Finding the genre of a song with Deep Learning

[Chip (OP)]

No attachments on this post.

LISTATT TRACE — currentMsgId(): 49592, 0 rows in attachments table

Finding the genre of a song with Deep Learning — A.I. Odyssey part. 1.pdf

You already have PyTorch working on Live and Clone, and ffmpeg for audio extraction. Spectrogram generation (librosa or plain scipy FFT) is CPU-light and would run fine on either box. Training a small CNN like this (2,000 songs, 128×128 slices) is a much lighter workload than what you've already put PyTorch through building/running CLIP and llama.cpp.

The real gap isn't infrastructure, it's that everything you've built for ISEmedia so far is inference-only (pretrained CLIP embeddings, pretrained SmolVLM captions) — you've never run a training loop. This article is a full training pipeline: labeled dataset → spectrogram slices → train/val/test split → CNN from scratch → checkpoint → inference. That's a different category of work than anything currently in ISEmedia, and you noted standalone audio files are currently skipped from AI processing entirely under the assumption they're just music.