mortm 4.0b5

音楽の旋律生成を実現したシステム

🎵 MORTM: Metric-Oriented Rhythmic Transformer for Music Generation

MORTM (Metric-Oriented Rhythmic Transformer for Music Generation) is a Transformer-based melody generation model that focuses on the metric structure of music. It generates musical sequences autoregressively, one bar at a time, while preserving rhythmic consistency. MORTM also includes V_MORTM for audio-based generation and BERTM for music classification tasks.

✨ Features

• Bar-level Autoregressive Generation: Normalizes each bar to 64 ticks and generates one bar at a time.

• High-Quality Music Generation: Uses a custom tokenizer to capture musical structure (pitch, duration, relative timing, bars), leading to coherent outputs.

• Efficient Transformer Architecture:

  • Decoder-Only (GPT-style): Optimized for autoregressive generation.
  • FlashAttention2 & ALiBi: Memory-efficient, high-speed attention with excellent long-sequence generalization.
  • Mixture of Experts (MoE): Sparsely activated FFN layers for higher capacity.

• Structured Tokenization: Tokens for Pitch, Duration, Position and structural tokens like <SME>, <TS>, <TE>. Normalizes to 96 ticks per bar.

• Multimodal (V_MORTM): Processes audio features (Mel spectrograms) directly.

• Classification (BERTM): A BERT-like encoder for music classification tasks.

• Versatile Applications: Melody generation, improvisation assistance, education, human-AI co-creation, audio style transfer.

🚀 Why MORTM?

• State-of-the-Art: Combines FlashAttention2, MoE, and ALiBi.
• Musical Understanding: Custom tokenizer captures core musical elements.
• Scalability: Supports diverse styles and long sequences.
• Audio Domain: V_MORTM for richer audio-based generation.
• Modular: Easy prototyping and comparative experiments.

🛠️ Installation

Prerequisites

• Python 3.8+
• NVIDIA GPU (for FlashAttention2)
• CUDA Toolkit (compatible with PyTorch)

1. Install PyTorch

Follow instructions at pytorch.org. Example:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

2. Install FlashAttention2

pip install flash-attn --no-build-isolation

3. Install Other Dependencies

pip install numpy einops pretty_midi midi2audio soundfile torchaudio PyYAML

• midi2audio requires FluidSynth and a soundfont (e.g., .sf2).

4. Optional: Gmail Notifications

pip install google-api-python-client google-auth-httplib2 google-auth-oauthlib

Requires OAuth2 setup (client_secret.json).

⚡ Quick Start

Data Preparation

Convert MIDI to tokenized .npz:

from mortm.train.tokenizer import Tokenizer, get_token_converter, TO_TOKEN
from mortm.convert import MIDI2Seq

tokenizer = Tokenizer(music_token=get_token_converter(TO_TOKEN))
converter = MIDI2Seq(tokenizer, "midi_dir", "your_midi.mid", program_list=[0], split_measure=12)
converter.convert()
converter.save("output_npz_dir")

# Save tokenizer vocabulary
tokenizer.save("vocab_output_dir")

Inference

MORTM: Melody Generation

import torch
import numpy as np
from mortm.models.mortm import MORTM, MORTMArgs
from mortm.train.tokenizer import Tokenizer, get_token_converter, TO_MUSIC
from mortm.de_convert import ct_token_to_midi
from mortm.models.modules.progress import _DefaultLearningProgress

DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer = Tokenizer(music_token=get_token_converter(TO_MUSIC), load_data="vocab_list.json")
args = MORTMArgs("configs/models/mortm/A.json")
model = MORTM(progress=_DefaultLearningProgress(), args=args)
model.load_state_dict(torch.load("trained_mortm.pth", map_location=DEVICE))
model.to(DEVICE).eval()

seed_ids = torch.tensor([tokenizer.get("<GEN>"), tokenizer.get("<TS>")], device=DEVICE)
with torch.no_grad():
    _, full_seq = model.top_p_sampling_measure(seed_ids, p=0.95, max_measure=8, temperature=0.7)

ct_token_to_midi(tokenizer, full_seq, "generated_melody.mid", program=0, tempo=120)

