|
|
|
|
|
|
|
|
|
import os |
|
import time as reqtime |
|
import datetime |
|
from pytz import timezone |
|
|
|
import torch |
|
from imagen_pytorch import Unet, Imagen, ImagenTrainer |
|
from imagen_pytorch.data import Dataset |
|
|
|
import spaces |
|
import gradio as gr |
|
|
|
import random |
|
import tqdm |
|
|
|
from midi_to_colab_audio import midi_to_colab_audio |
|
import TMIDIX |
|
|
|
|
|
|
|
@spaces.GPU |
|
def GenerateSong(input_melody_seed_number): |
|
print('=' * 70) |
|
print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT))) |
|
start_time = reqtime.time() |
|
|
|
print('Loading model...') |
|
|
|
SEQ_LEN = 2560 |
|
PAD_IDX = 514 |
|
DEVICE = 'cuda' |
|
|
|
|
|
|
|
model = TransformerWrapper( |
|
num_tokens = PAD_IDX+1, |
|
max_seq_len = SEQ_LEN, |
|
attn_layers = Decoder(dim = 1024, depth = 24, heads = 16, attn_flash = True) |
|
) |
|
|
|
model = AutoregressiveWrapper(model, ignore_index = PAD_IDX) |
|
|
|
model.to(DEVICE) |
|
print('=' * 70) |
|
|
|
print('Loading model checkpoint...') |
|
|
|
model.load_state_dict( |
|
torch.load('Melody2Song_Seq2Seq_Music_Transformer_Trained_Model_28482_steps_0.719_loss_0.7865_acc.pth', |
|
map_location=DEVICE)) |
|
print('=' * 70) |
|
|
|
model.eval() |
|
|
|
if DEVICE == 'cpu': |
|
dtype = torch.bfloat16 |
|
else: |
|
dtype = torch.bfloat16 |
|
|
|
ctx = torch.amp.autocast(device_type=DEVICE, dtype=dtype) |
|
|
|
print('Done!') |
|
print('=' * 70) |
|
seed_melody = seed_melodies_data[input_melody_seed_number] |
|
print('Input melody seed number:', input_melody_seed_number) |
|
print('-' * 70) |
|
|
|
|
|
|
|
print('=' * 70) |
|
|
|
print('Sample output events', seed_melody[:16]) |
|
print('=' * 70) |
|
print('Generating...') |
|
|
|
x = (torch.tensor(seed_melody, dtype=torch.long, device='cuda')[None, ...]) |
|
|
|
with ctx: |
|
with torch.inference_mode(): |
|
out = model.generate(x, |
|
1024, |
|
temperature=0.9, |
|
return_prime=False, |
|
verbose=False) |
|
|
|
output = out[0].tolist() |
|
|
|
print('=' * 70) |
|
print('Done!') |
|
print('=' * 70) |
|
|
|
|
|
print('Rendering results...') |
|
|
|
print('=' * 70) |
|
print('Sample INTs', output[:15]) |
|
print('=' * 70) |
|
|
|
out1 = output |
|
|
|
if len(out1) != 0: |
|
|
|
song = out1 |
|
song_f = [] |
|
|
|
time = 0 |
|
dur = 0 |
|
vel = 90 |
|
pitch = 0 |
|
channel = 0 |
|
|
|
patches = [0] * 16 |
|
patches[3] = 40 |
|
|
|
for ss in song: |
|
|
|
if 0 < ss < 128: |
|
|
|
time += (ss * 32) |
|
|
|
if 128 < ss < 256: |
|
|
|
dur = (ss-128) * 32 |
|
|
|
if 256 < ss < 512: |
|
|
|
pitch = (ss-256) % 128 |
|
|
|
channel = (ss-256) // 128 |
|
|
|
if channel == 1: |
|
channel = 3 |
|
vel = 110 + (pitch % 12) |
|
song_f.append(['note', time, dur, channel, pitch, vel, 40]) |
|
|
|
else: |
|
vel = 80 + (pitch % 12) |
|
channel = 0 |
|
song_f.append(['note', time, dur, channel, pitch, vel, 0]) |
|
|
|
fn1 = "Melody2Song-Seq2Seq-Music-Transformer-Composition" |
|
|
|
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f, |
|
output_signature = 'Melody2Song Seq2Seq Music Transformer', |
|
output_file_name = fn1, |
|
track_name='Project Los Angeles', |
|
list_of_MIDI_patches=patches |
|
) |
|
|
|
new_fn = fn1+'.mid' |
|
|
|
|
|
audio = midi_to_colab_audio(new_fn, |
|
soundfont_path=soundfont, |
|
sample_rate=16000, |
|
volume_scale=10, |
|
output_for_gradio=True |
|
) |
|
|
|
print('Done!') |
|
print('=' * 70) |
|
|
|
|
|
|
|
output_midi_title = str(fn1) |
|
output_midi_summary = str(song_f[:3]) |
|
output_midi = str(new_fn) |
|
output_audio = (16000, audio) |
|
|
|
output_plot = TMIDIX.plot_ms_SONG(song_f, plot_title=output_midi, return_plt=True) |
|
|
|
print('Output MIDI file name:', output_midi) |
|
print('Output MIDI title:', output_midi_title) |
|
print('Output MIDI summary:', output_midi_summary) |
|
print('=' * 70) |
|
|
|
|
|
|
|
|
|
print('-' * 70) |
|
print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT))) |
|
print('-' * 70) |
|
print('Req execution time:', (reqtime.time() - start_time), 'sec') |
|
|
|
return output_midi_title, output_midi_summary, output_midi, output_audio, output_plot |
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
|
PDT = timezone('US/Pacific') |
|
|
|
print('=' * 70) |
|
print('App start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT))) |
|
print('=' * 70) |
|
|
|
soundfont = "SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2" |
|
|
|
app = gr.Blocks() |
|
|
|
with app: |
|
|
|
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Imagen POP Music Medley Diffusion Transformer</h1>") |
|
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Generate unique POP music medleys with Imagen diffusion transformer</h1>") |
|
gr.Markdown("\n\n" |
|
"This is a demo for MIDI Images dataset\n\n" |
|
"Please see [MIDI Images](https://huggingface.co/datasets/asigalov61/MIDI-Images) Hugging Face repo for more information\n\n" |
|
) |
|
|
|
input_num_medley_comps = gr.Slider(1, 64, value=8, step=1, label="Number of medley compositions") |
|
|
|
run_btn = gr.Button("Generate POP Medley", variant="primary") |
|
|
|
gr.Markdown("## Generation results") |
|
|
|
output_midi_title = gr.Textbox(label="Output MIDI title") |
|
output_midi_summary = gr.Textbox(label="Output MIDI summary") |
|
output_audio = gr.Audio(label="Output MIDI audio", format="wav", elem_id="midi_audio") |
|
output_plot = gr.Plot(label="Output MIDI score plot") |
|
output_midi = gr.File(label="Output MIDI file", file_types=[".mid"]) |
|
|
|
run_event = run_btn.click(GenerateSong, [input_melody_seed_number], |
|
[output_midi_title, output_midi_summary, output_midi, output_audio, output_plot]) |
|
|
|
app.queue().launch() |
