import spaces
import random
import argparse
import glob
import json
import os
import time
from concurrent.futures import ThreadPoolExecutor
import gradio as gr
import numpy as np
import torch
import torch.nn.functional as F
import tqdm
from huggingface_hub import hf_hub_download
from transformers import DynamicCache
import MIDI
from midi_model import MIDIModel, MIDIModelConfig
from midi_synthesizer import MidiSynthesizer
MAX_SEED = np.iinfo(np.int32).max
in_space = os.getenv("SYSTEM") == "spaces"
@torch.inference_mode()
def generate(model: MIDIModel, prompt=None, batch_size=1, max_len=512, temp=1.0, top_p=0.98, top_k=20,
disable_patch_change=False, disable_control_change=False, disable_channels=None, generator=None):
tokenizer = model.tokenizer
if disable_channels is not None:
disable_channels = [tokenizer.parameter_ids["channel"][c] for c in disable_channels]
else:
disable_channels = []
max_token_seq = tokenizer.max_token_seq
if prompt is None:
input_tensor = torch.full((1, max_token_seq), tokenizer.pad_id, dtype=torch.long, device=model.device)
input_tensor[0, 0] = tokenizer.bos_id # bos
input_tensor = input_tensor.unsqueeze(0)
input_tensor = torch.cat([input_tensor] * batch_size, dim=0)
else:
if len(prompt.shape) == 2:
prompt = prompt[None, :]
prompt = np.repeat(prompt, repeats=batch_size, axis=0)
elif prompt.shape[0] == 1:
prompt = np.repeat(prompt, repeats=batch_size, axis=0)
elif len(prompt.shape) != 3 or prompt.shape[0] != batch_size:
raise ValueError(f"invalid shape for prompt, {prompt.shape}")
prompt = prompt[..., :max_token_seq]
if prompt.shape[-1] < max_token_seq:
prompt = np.pad(prompt, ((0, 0), (0, 0), (0, max_token_seq - prompt.shape[-1])),
mode="constant", constant_values=tokenizer.pad_id)
input_tensor = torch.from_numpy(prompt).to(dtype=torch.long, device=model.device)
cur_len = input_tensor.shape[1]
bar = tqdm.tqdm(desc="generating", total=max_len - cur_len)
cache1 = DynamicCache()
past_len = 0
with bar:
while cur_len < max_len:
end = [False] * batch_size
hidden = model.forward(input_tensor[:, past_len:], cache=cache1)[:, -1]
next_token_seq = None
event_names = [""] * batch_size
cache2 = DynamicCache()
for i in range(max_token_seq):
mask = torch.zeros((batch_size, tokenizer.vocab_size), dtype=torch.int64, device=model.device)
for b in range(batch_size):
if end[b]:
mask[b, tokenizer.pad_id] = 1
continue
if i == 0:
mask_ids = list(tokenizer.event_ids.values()) + [tokenizer.eos_id]
if disable_patch_change:
mask_ids.remove(tokenizer.event_ids["patch_change"])
if disable_control_change:
mask_ids.remove(tokenizer.event_ids["control_change"])
mask[b, mask_ids] = 1
else:
param_names = tokenizer.events[event_names[b]]
if i > len(param_names):
mask[b, tokenizer.pad_id] = 1
continue
param_name = param_names[i - 1]
mask_ids = tokenizer.parameter_ids[param_name]
if param_name == "channel":
mask_ids = [i for i in mask_ids if i not in disable_channels]
mask[b, mask_ids] = 1
mask = mask.unsqueeze(1)
x = next_token_seq
if i != 0:
hidden = None
x = x[:, -1:]
logits = model.forward_token(hidden, x, cache=cache2)[:, -1:]
scores = torch.softmax(logits / temp, dim=-1) * mask
samples = model.sample_top_p_k(scores, top_p, top_k, generator=generator)
if i == 0:
next_token_seq = samples
for b in range(batch_size):
if end[b]:
continue
eid = samples[b].item()
if eid == tokenizer.eos_id:
end[b] = True
else:
