# https://huggingface.co/spaces/asigalov61/Advanced-MIDI-Classifier
import os
import time as reqtime
import datetime
from pytz import timezone
import torch
import spaces
import gradio as gr
from x_transformer_1_23_2 import *
import random
from statistics import mode
import tqdm
from midi_to_colab_audio import midi_to_colab_audio
import TMIDIX
import matplotlib.pyplot as plt
in_space = os.getenv("SYSTEM") == "spaces"
# =================================================================================================
@spaces.GPU
def ClassifyMIDI(input_midi):
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 = 1024
PAD_IDX = 14627
DEVICE = 'cuda' # 'cuda'
# instantiate the model
model = TransformerWrapper(
num_tokens = PAD_IDX+1,
max_seq_len = SEQ_LEN,
attn_layers = Decoder(dim = 1024, depth = 12, 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('Annotated_MIDI_Dataset_Classifier_Trained_Model_21269_steps_0.4335_loss_0.8716_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)
#==================================================================
print('=' * 70)
fn = os.path.basename(input_midi.name)
fn1 = fn.split('.')[0]
print('-' * 70)
print('Input file name:', fn)
print('=' * 70)
print('Loading MIDI file...')
midi_name = fn
raw_score = TMIDIX.midi2single_track_ms_score(open(input_midi.name, 'rb').read())
escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
escore = [e for e in TMIDIX.augment_enhanced_score_notes(escore_notes, timings_divider=32) if e[6] < 80]
cscore = TMIDIX.chordify_score([1000, escore])
#=======================================================
# MAIN PROCESSING CYCLE
#=======================================================
melody_chords = []
pe = cscore[0][0]
for c in cscore:
pitches = []
for e in c:
if e[4] not in pitches:
dtime = max(0, min(127, e[1]-pe[1]))
dur = max(1, min(127, e[2]))
ptc = max(1, min(127, e[4]))
melody_chords.append([dtime, dur, ptc])
pitches.append(ptc)
pe = e
#==============================================================
seq = []
input_data = []
notes_counter = 0
for mm in melody_chords:
time = mm[0]
dur = mm[1]
ptc = mm[2]
seq.extend([time, dur+128, ptc+256])
notes_counter += 1
for i in range(0, len(seq)-SEQ_LEN-4, (SEQ_LEN-4) // 2):
schunk = seq[i:i+SEQ_LEN-4]
input_data.append([14624] + schunk + [14625])
print('Done!')
print('=' * 70)
#==============================================================
print('Composition has', notes_counter, 'notes')
print('=' * 70)
print('Composition was split into' , len(input_data), 'chunks', 'of 340 notes each with 170 notes overlap')
print('Number of notes in all composition chunks:', len(input_data) * 340)
number_of_batches = 100 # @param {type:"slider", min:1, max:100, step:1}
# @markdown NOTE: You can increase the number of batches on high-ram GPUs for better classification
print('=' * 70)
print('Annotated MIDI Dataset Classifier')
print('=' * 70)
print('Classifying...')
torch.cuda.empty_cache()
model.eval()
results = []
for input in input_data:
x = torch.tensor([input[:1022]] * number_of_batches, dtype=torch.long, device='cuda')
with ctx:
out = model.generate(x,
1,
temperature=0.3,
return_prime=False,
verbose=False)
y = out.tolist()
output = [l[0] for l in y]
result = mode(output)
results.append(result)
all_results_labels = [classifier_labels[0][r-384] for r in results]
final_result = mode(results)
print('Done!')
print('=' * 70)
print('Most common classification label:', classifier_labels[0][final_result-384])
print('Most common classification label ratio:' , results.count(final_result) / len(results))
print('Most common classification label index', final_result)
print('=' * 70)
print('All classification labels summary:')
print('=' * 70)
for i, a in enumerate(all_results_labels):
print('Notes', i*170, '-', (i*170)+340, '===', a)
print('=' * 70)
print('Done!')
print('=' * 70)
#===============================================================================
print('Rendering results...')
print('=' * 70)
print('Sample INTs', output[:15])
print('=' * 70)
out1 = output
fn1 = "Melody2Song-Seq2Seq-Music-Transformer-Composition"
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
output_signature = 'Advanced MIDI Classifier',
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"
print('Loading Annotated MIDI Dataset processed scores...')
processed_scores = TMIDIX.Tegridy_Any_Pickle_File_Reader('processed_scores')
print(processed_scores[0][0])
print('=' * 70)
print('Loading Annotated MIDI Dataset Classifier Songs Artists Labels...')
classifier_labels = TMIDIX.Tegridy_Any_Pickle_File_Reader('Annotated_MIDI_Dataset_Classifier_Songs_Artists_Labels')
print(classifier_labels[0])
print('=' * 70)
app = gr.Blocks()
with app:
gr.Markdown("Advanced MIDI Classifier
")
gr.Markdown("Detailed MIDI classification with transformers
")
gr.Markdown(
"\n\n")
input_midi = gr.File(label="Input MIDI", file_types=[".midi", ".mid", ".kar"])
run_btn = gr.Button("classify", variant="primary")
gr.Markdown("## Classification 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(ClassifyMIDI, [input_midi],
[output_midi_title, output_midi_summary, output_midi, output_audio, output_plot])
app.queue().launch()