BERTM: Music Classification

import torch
import numpy as np
import torch.nn.functional as F
from mortm.models.bertm import BERTM, MORTMArgs as BERTMArgs
from mortm.train.tokenizer import Tokenizer, get_token_converter, TO_MUSIC
from mortm.models.modules.progress import _DefaultLearningProgress

DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer = Tokenizer(music_token=get_token_converter(TO_MUSIC), load_data="vocab_list.json")
args = BERTMArgs("configs/models/bertm/class_file.json")
model = BERTM(progress=_DefaultLearningProgress(), args=args)
model.load_state_dict(torch.load("trained_bertm.pth", map_location=DEVICE))
model.to(DEVICE).eval()

input_npz = np.load("input_music.npz")['array1']
input_ids = torch.tensor(input_npz, dtype=torch.long, device=DEVICE).unsqueeze(0)
with torch.no_grad():
    logits = model(input_ids)
    probs = F.softmax(logits, dim=-1)
pred = "Human" if probs.argmax() == 0 else "AI"
print(f"Prediction: {pred}, Probabilities: {probs.squeeze().tolist()}")

Training

Train MORTM

python train_mortm_example.py --model_config configs/models/mortm/A.json \
    --train_config configs/train/pre_training.json \
    --root_directory path/to/npz_dataset \
    --save_directory out/models_mortm \
    --version MyMORTM_v1

Train V_MORTM

python train_v_mortm_example.py --model_config configs/models/v_mortm/A.json \
    --train_config configs/train/pre_training.json \
    --root_directory path/to/wav_dataset \
    --save_directory out/models_v_mortm \
    --version MyV_MORTM_v1

Train BERTM

python train_bertm_example.py --model_config configs/models/bertm/class_file.json \
    --train_config configs/train/pre_training.json \
    --human_dir path/to/human_npz \
    --ai_dir path/to/ai_npz \
    --save_directory out/models_bertm \
    --version MyBERTM_v1

Troubleshooting

• load_state_dict errors: Check config and map_location.
• Inference errors: Ensure correct tensor shapes and vocab.
• CUDA OOM: Reduce batch size or use smaller model.
• FlashAttention2 issues: Verify CUDA and compiler compatibility.

Token Format (Example)

<Gen> <TS>
Pitch=64 Duration=8 Position=0
Pitch=66 Duration=8 Position=8
...
<TE> <SME>

• Pitch: MIDI note number (e.g., 64 = E4)
• Duration: Length in ticks (8 ticks = eighth note)
• Position: Start position within the bar (0–95)
• <SME>: End of bar
• <TS>/<TE>: Track start/end tokens

Training & Generation Workflow

1. Convert MIDI data into normalized token sequences.
2. Input one bar of tokens to the decoder.
3. Autoregressively generate the next bar.
4. Evaluation Metrics:

• Does each bar sum to 64 ticks?
• Frequency of out-of-scale notes.
• Musicality and phrasing smoothness.

Model Variants

Defined in JSON configs (configs/models/...). Example:

Parameter	Value	Description
d_model	512	Embedding dimension
num_heads	8	Attention heads
num_layers	12	Decoder layers
dim_feedforward	2048	FFN dimension
num_experts	16	MoE experts
topk_experts	2	Active experts per token
vocab_size	...	From vocab_list.json

Model	Layers	Experts	Shared Experts	Embedding Dim	Heads
MORTM-C	12	6	1	512	8
MORTM-B	12	12	1	512	8
MORTM-A	12	16	1	512	8
MORTM-S	12	24	1	512	8
MORTM-SS	12	64	1	512	8

🛠️ Techniques

• FlashAttention2: Efficient exact attention.
• ALiBi: Linear biases for relative positions.
• MoE: Sparse mixture-of-experts.
• Absolute PE: Optional for phrase preservation.

License

MIT License

References

• Vaswani et al., "Attention is All You Need"
• Dao et al., "FlashAttention: Fast and Memory-Efficient Exact Attention"
• Press et al., "Train Short, Test Long: Attention with Linear Biases"
• Shazeer et al., "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer"
• Huang et al., "Music Transformer"

Author

Takaaki Nagoshi Graduate School of Integrated Basic Sciences, Nihon University [e-mail]