Char-RNN Text Generation

Built RNN and LSTM cells from scratch in PyTorch — explicit input/hidden/bias weight matrices and gate computations (input, forget, output with tanh/sigmoid) — instead of relying on the built-in nn.RNN / nn.LSTM modules.

Assembled character-level language models on top of those cells and trained them for autoregressive text generation.

Engineered the sequence-construction data pipeline (windowed character sequences) and a sampling-based generation loop.

Contrasted the basic RNN against the LSTM to demonstrate how gating mitigates vanishing gradients on longer dependencies.