ML_Music / Assignment 2

CSE153 music generation assignment with two sub-tasks:

• Task 1: Unconditioned symbolic generation — LSTM models for pitch and duration trained on classical piano MIDI.
• Task 2: Conditioned symbolic generation — builds on Task 1 with arpeggio-focused filtering and constrained decoding.

Data Setup

1. Place classical piano MIDI files under dataset/ (notebook reads dataset/**/*.mid).
2. Adjust preprocessing constants in the notebook if needed (e.g., MIN_SEQ_LEN, MAX_SEQ_LEN, TRANSPOSE_RANGE).

Task 1: Unconditioned Workflow

1. Preprocessing (Task1 section in notebook)
   • Parse MIDI, quantize to 1/16 notes, drop short sequences, split at length 128, augment with ±3/±6 semitone transposition.
   • Build pitch_to_id / duration_to_id; save to Task1_preprocessed/.
2. EDA
   • Run EDA cells for pitch-class distribution, duration distribution, entropy, length histograms, and summary stats.
3. Training & Generation
   • Train separate pitch/duration LSTMs (embed 128, hidden 256, 2 layers, dropout 0.3, label smoothing 0.1).
   • Best models (by val perplexity) saved to Task1_outputs/; generates generated_music.mid.
4. Baselines & Metrics
   • Trigram baseline vs. LSTM: pitch perplexity ~10 vs. trigram ~498, showing stronger sequence modeling.

Task 2: Arpeggio-Conditioned Workflow

1. Preprocessing
   • Keep only arpeggio-like sequences: durations in {1,2,4}, intervals ≤24 semitones, min length 32; apply the same transposition augmentation.
   • Save processed data and vocab to Task2_preprocessed/.
2. Training & Generation
   • Train pitch/duration LSTMs (same hyperparams as Task 1) with constrained decoding:
     • nucleus/top-k sampling (top_p=0.9, top_k=50), n-gram repeat blocking, repeat-threshold control;
     • progressive relaxation of duration set and top-p if sampling fails.
   • Outputs Task2_outputs/generated_music.mid.
3. Baselines & Metrics
   • Against n-gram baselines: typical test perplexity ~9.6 (pitch) and ~1.2 (duration), outperforming unstructured baselines.

Tunable Knobs

• Faster runs: lower NUM_EPOCHS, HIDDEN_SIZE, batch size, or subsample MIDI files.
• More diversity: raise TEMPERATURE, TOP_P, or TOP_K; if quality drops, tighten n-gram/repeat constraints.
• Longer pieces: raise MAX_GENERATION_LENGTH, noting stability limits on very long sequences.

References & Data

• Classical piano MIDI dataset (Kaggle, sourced from piano-midi.de, 19 composers).
• Related work: Melissa Jalali Monfared’s LSTM example; Sulun et al. (2022) transformer with continuous emotion conditioning.