• Home
• Introduction
• Changelog
• Migration
• Benchmarks
• Contribute
• Usage
• Notebooks

Changelog

Conventions

PiNN follows the PEP 440 scheme of versioning. Specifically, the versions are tagged as major.minor.micro. micro updates with same minor version are expected to be backward-compactible, i.e., the models trained on old micro version can be used newer ones. minor updates can break backward-compatibility, if this happens it should be documented in this change log.

v2.x.y

v2.1.0

• Backend:
  • TensorFlow window moved from 2.6–2.9 to 2.15 (Python 3.9–3.11, NumPy 1.x). This is the last TF release that still ships tf.estimator and the legacy Keras optimizers (both removed in 2.16), and the first in-window release that can use NVIDIA Hopper / GH200 (sm_90). See migration.
  • Existing 2.x parameter files and trained models remain usable; YAML optimizer names (Adam, SGD, EKF, …) are unchanged.
• Compatibility (TF 2.15):
  • Optimizers now request the legacy Keras optimizer automatically so the estimator graph-mode tf.gradients + apply_gradients loop still works.
  • ASE calculator disables tf.data prefetch/autotune; otherwise calculate() could return the previous step's energy/forces.
  • Default ASE raised to ≥3.25.0 in setup.py / environment.yml (Bussi NVT thermostat; 3.22.0 does not provide it). Containers install PiNN from those files and do not re-pin ASE.
  • I/O: TFRecord spec from public dataset.element_spec; ANI-1 loader uses h5py 3 (dataset[()]); removed np.int / np.float aliases.
• Containers:
  • CPU image: tensorflow/tensorflow:2.15.0 (runtime CLI; no Jupyter).
  • GPU image: NGC nvcr.io/nvidia/tensorflow:24.03-tf2-py3 (TF 2.15 + CUDA 12.4, multi-arch amd64/arm64 — the only aarch64 GPU TF 2.15).
  • In-repo Singularity / Singularity.gpu defs for clusters without Docker.
• CI / packaging:
  • Test matrix is Python 3.9–3.11 × TensorFlow 2.15.
  • Docker CPU image push and docs deploy run only from Teoroo-CMC/PiNN on master / version tags; forks still build the CPU image. GPU images are built from Dockerfile.gpu / Singularity.gpu outside GitHub Actions.
  • setup.py extras: .[cpu] / .[gpu] pin tensorflow>=2.15,<2.16.
• Workflow:
  • Nextflow arrhenius profile (GPU train / CPU data-prep via Apptainer).
  • figshare dataset downloads go through the API (article-zip returned HTTP 202).

v2.0.0

• New Network:
  • PiNet2: new equivariant neural network;
• New Models:
  • PiNet-dipole: dipole moment prediction model;
  • PiNet-\(\chi\): machine learning charge response kernel model;
• New workflow:
  • nextflow: training pipeline is now managed by Nextflow
• New tools:
  • pinn report: extract results from work directory or model folder

v1.x.y

v1.1.0

• Refactor layers (breaks compatibility with saved models <v1.1.0).

v1.0.0

• Refactor to TensorFlow 2.

v0.x.y

v0.3.0

• Initial public release