event_names[b] = tokenizer.id_events[eid]
else:
next_token_seq = torch.cat([next_token_seq, samples], dim=1)
if all([len(tokenizer.events[event_names[b]]) == i for b in range(batch_size) if not end[b]]):
break
if next_token_seq.shape[1] < max_token_seq:
next_token_seq = F.pad(next_token_seq, (0, max_token_seq - next_token_seq.shape[1]),
"constant", value=tokenizer.pad_id)
next_token_seq = next_token_seq.unsqueeze(1)
input_tensor = torch.cat([input_tensor, next_token_seq], dim=1)
past_len = cur_len
cur_len += 1
bar.update(1)
yield next_token_seq[:, 0].cpu().numpy()
if all(end):
break
def create_msg(name, data):
return {"name": name, "data": data}
def send_msgs(msgs):
return json.dumps(msgs)
def get_duration(model_name, tab, mid_seq, continuation_state, continuation_select, instruments, drum_kit, bpm,
time_sig, key_sig, mid, midi_events, reduce_cc_st, remap_track_channel, add_default_instr,
remove_empty_channels, seed, seed_rand, gen_events, temp, top_p, top_k, allow_cc):
t = gen_events // 23
if "large" in model_name:
t = gen_events // 14
return t + 5
@spaces.GPU(duration=get_duration)
def run(model_name, tab, mid_seq, continuation_state, continuation_select, instruments, drum_kit, bpm, time_sig,
key_sig, mid, midi_events, reduce_cc_st, remap_track_channel, add_default_instr, remove_empty_channels,
seed, seed_rand, gen_events, temp, top_p, top_k, allow_cc):
model = models[model_name]
model.to(device=opt.device)
tokenizer = model.tokenizer
bpm = int(bpm)
if time_sig == "auto":
time_sig = None
time_sig_nn = 4
time_sig_dd = 2
else:
time_sig_nn, time_sig_dd = time_sig.split('/')
time_sig_nn = int(time_sig_nn)
time_sig_dd = {2: 1, 4: 2, 8: 3}[int(time_sig_dd)]
if key_sig == 0:
key_sig = None
key_sig_sf = 0
key_sig_mi = 0
else:
key_sig = (key_sig - 1)
key_sig_sf = key_sig // 2 - 7
key_sig_mi = key_sig % 2
gen_events = int(gen_events)
max_len = gen_events
if seed_rand:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator(opt.device).manual_seed(seed)
disable_patch_change = False
disable_channels = None
if tab == 0:
i = 0
mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)]
if tokenizer.version == "v2":
if time_sig is not None:
mid.append(tokenizer.event2tokens(["time_signature", 0, 0, 0, time_sig_nn - 1, time_sig_dd - 1]))
if key_sig is not None:
mid.append(tokenizer.event2tokens(["key_signature", 0, 0, 0, key_sig_sf + 7, key_sig_mi]))
if bpm != 0:
mid.append(tokenizer.event2tokens(["set_tempo", 0, 0, 0, bpm]))
patches = {}
if instruments is None:
instruments = []
for instr in instruments:
patches[i] = patch2number[instr]
i = (i + 1) if i != 8 else 10
if drum_kit != "None":
patches[9] = drum_kits2number[drum_kit]
for i, (c, p) in enumerate(patches.items()):
mid.append(tokenizer.event2tokens(["patch_change", 0, 0, i + 1, c, p]))
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
if len(instruments) > 0:
disable_patch_change = True
disable_channels = [i for i in range(16) if i not in patches]
elif tab == 1 and mid is not None:
eps = 4 if reduce_cc_st else 0
mid = tokenizer.tokenize(MIDI.midi2score(mid), cc_eps=eps, tempo_eps=eps,
remap_track_channel=remap_track_channel,
add_default_instr=add_default_instr,
remove_empty_channels=remove_empty_channels)
mid = mid[:int(midi_events)]
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
elif tab == 2 and mid_seq is not None:
