app.py

# 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 classify_GPU(input_data):

    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)

    #==================================================================

    number_of_batches = 1 # @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,
                              filter_logits_fn=top_k,
                              filter_kwargs={'k': 1},
                              return_prime=False,
                              verbose=False)
    
        y = out.tolist()
    
        output = [l[0] for l in y]
        result = mode(output)
    
        results.append(result)

    return output, results
                       
# =================================================================================================

def ClassifyMIDI(input_midi):

    SEQ_LEN = 1024
    PAD_IDX = 14627
    
    print('=' * 70)
    print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    start_time = reqtime.time()

    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) // 4):
      schunk = seq[i:i+SEQ_LEN-4]
      input_data.append([14624] + schunk + [14625])
    
    print('Done!')
    print('=' * 70)
    
    #==============================================================

    classification_summary_string = '=' * 70
    classification_summary_string += '\n'
    
    print('Composition has', notes_counter, 'notes')
    print('=' * 70)
    print('Composition was split into' , len(input_data), 'chunks of 340 notes each with 255 notes overlap')
    print('Number of notes in all composition chunks:', len(input_data) * 340)

    classification_summary_string += 'Composition has ' + str(notes_counter) + ' notes\n'
    classification_summary_string += '=' * 70
    classification_summary_string += '\n'
    classification_summary_string += 'Composition was split into ' + str(len(input_data)) + ' chunks of 340 notes each with 170 notes overlap\n'
    classification_summary_string += 'Number of notes in all composition chunks: ' + str(len(input_data) * 340) + '\n'
    classification_summary_string += '=' * 70
    classification_summary_string += '\n'

    output, results = classify_GPU(input_data)
    
    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)

    classification_summary_string += 'Most common classification label: ' + str(classifier_labels[0][final_result-384]) + '\n'
    classification_summary_string += 'Most common classification label ratio: ' + str(results.count(final_result) / len(results)) + '\n'
    classification_summary_string += 'Most common classification label index '+ str(final_result) + '\n'
    classification_summary_string += '=' * 70
    classification_summary_string += '\n'
    
    print('All classification labels summary:')
    print('=' * 70)
    
    for i, a in enumerate(all_results_labels):
        print('Notes', i*85, '-', (i*85)+340, '===', a)
        classification_summary_string += 'Notes ' + str(i*85) + ' - ' + str((i*85)+340) + ' === ' + str(a) + '\n'
    
    classification_summary_string += '=' * 70
    classification_summary_string += '\n'
    
    print('=' * 70)
    print('Done!')
    print('=' * 70)
    
    #===============================================================================
    print('Rendering results...')

    score_idx = processed_scores_labels.index(classifier_labels[0][final_result-384])

    output_score = processed_scores[score_idx][1][:6000]
    
    print('=' * 70)
    print('Sample INTs', results[:15])
    print('=' * 70)
    
    fn1 = processed_scores[score_idx][0]

    output_score = TMIDIX.recalculate_score_timings(output_score)

    output_score, patches, overflow_patches = TMIDIX.patch_enhanced_score_notes(output_score)
    
    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(output_score,
                                                              output_signature = 'Advanced MIDI Classifier',
                                                              output_file_name = fn1,
                                                              track_name='Project Los Angeles',
                                                              list_of_MIDI_patches=patches,
                                                              timings_multiplier=16
                                                              )
    
    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 = classification_summary_string
    output_midi = str(new_fn)
    output_audio = (16000, audio)
    
    output_plot = TMIDIX.plot_ms_SONG(output_score, plot_title=output_midi, return_plt=True, timings_multiplier=16)

    print('Output MIDI file name:', output_midi)
    print('Output MIDI title:', output_midi_title)
    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')
    processed_scores_labels = [l[0] for l in processed_scores]
    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('=' * 70)

    app = gr.Blocks()
    with app:
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Advanced MIDI Classifier</h1>")
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Detailed MIDI classification with transformers</h1>")
        gr.Markdown(
            "![Visitors](https://api.visitorbadge.io/api/visitors?path=asigalov61.Advanced-MIDI-Classifier&style=flat)\n\n"
            "This is a demo for Annotated MIDI Dataset\n\n"
            "Check out [Annotated MIDI Dataset](https://huggingface.co/datasets/asigalov61/Annotated-MIDI-Dataset) on Hugging Face!\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="Best classification match MIDI title")
        output_midi_summary = gr.Textbox(label="MIDI classification summary")
        output_audio = gr.Audio(label="Best classification match MIDI audio", format="wav", elem_id="midi_audio")
        output_plot = gr.Plot(label="Best classification match MIDI score plot")
        output_midi = gr.File(label="Best classification match 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()