Skip to content

README.md

Latest commit

 

History

History
120 lines (96 loc) · 4.55 KB

README.md

File metadata and controls

120 lines (96 loc) · 4.55 KB
Raw

DeepLearning-Project

One-line summary

A compact research codebase for training and evaluating sequential models (RNNs) on lyric/text generation tasks. The repository contains training and generation scripts, model implementations, preprocessing utilities, and example datasets to reproduce baseline experiments.

Table of contents

  • Highlights
  • Architecture (components & responsibilities)
  • Pipeline (data → model → output)
  • Quickstart
  • Repository layout
  • Reproducibility & experiments
  • Recommendations & next steps
  • Contact

Highlights

  • Clear, modular baseline implementations of RNN architectures suitable for quick experimentation and comparison.
  • Separate training and generation workflows to keep experiments reproducible and composable.
  • Utilities for text and MIDI feature handling (integration-ready for multimodal experiments).

Architecture

The following table summarizes the main components, their location in the repo and their responsibilities.

Component Location Responsibility
Data sets data/sets/ Raw CSV datasets for training and testing (lyrics).
Preprocessing utils utils/text_utils.py Tokenization, vocabulary creation, batching helpers.
Feature extraction utils/midi_features.py Extract and normalize MIDI-derived features (if used).
Model implementations models/ RNN_baseline.py, RNN_baseline_V1.py, RNN_baseline_V2.py — core architectures and forward logic.
Training loop train.py Loads data, trains model, handles checkpoints and logging.
Generation / sampling generate.py Loads a trained checkpoint and generates text from a prompt/seed.

Pipeline (high-level)

This project follows a linear pipeline designed for experiment reproducibility and easy swapping of components.

  1. Data collection: add or edit CSVs under data/sets/.
  2. Preprocessing: utils/text_utils.py tokenizes text, builds or loads vocab, and produces batches.
  3. Model definition: pick or modify an RNN in models/.
  4. Training: train.py runs epochs, logs metrics, and saves checkpoints.
  5. Generation: generate.py loads a checkpoint and produces output samples.
  6. Evaluation: compare outputs against test set and qualitative inspection.

ASCII flow diagram

Data (CSV) --> Preprocessing --> Batches --> Trainer --> Checkpoints | v Generation

Key pipeline notes

  • Checkpointing: train.py should save model + optimizer state so generation can resume from any saved state.
  • Configs: keep hyperparameters in a single place (recommended: config.yaml or CLI args) to reproduce experiments.
  • Determinism: log random seeds and library versions for exact reproduction.

Quickstart (Windows example)

  1. Create and activate a virtual environment:
python -m venv .venv
.\.venv\Scripts\activate
  1. (Optional) Install dependencies (create requirements.txt if absent):
pip install torch pandas numpy
  1. Train a baseline model (defaults in train.py):
python train.py
  1. Generate samples from a saved checkpoint:
python generate.py --checkpoint path/to/checkpoint.pt --seed "start prompt"

Repository layout (concise)

Path Purpose
train.py Training loop and CLI for experiments
generate.py Generation / sampling script
models/ RNN implementations and model definitions
data/sets/ Example datasets (lyrics_train_set.csv, lyrics_test_set.csv)
utils/ Preprocessing and feature utilities (text_utils.py, midi_features.py)

Reproducibility & experiments

  • Save: model checkpoints, training logs (loss/metrics), and a copy of the config used for each run.
  • Document: random seeds, Python and library versions, and the exact dataset files used.
  • Evaluate: include both quantitative metrics and qualitative samples in any report.

Recommendations & next steps

  • Add a requirements.txt or pyproject.toml to lock dependencies.
  • Add a config.yaml or expand CLI arg parsing to centralize hyperparameters and dataset paths.
  • Add automated example notebooks or a scripts/ folder with reproducible experiment runners.
  • Consider adding unit tests for preprocessing utilities to ensure stable pipelines.

Contact

For questions or collaboration, open an issue in this repository or contact the maintainer directly via the repository profile.

If you want, I can:

  • produce a concise English one-paragraph pitch for sharing on LinkedIn/GitHub profile,
  • create a short CV-friendly bullet summary,
  • or add requirements.txt and a minimal config.yaml to the repo now.