|
|
import argparse |
|
|
import glob |
|
|
import os.path |
|
|
|
|
|
import torch |
|
|
import torch.nn.functional as F |
|
|
|
|
|
import gradio as gr |
|
|
import numpy as np |
|
|
import onnxruntime as rt |
|
|
import tqdm |
|
|
import json |
|
|
|
|
|
from midi_synthesizer import synthesis |
|
|
import TMIDIX |
|
|
|
|
|
in_space = os.getenv("SYSTEM") == "spaces" |
|
|
|
|
|
|
|
|
|
|
|
def generate( |
|
|
start_tokens, |
|
|
seq_len, |
|
|
max_seq_len = 2048, |
|
|
temperature = 0.9, |
|
|
verbose=True, |
|
|
return_prime=False, |
|
|
): |
|
|
|
|
|
out = torch.LongTensor([start_tokens]) |
|
|
|
|
|
st = len(start_tokens) |
|
|
|
|
|
if verbose: |
|
|
print("Generating sequence of max length:", seq_len) |
|
|
|
|
|
for s in range(seq_len): |
|
|
x = out[:, -max_seq_len:] |
|
|
|
|
|
torch_in = x.tolist()[0] |
|
|
|
|
|
logits = torch.FloatTensor(session.run(None, {'input': [torch_in]})[0])[:, -1] |
|
|
|
|
|
filtered_logits = logits |
|
|
|
|
|
probs = F.softmax(filtered_logits / temperature, dim=-1) |
|
|
|
|
|
sample = torch.multinomial(probs, 1) |
|
|
|
|
|
out = torch.cat((out, sample), dim=-1) |
|
|
|
|
|
if verbose: |
|
|
if s % 32 == 0: |
|
|
print(s, '/', seq_len) |
|
|
|
|
|
if return_prime: |
|
|
return out[:, :] |
|
|
|
|
|
else: |
|
|
return out[:, st:] |
|
|
|
|
|
|
|
|
|
|
|
def GenerateMIDI(params): |
|
|
|
|
|
melody_chords_f = generate([3087, 3073+1, 3075+1], 512) |
|
|
|
|
|
melody_chords_f = melody_chords_f.tolist()[0] |
|
|
|
|
|
print('=' * 70) |
|
|
print('Sample INTs', melody_chords_f[:12]) |
|
|
print('=' * 70) |
|
|
|
|
|
if len(melody_chords_f) != 0: |
|
|
|
|
|
song = melody_chords_f |
|
|
song_f = [] |
|
|
time = 0 |
|
|
dur = 0 |
|
|
vel = 0 |
|
|
pitch = 0 |
|
|
channel = 0 |
|
|
|
|
|
for ss in song: |
|
|
|
|
|
if ss > 0 and ss < 256: |
|
|
|
|
|
time += ss * 8 |
|
|
|
|
|
if ss >= 256 and ss < 1280: |
|
|
|
|
|
dur = ((ss-256) // 8) * 32 |
|
|
vel = (((ss-256) % 8)+1) * 15 |
|
|
|
|
|
if ss >= 1280 and ss < 2816: |
|
|
channel = (ss-1280) // 128 |
|
|
pitch = (ss-1280) % 128 |
|
|
|
|
|
song_f.append(['note', time, dur, channel, pitch, vel ]) |
|
|
|
|
|
detailed_stats = TMIDIX.Tegridy_SONG_to_MIDI_Converter(song_f, |
|
|
output_signature = 'Allegro Music Transformer', |
|
|
output_file_name = 'Allegro-Music-Transformer-Music-Composition', |
|
|
track_name='Project Los Angeles', |
|
|
list_of_MIDI_patches=[0, 24, 32, 40, 42, 46, 56, 71, 73, 0, 53, 19, 0, 0, 0, 0], |
|
|
number_of_ticks_per_quarter=500) |
|
|
print('=' * 70) |
|
|
|
|
|
|
|
|
|
|
|
def cancel_run(mid_seq): |
|
|
if mid_seq is None: |
|
|
return None, None |
|
|
mid = tokenizer.detokenize(mid_seq) |
|
|
with open(f"output.mid", 'wb') as f: |
|
|
f.write(MIDI.score2midi(mid)) |
|
|
audio = synthesis(MIDI.score2opus(mid), soundfont_path) |
|
|
return "output.mid", (44100, audio), [create_msg("visualizer_end", None)] |
|
|
|
|
|
def load_javascript(dir="javascript"): |
|
|
scripts_list = glob.glob(f"*.js") |
|
|
javascript = "" |
|
|
for path in scripts_list: |
|
|
with open(path, "r", encoding="utf8") as jsfile: |
|
|
javascript += f"\n<!-- {path} --><script>{jsfile.read()}</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 |
|
|
|
|
|
|
|
|
class JSMsgReceiver(gr.HTML): |
|
|
|
|
|
def __init__(self, **kwargs): |
|
|
super().__init__(elem_id="msg_receiver", visible=False, **kwargs) |
|
|
|
|
|
def postprocess(self, y): |
|
|
if y: |
|
|
y = f"<p>{json.dumps(y)}</p>" |
|
|
return super().postprocess(y) |
|
|
|
|
|
def get_block_name(self) -> str: |
|
|
return "html" |
|
|
|
|
|
|
|
|
|
|
|
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") |
|
|
opt = parser.parse_args() |
|
|
|
|
|
session = rt.InferenceSession('Allegro_Music_Transformer_Small_Trained_Model_56000_steps_0.9399_loss_0.7374_acc.onnx', providers=['CUDAExecutionProvider']) |
|
|
|
|
|
load_javascript() |
|
|
app = gr.Blocks() |
|
|
with app: |
|
|
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Allegro Music Transformer</h1>") |
|
|
gr.Markdown("\n\n" |
|
|
"Full-attention multi-instrumental music transformer featuring asymmetrical encoding with octo-velocity, and chords counters tokens, optimized for speed and performance\n\n" |
|
|
"Check out [Allegro Music Transformer](https://github.com/asigalov61/Allegro-Music-Transformer) on GitHub!\n\n" |
|
|
"[Open In Colab]" |
|
|
"(https://colab.research.google.com/github/asigalov61/Allegro-Music-Transformer/blob/main/Allegro_Music_Transformer_Composer.ipynb)" |
|
|
" for faster execution and endless generation" |
|
|
) |
|
|
|
|
|
js_msg = JSMsgReceiver() |
|
|
|
|
|
run_btn = gr.Button("generate", variant="primary") |
|
|
stop_btn = gr.Button("stop and output") |
|
|
|
|
|
output_audio = gr.Audio(label="output audio", format="mp3", elem_id="midi_audio") |
|
|
output_midi = gr.File(label="output midi", file_types=[".mid"]) |
|
|
run_event = run_btn.click(GenerateMIDI, [output_midi_seq, output_midi, output_audio, js_msg]) |
|
|
stop_btn.click(cancel_run, output_midi_seq, [output_midi, output_audio, js_msg], cancels=run_event, queue=False) |
|
|
|
|
|
app.queue(2).launch(server_port=opt.port, share=opt.share, inbrowser=True) |
