I built MaRN: a PyTorch library for training neural networks through low-dimensional parameter mappings [P]

I built MaRN (Mapping Networks), a PyTorch library that lets you optimize a compact latent representation instead of directly training every model parameter.

Some results from my current benchmarks:

• MNIST CNN: 107,998 → 1,872 trainable parameters (57.7× reduction), with 91.80% accuracy.
• LSTM forecasting: 12,051 → 2,048 parameters, with validation MSE of 0.00006.
• CNN2 + pruning: 204 trainable parameters, with 81.25% accuracy.

These results come with trade-offs: mapped models can train substantially slower, and performance varies by task. The benchmarks are exploratory, using synthetic data for some tasks, and aren't evidence of general superiority over direct training.

The library includes global and layer-wise mappings, regularization options, and pruning/LRD integrations.

Code: https://github.com/arjunmnath/MaRN
Docs: https://marn.readthedocs.io/

I'd value feedback on the approach, benchmark design, and where this kind of parameter-efficient optimization could be useful.