mid = np.asarray(mid_seq, dtype=np.int64)
if continuation_select > 0:
continuation_state.append(mid_seq)
mid = np.repeat(mid[continuation_select - 1:continuation_select], repeats=OUTPUT_BATCH_SIZE, axis=0)
mid_seq = mid.tolist()
else:
continuation_state.append(mid.shape[1])
else:
continuation_state = [0]
mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)]
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
if mid is not None:
max_len += mid.shape[1]
init_msgs = [create_msg("progress", [0, gen_events])]
if not (tab == 2 and continuation_select == 0):
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
init_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events])]
yield mid_seq, continuation_state, seed, send_msgs(init_msgs)
midi_generator = generate(model, mid, batch_size=OUTPUT_BATCH_SIZE, max_len=max_len, temp=temp,
top_p=top_p, top_k=top_k, disable_patch_change=disable_patch_change,
disable_control_change=not allow_cc, disable_channels=disable_channels,
generator=generator)
events = [list() for i in range(OUTPUT_BATCH_SIZE)]
t = time.time() + 1
for i, token_seqs in enumerate(midi_generator):
token_seqs = token_seqs.tolist()
for j in range(OUTPUT_BATCH_SIZE):
token_seq = token_seqs[j]
mid_seq[j].append(token_seq)
events[j].append(tokenizer.tokens2event(token_seq))
if time.time() - t > 0.5:
msgs = [create_msg("progress", [i + 1, gen_events])]
for j in range(OUTPUT_BATCH_SIZE):
msgs += [create_msg("visualizer_append", [j, events[j]])]
events[j] = list()
yield mid_seq, continuation_state, seed, send_msgs(msgs)
t = time.time()
yield mid_seq, continuation_state, seed, send_msgs([])
def finish_run(model_name, mid_seq):
if mid_seq is None:
outputs = [None] * OUTPUT_BATCH_SIZE
return *outputs, []
tokenizer = models[model_name].tokenizer
outputs = []
end_msgs = [create_msg("progress", [0, 0])]
if not os.path.exists("outputs"):
os.mkdir("outputs")
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
mid = tokenizer.detokenize(mid_seq[i])
with open(f"outputs/output{i + 1}.mid", 'wb') as f:
f.write(MIDI.score2midi(mid))
outputs.append(f"outputs/output{i + 1}.mid")
end_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events]),
create_msg("visualizer_end", i)]
return *outputs, send_msgs(end_msgs)
def synthesis_task(mid):
return synthesizer.synthesis(MIDI.score2opus(mid))
def render_audio(model_name, mid_seq, should_render_audio):
if (not should_render_audio) or mid_seq is None:
outputs = [None] * OUTPUT_BATCH_SIZE
return tuple(outputs)
tokenizer = models[model_name].tokenizer
outputs = []
if not os.path.exists("outputs"):
os.mkdir("outputs")
audio_futures = []
for i in range(OUTPUT_BATCH_SIZE):
mid = tokenizer.detokenize(mid_seq[i])
audio_future = thread_pool.submit(synthesis_task, mid)
audio_futures.append(audio_future)
for future in audio_futures:
outputs.append((44100, future.result()))
if OUTPUT_BATCH_SIZE == 1:
return outputs[0]
return tuple(outputs)
def undo_continuation(model_name, mid_seq, continuation_state):
if mid_seq is None or len(continuation_state) < 2:
return mid_seq, continuation_state, send_msgs([])
tokenizer = models[model_name].tokenizer
if isinstance(continuation_state[-1], list):
mid_seq = continuation_state[-1]
else:
mid_seq = [ms[:continuation_state[-1]] for ms in mid_seq]
continuation_state = continuation_state[:-1]
end_msgs = [create_msg("progress", [0, 0])]
