I built MaRN: a PyTorch library for training neural networks through low-dimensional parameter mappings [P]
retained text
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:** Reduced a **537,748-parameter CNN to 4,080 trainable parameters** (**131.8× reduction**), with accuracy decreasing from **99.07% → 98.10%** (**−0.97 pp**). A smaller **107,998-parameter CNN** was reduced to **1,872 trainable parameters** (**57.7× reduction**), with accuracy decreasing from **98.83% → 97.18%** (**−1.65 pp**). 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](https://github.com/arjunmnath/MaRN) **Docs:** [https://marn.readthedocs.io/](https://marn.readthedocs.io/) I'd value feedback on the approach, benchmark design, and where this kind of parameter-efficient optimization could be useful.
document fetched 2026-10-10 00:28 · canonical https://reddit.com/r/MachineLearning/comments/1x1fjrv/i_built_marn_a_pytorch_library_for_training