projectlosangeles's picture
Update app.py
4f2736c verified
raw
history blame
12 kB
#============================================================================================
# https://huggingface.co/spaces/projectlosangeles/Godzilla-Piano-Chords-Texturing-Transformer
#============================================================================================
print('=' * 70)
print('Godzilla Piano Chords Texturing Transformer Gradio App')
print('=' * 70)
print('Loading core Godzilla Piano Chords Texturing Transformer modules...')
import os
import copy
import time as reqtime
import datetime
from pytz import timezone
print('=' * 70)
print('Loading main Godzilla Piano Chords Texturing 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_flash_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 Godzilla Piano Chords Texturing 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 = 'Godzilla_Piano_Chords_Texturing_Trained_Model_18001_steps_0.8099_loss_0.7677_acc.pth'
SOUDFONT_PATH = 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'
MAX_MELODY_NOTES = 64
MAX_GEN_TOKS = 3072
#==================================================================================
print('=' * 70)
print('Instantiating model...')
device_type = 'cuda'
dtype = 'bfloat16'
ptdtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)
SEQ_LEN = 1536
PAD_IDX = 708
model = TransformerWrapper(num_tokens = PAD_IDX+1,
max_seq_len = SEQ_LEN,
attn_layers = Decoder(dim = 2048,
depth = 8,
heads = 32,
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/Godzilla-Piano-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 load_midi(input_midi):
raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
sp_escore_notes = TMIDIX.solo_piano_escore_notes(escore_notes)
zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
escore = TMIDIX.augment_enhanced_score_notes(zscore, timings_divider=32)
escore = TMIDIX.fix_escore_notes_durations(escore)
cscore = TMIDIX.chordify_score([1000, escore])
score = []
chords = []
pc = cscore[0]
for c in cscore:
tones_chord = sorted(set([e[4] % 12 for e in c]))
if tones_chord not in TMIDIX.ALL_CHORDS_SORTED:
tones_chord = TMIDIX.check_and_fix_tones_chord(tones_chord, use_full_chords=False)
chord_tok = TMIDIX.ALL_CHORDS_SORTED.index(tones_chord)
chords.append(chord_tok+384)
score.append(chord_tok+384)
score.append(max(0, min(127, c[0][1]-pc[0][1])))
for n in c:
score.extend([max(1, min(127, n[2]))+128, max(1, min(127, n[4]))+256])
pc = c
print('Done!')
print('=' * 70)
print('Score has', len(chords), 'chords')
print('Score hss', len(score), 'tokens')
print('=' * 70)
return score, chords
#==================================================================================
@spaces.GPU
def Generate_Chords_Textures(input_midi,
input_melody,
melody_patch,
use_nth_note,
model_temperature,
model_sampling_top_p
):
#===============================================================================
print('=' * 70)
print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
start_time = reqtime.time()
print('=' * 70)
print('=' * 70)
print('Requested settings:')
print('=' * 70)
fn = os.path.basename(input_midi)
fn1 = fn.split('.')[0]
print('Input MIDI file name:', fn)
print('Source melody patch:', melody_patch)
print('Use nth melody note:', use_nth_note)
print('Model temperature:', model_temperature)
print('Model top p:', model_sampling_top_p)
print('=' * 70)
#==================================================================
print('Loading MIDI...')
score, chords = load_midi(input_midi)
print('Sample score chords', chords[:10])
print('Sample score tokens', score[:10])
#==================================================================
print('=' * 70)
print('Generating...')
x = torch.LongTensor([705] + chords[:128] + [706]).cuda()
with ctx:
out = model.generate(x,
1024,
temperature=model_temperature,
filter_logits_fn=top_p,
filter_kwargs={'thres': model_sampling_top_p},
return_prime=False,
eos_token=707,
verbose=False)
final_song = out.tolist()
#==================================================================
print('=' * 70)
print('Done!')
print('=' * 70)
#===============================================================================
print('Rendering results...')
print('=' * 70)
print('Sample INTs', final_song[:15])
print('=' * 70)
song_f = []
if len(final_song) != 0:
time = 0
dur = 1
vel = 90
pitch = 60
channel = 0
patch = 0
patches = [0] * 16
for m in song:
if 0 <= m < 128:
time += m * 32
elif 128 < m < 256:
dur = (m-128) * 32
elif 256 < m < 384:
pitch = (m-256)
song_f.append(['note', time, dur, 0, pitch, max(40, pitch), 0])
fn1 = "Godzilla-Piano-Chords-Texturing-Transformer-Composition"
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
output_signature = 'Godzilla Piano Chords Texturing Transformer',
output_file_name = fn1,
track_name='Project Los Angeles',
list_of_MIDI_patches=patches_map
)
new_fn = fn1+'.mid'
audio = midi_to_colab_audio(new_fn,
soundfont_path=SOUDFONT_PATH,
sample_rate=16000,
volume_scale=10,
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('Output MIDI file name:', output_midi)
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_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("<h1 style='text-align: center; margin-bottom: 1rem'>Godzilla Piano Chords Texturing Transformer</h1>")
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Solo Piano chords Texturing Transformer music transformer</h1>")
gr.HTML("""
<p>
<a href="https://huggingface.co/spaces/projectlosangeles/Godzilla-Piano-Chords-Texturing-Transformer?duplicate=true">
<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-md.svg" alt="Duplicate in Hugging Face">
</a>
</p>
for faster execution and endless generation!
""")
#==================================================================================
gr.Markdown("## Upload source melody MIDI or enter a search query for a sample melody below")
gr.Markdown("### PLEASE NOTE: The demo is limited and will only texture first 128 chords of the MIDI file")
input_midi = gr.File(label="Input MIDI",
file_types=[".midi", ".mid", ".kar"]
)
gr.Markdown("## Generation options")
model_temperature = gr.Slider(0.1, 1, value=0.9, step=0.01, label="Model temperature")
model_sampling_top_p = gr.Slider(0.1, 0.99, value=0.96, step=0.01, label="Model sampling top p value")
generate_btn = gr.Button("Generate", variant="primary")
gr.Markdown("## Generation results")
output_title = gr.Textbox(label="MIDI melody title")
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_Chords_Textures,
[input_midi,
model_temperature,
model_sampling_top_p
],
[output_audio,
output_plot,
output_midi
]
)
gr.Examples(
[["Sharing The Night Together.kar", 0.9, 0.96]
],
[input_midi,
model_temperature,
model_sampling_top_p
],
[output_audio,
output_plot,
output_midi
],
Generate_Chords_Textures
)
#==================================================================================
demo.launch()
#==================================================================================