You can deploy this model and start earning money today!

xingjianll/
midi-gpt2

Name: midi-gpt2
Availability: InStock

safetensors

gpt2

Views

No views yet

https://github.com/xingjianll/symbolic-music-generation/

python

from miditok import PerTok, TokenizerConfig
from transformers import AutoModelForCausalLM, GenerationConfig
import torch
import io
import mido
import rtmidi

# Load tokenizer config and tokenizer
config = TokenizerConfig(
    num_velocities=8,
    use_velocities=True,
    use_chords=False,
    use_rests=True,
    use_tempos=True,
    use_time_signatures=False,
    use_sustain_pedals=False,
    use_pitch_bends=False,
    use_pitch_intervals=False,
    use_programs=False,
    use_pitchdrum_tokens=False,
    ticks_per_quarter=320,
    use_microtiming=False,
    max_microtiming_shift=0.125
)
tokenizer = PerTok(config)
tokenizer.from_pretrained("xingjianll/midi-tokenizer")

# Load model and generation config
model = AutoModelForCausalLM.from_pretrained("xingjianll/midi-gpt2")
gen_config = GenerationConfig.from_pretrained("xingjianll/midi-gpt2")
model.eval()

# Generate token sequence
with torch.no_grad():
    input_ids = torch.tensor([[tokenizer["BOS_None"]]], dtype=torch.long)
    output = model.generate(
        input_ids=input_ids,
        generation_config=gen_config
    )

# Decode tokens to MIDI bytes
generated_ids = output[0].tolist()
midi_bytes = tokenizer.decode([generated_ids]).dumps_midi()
midi_file = io.BytesIO(midi_bytes)

# Optionally save to disk
# with open("output.mid", "wb") as f:
#     f.write(midi_bytes)

# Setup MIDI output to GarageBand
midiout = rtmidi.MidiOut()
available_ports = midiout.get_ports()
print("Available MIDI ports:", available_ports)

# Choose GarageBand's virtual port
port = mido.open_output('GarageBand Virtual In')

# Play MIDI via GarageBand
midi = mido.MidiFile(file=midi_file)
for msg in midi.play():
    print(msg)
    port.send(msg)