Skip to content

Minnt

Minnt is Milan's Minimalistic Neural Network Trainer for PyTorch inspired by Keras.

See the Examples. Install using:

pip install minnt --extra-index-url=https://download.pytorch.org/whl/cu128 torch torchvision

The central class in Minnt is the TrainableModule providing:

  • high-level training, evaluation, and prediction API, including automatic device management;
  • serialization and deserialization of weights (optionally including the optimizer) and configuration;
  • automatic logging via various loggers;
  • easy to use multi-GPU single-node training.

The TransformedDataset class allows applying both per-example and per-batch transformation functions on a given dataset, and simplifies the creation of a corresponding dataloader (in a multi-GPU setting if required).

Furthermore, the package contains a collection of losses and metrics; however, losses from PyTorch and metrics from torchmetrics can also be used directly.

Finally, several utilities are provided.

The Minnt version 

__version__ = '0.0.4-alpha'