scarlet0703/EWC-DR` (commit adcaaeb9) on Split MNIST.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# The importance is the reversed-logit gradient squared, averaged over the
+# task loader and clamped at `omegamax`, merged with alpha = known/total.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/ewc_dr.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: ewc_dr
+  lambda: 1.0
+  omegamax: 0.0001            # official clamp

gmumcl-1.0.0/configs/baselines/gpm.yaml

@@ -0,0 +1,33 @@
+# GPM, port of `sahagobinda/GPM` (commit 1a238ec9) on Split MNIST
+# class-incremental. The official thresholds (0.97 + 0.003 * task) and row
+# cap are kept; the training protocol follows the Hsu-adapted setup used by
+# the other rows so the numbers are comparable.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/gpm.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: gpm
+  threshold_base: 0.97        # official energy thresholds
+  threshold_step: 0.003
+  max_rows: 125

gmumcl-1.0.0/configs/baselines/intact.yaml

@@ -0,0 +1,32 @@
+# InTAct, port of `pkrukowski1/InTAct` (commit 83b90c55) on Split MNIST
+# class-incremental. InTAct paper protocol: 2x400 MLP, batch 512, 5 epochs,
+# Adam lr=2e-4, and the published class-IL coefficients.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/intact.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 512
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 5
+  lr: 0.0002
+  device: cuda
+
+method:
+  name: intact
+  var_scale: 4.2              # published class-IL coefficients
+  lambda_int_drift: 6.0
+  lambda_feat: 779.0

gmumcl-1.0.0/configs/baselines/kfac.yaml

@@ -0,0 +1,31 @@
+# K-FAC, NCL-paper baseline (Kao et al., 2021) on Split MNIST class-incremental.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# `lambda` is not tuned here (the NCL paper tunes it per experiment).
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/kfac.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: kfac
+  lambda: 1.0
+  kfac_samples: 256           # output-gradient samples per task

gmumcl-1.0.0/configs/baselines/l2.yaml

@@ -0,0 +1,29 @@
+# L2, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/l2.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: l2
+  lambda: 100.0               # Hsu et al. grid-search optimum

gmumcl-1.0.0/configs/baselines/lwf.yaml

@@ -0,0 +1,30 @@
+# LwF, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/lwf.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: lwf
+  lambda: 1.0                 # distillation weight (Li and Hoiem default)
+  temperature: 2.0

gmumcl-1.0.0/configs/baselines/mas.yaml

@@ -0,0 +1,29 @@
+# MAS, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/mas.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: mas
+  lambda: 1.0                 # Hsu et al. grid-search optimum

gmumcl-1.0.0/configs/baselines/naive.yaml

@@ -0,0 +1,28 @@
+# Naive fine-tuning, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/naive.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: naive

gmumcl-1.0.0/configs/baselines/ncl.yaml

@@ -0,0 +1,36 @@
+# NCL (no opt), NCL-paper baseline (Kao et al., NeurIPS 2021) on Split MNIST
+# class-incremental. Protocol (paper Appendix E): batch 256, 2000 iterations
+# per task, SGD with momentum 0.9 and lr 0.05, single 10-way output layer.
+# Unit Gaussian prior (p_w^-2 = 1), projection stabiliser 1e-10, lambda 1.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/ncl.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 256
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: false
+
+trainer:
+  optimizer: sgd
+  momentum: 0.9
+  lr: 0.05
+  iterations: 2000
+  device: cuda
+
+method:
+  name: ncl
+  data_size: 1.0              # unit prior: NCL (no opt)
+  projection_alpha: 0.0000000001
+  lambda: 1.0
+  kfac_samples: 256

gmumcl-1.0.0/configs/baselines/online_ewc.yaml

@@ -0,0 +1,30 @@
+# Online EWC, Hsu et al. (2018) Split MNIST class-incremental baseline.
+# Protocol: 4 epochs/task, batch 128, Adam lr=1e-3, all training data.
+# See docs/results.md for the full protocol and documented deviations.
+# Run: python scripts/run.py --config configs/baselines/online_ewc.yaml
+seed: 0
+
+data:
+  name: mnist
+  root: data
+  classes_per_task: 2
+  calibration_fraction: 0.0
+  min_calibration: 0
+  batch_size: 128
+
+model:
+  name: mlp
+  input_dim: 784
+  hidden: 400
+  layers: 2
+  growing_head: true
+
+trainer:
+  epochs: 4
+  lr: 0.001
+  device: cuda
+
+method:
+  name: online_ewc
+  gamma: 1.0                  # Hsu et al. accumulate without decay
+  lambda: 100.0               # Hsu et al. grid-search optimum