suyashnpande/cinematic-music-descriptor-v2-module3

Cinematic Music Descriptor — Module 3 – Music Descriptor Heads

Multi-head MLP that takes Module2 context vectors and predicts all music descriptors for a scene: tempo, tonality, orchestration, etc.

Label Schema

Regression

• tempo_bpm: 45–170 BPM
• musical_valence: -0.93 to 0.68

Classification

• tonality: ['atonal', 'major', 'minor']
• harmonic_style: ['atonal', 'chromatic', 'cluster', 'diatonic', 'modal', 'pentatonic', 'whole_tone']
• dynamic_shape_m4: ['crescendo', 'diminuendo', 'flat', 'subito_forte', 'subito_piano', 'sustained', 'swell', 'terraced']
• rhythm_style: ['drive', 'off', 'ostinato', 'pulse', 'rubato', 'sparse']
• texture: ['ambient', 'chamber', 'full', 'hybrid', 'solo']

Multi-label

• orchestration: ['ambient_pad', 'brass', 'choir', 'electronic', 'ethnic', 'guitar', 'harp', 'organ', 'percussion', 'piano', 'solo_voice', 'strings', 'synth', 'woodwinds']

Training Details

• Base model: roberta-base
• Dataset: ~11,000 scenes from 60–80 movies
• Framework: PyTorch + HuggingFace Transformers
• Logging: Weights & Biases

Usage

import torch
from huggingface_hub import hf_hub_download
import config as C

# Download weights  (repo ID is built from HF_REPO_ID in config.py)
# e.g. "suyashnpande/cinematic-music-descriptor-v2-module3"
repo_id = f"{C.HF_REPO_ID}-module3"
path = hf_hub_download(repo_id=repo_id,
                       filename="module3_final.pt",
                       token=C.HF_READ_TOKEN)

Citation

If you use this model, please cite the project.