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
end_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events]),
create_msg("visualizer_end", i)]
return mid_seq, continuation_state, send_msgs(end_msgs)
def load_javascript(dir="javascript"):
scripts_list = glob.glob(f"{dir}/*.js")
javascript = ""
for path in scripts_list:
with open(path, "r", encoding="utf8") as jsfile:
js_content = jsfile.read()
js_content = js_content.replace("const MIDI_OUTPUT_BATCH_SIZE=4;",
f"const MIDI_OUTPUT_BATCH_SIZE={OUTPUT_BATCH_SIZE};")
javascript += f"\n<!-- {path} --><script>{js_content}</script>"
template_response_ori = gr.routes.templates.TemplateResponse
def template_response(*args, **kwargs):
res = template_response_ori(*args, **kwargs)
res.body = res.body.replace(
b'</head>', f'{javascript}</head>'.encode("utf8"))
res.init_headers()
return res
gr.routes.templates.TemplateResponse = template_response
def hf_hub_download_retry(repo_id, filename):
print(f"downloading {repo_id} {filename}")
retry = 0
err = None
while retry < 30:
try:
return hf_hub_download(repo_id=repo_id, filename=filename)
except Exception as e:
err = e
retry += 1
if err:
raise err
number2drum_kits = {-1: "None", 0: "Standard", 8: "Room", 16: "Power", 24: "Electric", 25: "TR-808", 32: "Jazz",
40: "Blush", 48: "Orchestra"}
patch2number = {v: k for k, v in MIDI.Number2patch.items()}
drum_kits2number = {v: k for k, v in number2drum_kits.items()}
key_signatures = ['C♭', 'A♭m', 'G♭', 'E♭m', 'D♭', 'B♭m', 'A♭', 'Fm', 'E♭', 'Cm', 'B♭', 'Gm', 'F', 'Dm',
'C', 'Am', 'G', 'Em', 'D', 'Bm', 'A', 'F♯m', 'E', 'C♯m', 'B', 'G♯m', 'F♯', 'D♯m', 'C♯', 'A♯m']
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
parser.add_argument("--port", type=int, default=7860, help="gradio server port")
parser.add_argument("--device", type=str, default="cpu", help="device to run model")
parser.add_argument("--batch", type=int, default=8, help="batch size")
parser.add_argument("--max-gen", type=int, default=1024, help="max")
opt = parser.parse_args()
OUTPUT_BATCH_SIZE = opt.batch
soundfont_path = hf_hub_download_retry(repo_id="skytnt/midi-model", filename="soundfont.sf2")
thread_pool = ThreadPoolExecutor(max_workers=OUTPUT_BATCH_SIZE)
synthesizer = MidiSynthesizer(soundfont_path)
models_info = {
"generic pretrain model (tv2o-medium) by skytnt": [
"skytnt/midi-model-tv2o-medium", {
"jpop": "skytnt/midi-model-tv2om-jpop-lora",
"touhou": "skytnt/midi-model-tv2om-touhou-lora"
}
],
"generic pretrain model (tv2o-large) by asigalov61": [
"asigalov61/Music-Llama", {}
],
"generic pretrain model (tv2o-medium) by asigalov61": [
"asigalov61/Music-Llama-Medium", {}
],
"generic pretrain model (tv1-medium) by skytnt": [
"skytnt/midi-model", {}
]
}
models = {}
if opt.device == "cuda":
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)
for name, (repo_id, loras) in models_info.items():
model = MIDIModel.from_pretrained(repo_id)
model.to(device="cpu", dtype=torch.float32)
models[name] = model
for lora_name, lora_repo in loras.items():
model = MIDIModel.from_pretrained(repo_id)
print(f"loading lora {lora_repo} for {name}")
model = model.load_merge_lora(lora_repo)
model.to(device="cpu", dtype=torch.float32)
models[f"{name} with {lora_name} lora"] = model
load_javascript()
app = gr.Blocks(theme=gr.themes.Soft())
with app:
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Midi Composer on CPU</h1>")
