#==================================================================================
# https://huggingface.co/spaces/asigalov61/Delta-Melody-Transformer
#==================================================================================
print('=' * 70)
print('Delta Melody Transformer Gradio App')
print('=' * 70)
print('Loading core Delta Melody Transformer modules...')
import os
import time as reqtime
import datetime
from pytz import timezone
print('=' * 70)
print('Loading main Delta Melody Transformer modules...')
os.environ['USE_FLASH_ATTENTION'] = '1'
import torch
torch.set_float32_matmul_precision('high')
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_cudnn_sdp(True)
from huggingface_hub import hf_hub_download
import TMIDIX
from midi_to_colab_audio import midi_to_colab_audio
from x_transformer_2_3_1 import *
import random
import tqdm
print('=' * 70)
print('Loading aux Delta Melody Transformer modules...')
import matplotlib.pyplot as plt
import gradio as gr
import spaces
print('=' * 70)
print('PyTorch version:', torch.__version__)
print('=' * 70)
print('Done!')
print('Enjoy! :)')
print('=' * 70)
#==================================================================================
MODEL_CHECKPOINT = 'Delta_Melody_Transformer_Medium_Trained_Model_4382_steps_0.2846_loss_0.9193_acc.pth'
SOUNDFONT = 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'
#==================================================================================
print('Loading SoundFont...')
soundfont = hf_hub_download(repo_id='projectlosangeles/soundfonts4u',
repo_type='dataset',
filename='SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'
)
print('Done!')
print('=' * 70)
#==================================================================================
print('=' * 70)
print('Instantiating Delta Melody Transformer model...')
device_type = 'cpu'
dtype = 'bfloat16'
ptdtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)
SEQ_LEN = 194
PAD_IDX = 705
model = TransformerWrapper(
num_tokens = PAD_IDX+1,
max_seq_len = SEQ_LEN,
attn_layers = Decoder(dim = 1024,
depth = 4,
heads = 8,
rotary_pos_emb = True,
attn_flash = True
)
)
model = AutoregressiveWrapper(model, ignore_index=PAD_IDX, pad_value=PAD_IDX)
print('=' * 70)
print('Loading model checkpoint...')
model_checkpoint = hf_hub_download(repo_id='asigalov61/Delta-Melody-Transformer', filename=MODEL_CHECKPOINT)
model.load_state_dict(torch.load(model_checkpoint, map_location=device_type, weights_only=True))
model = torch.compile(model, mode='max-autotune')
model.to(device_type)
model.eval()
print('=' * 70)
print('Done!')
print('=' * 70)
print('Model will use', dtype, 'precision...')
print('=' * 70)
#==================================================================================
def generate_melody_seq(inp, model_temp=0.9, model_top_k=15):
x = torch.LongTensor(inp).to(device_type)
with ctx:
out = model.generate(x,
194,
temperature=model_temp,
filter_logits_fn=top_k,
filter_kwargs={'k': model_top_k},
return_prime=False,
eos_token=704,
verbose=True)
out = out.tolist()
return out
#==================================================================================
def Generate_Melody(melody_len,
melody_start_pitch,
melody_min_run_time,
melody_patch,
model_temperature,
model_sampling_top_k
):
#===============================================================================
print('=' * 70)
print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
start_time = reqtime.time()
print('=' * 70)
print('Requested settings:')
print('=' * 70)
print('Melody len:', melody_len)
print('Melody start pitch:', melody_start_pitch)
print('Melody min run time:', melody_min_run_time)
print('Melody patch:', melody_patch)
print('Model temperature:', model_temperature)
print('Model top k:', model_sampling_top_k)
print('=' * 70)
#==================================================================
print('=' * 70)
print('Generating...')
#==================================================================
inp = [512+melody_len]
mel_run_time = -1
tries = 0
while mel_run_time < melody_min_run_time and tries < 50:
out = generate_melody_seq(inp,
model_temperature,
model_sampling_top_k
)
mel_run_time = (sum(t for t in out if t < 128) * 32) / 1000
tries += 1
print('Generation completed in', tries, 'tries')
print('=' * 70)
print('Done!')
print('=' * 70)
#===============================================================================
print('Rendering results...')
print('=' * 70)
print('Sample INTs', out[:15])
print('=' * 70)
song_f = []
time = 0
dur = 1
vel = 90
pitch = 60
channel = 0
pat = melody_patch
patches = [0] * 16
patches[0] = melody_patch
for ss in out:
if 0 <= ss < 128:
time += ss * 32
if 128 < ss < 256:
dur = (ss-128) * 32
if 256 < ss < 512:
dptc = (ss-256-128)
pitch += dptc
song_f.append(['note', time, dur, channel, pitch, vel, pat])
fn1 = "Delta-Melody-Transformer-Composition"
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
output_signature = 'Delta Melody 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,
output_for_gradio=True
)
print('Done!')
print('=' * 70)
#========================================================
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('-' * 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_audio, output_plot, output_midi
#==================================================================================
PDT = timezone('US/Pacific')
print('=' * 70)
print('App start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
print('=' * 70)
#==================================================================================
with gr.Blocks() as demo:
#==================================================================================
gr.Markdown("Delta Melody Transformer
")
gr.Markdown("Generate unique coherent melodies with transformers
")
gr.HTML("""
for faster execution and endless generation!
""")
#==================================================================================
gr.Markdown("## Generation options")
melody_len = gr.Slider(24, 64, value=32, step=1, label="Desired melody length")
melody_start_pitch = gr.Slider(48, 84, value=66, step=1, label="Desired melody start MIDI pitch")
melody_min_run_time = gr.Slider(0, 20, value=0, step=1, label="Minimum melody run time in seconds")
melody_patch = gr.Slider(0, 127, value=40, step=1, label="Desired melody MIDI patch")
model_temperature = gr.Slider(0.1, 1, value=0.9, step=0.01, label="Model temperature")
model_sampling_top_k = gr.Slider(1, 100, value=5, step=1, label="Model sampling top k value")
generate_btn = gr.Button("Generate", variant="primary")
gr.Markdown("## Generation results")
output_audio = gr.Audio(label="MIDI audio", format="wav", elem_id="midi_audio")
output_plot = gr.Plot(label="MIDI score plot")
output_midi = gr.File(label="MIDI file", file_types=[".mid"])
generate_btn.click(Generate_Melody,
[melody_len,
melody_start_pitch,
melody_min_run_time,
melody_patch,
model_temperature,
model_sampling_top_k
],
[output_audio,
output_plot,
output_midi,
]
)
#==================================================================================
demo.launch()
#==================================================================================