New Neural Net Framework (github.com)

Sequential is a new, lightweight neural-net framework purpose-built for multi-step time-series forecasting. Implemented in pure Python and NumPy, it supports feed‑forward, RNN, LSTM and decoder‑only Transformer models and handles both many‑to‑one and many‑to‑many prediction windows (any target length up to the input window). The project emphasizes explora­tibility and teachability: limited, well-documented modules make it easy to audit, extend, and understand how layers and training behave, positioning Sequential as a simpler, domain-focused alternative to large, general-purpose stacks like Keras/TensorFlow/PyTorch. On the technical side, Sequential requires Python >= 3.12 and NumPy >= 2.0.0 and installs from the repo (with optional CLI/Jupyter and dev extras). It automates preprocessing, fitting, validation and forecasting via train.py and forecast.py (JSON config-driven), saves fitted models and CSV/plot outputs, and includes notebook templates for recurrent and Transformer workflows. Notable implementation details: recurrent layers always return full sequences (avoiding return_sequences fiddles), built-in time‑dimension compression for shorter outputs, autoregressive masking for Transformer decoder training, and utilities for reproducible runs (np.random.seed) plus pytest test coverage. Benchmarks and a development blog series are forthcoming.