gr.Markdown("\n\n"
"Midi event transformer for symbolic music generation\n\n"
"Demo for [SkyTNT/midi-model](https://github.com/SkyTNT/midi-model)\n\n"
"[Open In Colab]"
"(https://colab.research.google.com/github/SkyTNT/midi-model/blob/main/demo.ipynb)"
" or [download windows app](https://github.com/SkyTNT/midi-model/releases)"
" for unlimited generation\n\n"
"**Update v1.3**: MIDITokenizerV2 and new MidiVisualizer\n\n"
"The current **best** model: generic pretrain model (tv2o-medium) by skytnt"
)
js_msg = gr.Textbox(elem_id="msg_receiver", visible=False)
js_msg.change(None, [js_msg], [], js="""
(msg_json) =>{
let msgs = JSON.parse(msg_json);
executeCallbacks(msgReceiveCallbacks, msgs);
return [];
}
""")
input_model = gr.Dropdown(label="select model", choices=list(models.keys()),
type="value", value=list(models.keys())[0])
tab_select = gr.State(value=0)
with gr.Tabs():
with gr.TabItem("custom prompt") as tab1:
input_instruments = gr.Dropdown(label="🪗instruments (auto if empty)", choices=list(patch2number.keys()),
multiselect=True, max_choices=15, type="value")
input_drum_kit = gr.Dropdown(label="🥁drum kit", choices=list(drum_kits2number.keys()), type="value",
value="None")
input_bpm = gr.Slider(label="BPM (beats per minute, auto if 0)", minimum=0, maximum=255,
step=1,
value=0)
input_time_sig = gr.Radio(label="time signature (only for tv2 models)",
value="auto",
choices=["auto", "4/4", "2/4", "3/4", "6/4", "7/4",
"2/2", "3/2", "4/2", "3/8", "5/8", "6/8", "7/8", "9/8", "12/8"]
)
input_key_sig = gr.Radio(label="key signature (only for tv2 models)",
value="auto",
choices=["auto"] + key_signatures,
type="index"
)
example1 = gr.Examples([
[[], "None"],
[["Acoustic Grand"], "None"],
[['Acoustic Grand', 'SynthStrings 2', 'SynthStrings 1', 'Pizzicato Strings',
'Pad 2 (warm)', 'Tremolo Strings', 'String Ensemble 1'], "Orchestra"],
[['Trumpet', 'Oboe', 'Trombone', 'String Ensemble 1', 'Clarinet',
'French Horn', 'Pad 4 (choir)', 'Bassoon', 'Flute'], "None"],
[['Flute', 'French Horn', 'Clarinet', 'String Ensemble 2', 'English Horn', 'Bassoon',
'Oboe', 'Pizzicato Strings'], "Orchestra"],
[['Electric Piano 2', 'Lead 5 (charang)', 'Electric Bass(pick)', 'Lead 2 (sawtooth)',
'Pad 1 (new age)', 'Orchestra Hit', 'Cello', 'Electric Guitar(clean)'], "Standard"],
[["Electric Guitar(clean)", "Electric Guitar(muted)", "Overdriven Guitar", "Distortion Guitar",
"Electric Bass(finger)"], "Standard"]
], [input_instruments, input_drum_kit])
with gr.TabItem("midi prompt") as tab2:
input_midi = gr.File(label="input midi", file_types=[".midi", ".mid"], type="binary")
input_midi_events = gr.Slider(label="use first n midi events as prompt", minimum=1, maximum=512,
step=1,
value=128)
input_reduce_cc_st = gr.Checkbox(label="reduce control_change and set_tempo events", value=True)
input_remap_track_channel = gr.Checkbox(
label="remap tracks and channels so each track has only one channel and in order", value=True)
input_add_default_instr = gr.Checkbox(
label="add a default instrument to channels that don't have an instrument", value=True)
input_remove_empty_channels = gr.Checkbox(label="remove channels without notes", value=False)
example2 = gr.Examples([[file, 128] for file in glob.glob("example/*.mid")],
[input_midi, input_midi_events])
with gr.TabItem("last output prompt") as tab3:
gr.Markdown("Continue generating on the last output.")
input_continuation_select = gr.Radio(label="select output to continue generating", value="all",
choices=["all"] + [f"output{i + 1}" for i in
range(OUTPUT_BATCH_SIZE)],
type="index"
)
undo_btn = gr.Button("undo the last continuation")
tab1.select(lambda: 0, None, tab_select, queue=False)
tab2.select(lambda: 1, None, tab_select, queue=False)
tab3.select(lambda: 2, None, tab_select, queue=False)
input_seed = gr.Slider(label="seed", minimum=0, maximum=2 ** 31 - 1,
step=1, value=0)
input_seed_rand = gr.Checkbox(label="random seed", value=True)
input_gen_events = gr.Slider(label="generate max n midi events", minimum=1, maximum=opt.max_gen,
step=1, value=opt.max_gen // 2)
with gr.Accordion("options", open=False):
input_temp = gr.Slider(label="temperature", minimum=0.1, maximum=1.2, step=0.01, value=1)
input_top_p = gr.Slider(label="top p", minimum=0.1, maximum=1, step=0.01, value=0.95)
input_top_k = gr.Slider(label="top k", minimum=1, maximum=128, step=1, value=20)
input_allow_cc = gr.Checkbox(label="allow midi cc event", value=True)
input_render_audio = gr.Checkbox(label="render audio after generation", value=True)
example3 = gr.Examples([[1, 0.94, 128], [1, 0.98, 20], [1, 0.98, 12]],
[input_temp, input_top_p, input_top_k])
run_btn = gr.Button("generate", variant="primary")
# stop_btn = gr.Button("stop and output")
output_midi_seq = gr.State()
output_continuation_state = gr.State([0])
midi_outputs = []
audio_outputs = []
with gr.Tabs(elem_id="output_tabs"):
for i in range(OUTPUT_BATCH_SIZE):
with gr.TabItem(f"output {i + 1}") as tab1:
output_midi_visualizer = gr.HTML(elem_id=f"midi_visualizer_container_{i}")
output_audio = gr.Audio(label="output audio", format="mp3", elem_id=f"midi_audio_{i}")
output_midi = gr.File(label="output midi", file_types=[".mid"])
midi_outputs.append(output_midi)
audio_outputs.append(output_audio)
run_event = run_btn.click(run, [input_model, tab_select, output_midi_seq, output_continuation_state,
input_continuation_select, input_instruments, input_drum_kit, input_bpm,
input_time_sig, input_key_sig, input_midi, input_midi_events,
input_reduce_cc_st, input_remap_track_channel,
input_add_default_instr, input_remove_empty_channels,
input_seed, input_seed_rand, input_gen_events, input_temp, input_top_p,
input_top_k, input_allow_cc],
[output_midi_seq, output_continuation_state, input_seed, js_msg],
concurrency_limit=10, queue=True)
finish_run_event = run_event.then(fn=finish_run,
inputs=[input_model, output_midi_seq],
outputs=midi_outputs + [js_msg],
queue=False)
finish_run_event.then(fn=render_audio,
inputs=[input_model, output_midi_seq, input_render_audio],
outputs=audio_outputs,
queue=False)
# stop_btn.click(None, [], [], cancels=run_event,
# queue=False)
undo_btn.click(undo_continuation, [input_model, output_midi_seq, output_continuation_state],
[output_midi_seq, output_continuation_state, js_msg], queue=False)
app.queue().launch(server_port=opt.port, share=opt.share, inbrowser=True, ssr_mode=False)
thread_pool.shutdown()