app.py

import argparse
import glob
import os.path

import gradio as gr

import pickle
import tqdm
import json

import MIDI
from midi_synthesizer import synthesis

import copy
from collections import Counter
import random
import statistics

import matplotlib.pyplot as plt

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

in_space = os.getenv("SYSTEM") == "spaces"

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

def run(midi, progress=gr.Progress()):

    mid_seq = MIDI.midi2score(midi)
    
    for m in progress.tqdm(meta_data):
        mid_seq = m[1][17:-1]
        mid_seq_ticks = m[1][16][1]
        mdata = m[1][:16]
        break

    x = []
    y = []
    c = []
    
    colors = ['red', 'yellow', 'green', 'cyan',
            'blue', 'pink', 'orange', 'purple',
            'gray', 'white', 'gold', 'silver',
            'lightgreen', 'indigo', 'maroon', 'turquoise']
    
    for s in [m for m in mid_seq if m[0] == 'note']:
        x.append(s[1])
        y.append(s[4])
        c.append(colors[s[3]])

    plt.close()
    plt.figure(figsize=(14,5))
    ax=plt.axes(title='MIDI Search Plot')
    ax.set_facecolor('black')
    
    plt.scatter(x,y, c=c)
    plt.xlabel("Time")
    plt.ylabel("Pitch")
 
    with open(f"output.mid", 'wb') as f:
        f.write(MIDI.score2midi([mid_seq_ticks, mid_seq]))
    audio = synthesis(MIDI.score2opus([mid_seq_ticks, mid_seq]), soundfont_path)
    yield mdata, "output.mid", (44100, audio), plt

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

def match_midi(midi, progress=gr.Progress()):
  
    print('=' * 70)
    print('Loading MIDI file...')
    
    #==================================================
    
    score = MIDI.midi2score(midi)
    
    events_matrix = []
    
    track_count = 0
    
    for s in score:
    
        if track_count > 0:
            track = s
            track.sort(key=lambda x: x[1])
            events_matrix.extend(track)
        else:
            midi_ticks = s
    
        track_count += 1
    
    events_matrix.sort(key=lambda x: x[1])
    
    mult_pitches_counts = []
    
    for i in range(-6, 6):
    
      events_matrix1 = []
    
      for e in events_matrix:
    
        ev = copy.deepcopy(e)
    
        if e[0] == 'note':
            if e[3] == 9:
                ev[4] = ((e[4] % 128) + 128)
            else:
              ev[4] = ((e[4] % 128) + i)
    
            events_matrix1.append(ev)
    
      pitches_counts = [[y[0],y[1]] for y in Counter([y[4] for y in events_matrix1 if y[0] == 'note']).most_common()]
      pitches_counts.sort(key=lambda x: x[0], reverse=True)
    
      mult_pitches_counts.append(pitches_counts)
    
    patches_list = sorted(list(set([y[3] for y in events_matrix if y[0] == 'patch_change'])))
    
    
    #==================================================
    
    ms_score = MIDI.midi2ms_score(midi)
    
    ms_events_matrix = []
    
    itrack1 = 1
    
    while itrack1 < len(ms_score):
        for event in ms_score[itrack1]:
            if event[0] == 'note':
                ms_events_matrix.append(event)
        itrack1 += 1
    
    ms_events_matrix.sort(key=lambda x: x[1])
    
    
    chords = []
    pe = ms_events_matrix[0]
    cho = []
    for e in ms_events_matrix:
        if (e[1] - pe[1]) == 0:
          if e[3] != 9:
            if (e[4] % 12) not in cho:
              cho.append(e[4] % 12)
        else:
          if len(cho) > 0:
            chords.append(sorted(cho))
          cho = []
          if e[3] != 9:
            if (e[4] % 12) not in cho:
              cho.append(e[4] % 12)
    
        pe = e
    
    if len(cho) > 0:
        chords.append(sorted(cho))
    
    ms_chords_counts = sorted([[list(key), val] for key,val in Counter([tuple(c) for c in chords if len(c) > 1]).most_common()], reverse=True, key = lambda x: x[1])
    
    times = []
    pt = ms_events_matrix[0][1]
    start = True
    for e in ms_events_matrix:
        if (e[1]-pt) != 0 or start == True:
            times.append((e[1]-pt))
            start = False
        pt = e[1]
    
    durs = [e[2] for e in ms_events_matrix]
    vels = [e[5] for e in ms_events_matrix]
    
    avg_time = int(sum(times) / len(times))
    avg_dur = int(sum(durs) / len(durs))
    
    mode_time = statistics.mode(times)
    mode_dur = statistics.mode(durs)
    
    median_time = int(statistics.median(times))
    median_dur = int(statistics.median(durs))
    
    #==================================================
    
    print('=' * 70)
    print('Done!')
    print('=' * 70)
    
    #==========================================================================================================

    #@title MIDI Pitches Search

    #@markdown Match ratio control option
    
    maximum_match_ratio_to_search_for = 1 #@param {type:"slider", min:0, max:1, step:0.01}
    
    #@markdown MIDI pitches search options
    
    pitches_counts_cutoff_threshold_ratio = 0 #@param {type:"slider", min:0, max:1, step:0.05}
    search_transposed_pitches = False #@param {type:"boolean"}
    skip_exact_matches = True #@param {type:"boolean"}
    
    #@markdown Additional search options
    
    add_pitches_counts_ratios = False #@param {type:"boolean"}
    add_timings_ratios = False #@param {type:"boolean"}
    add_durations_ratios = False #@param {type:"boolean"}
    
    print('=' * 70)
    print('MIDI Pitches Search')
    print('=' * 70)
    
    final_ratios = []
    
    for d in progress.tqdm(meta_data):
    
      try:
        p_counts = d[1][10][1]
        p_counts.sort(reverse = True, key = lambda x: x[1])
        max_p_count = p_counts[0][1]
        trimmed_p_counts = [y for y in p_counts if y[1] >= (max_p_count * pitches_counts_cutoff_threshold_ratio)]
        total_p_counts = sum([y[1] for y in trimmed_p_counts])
    
        if search_transposed_pitches:
          search_pitches = mult_pitches_counts
        else:
          search_pitches = [mult_pitches_counts[6]]
    
        #===================================================
    
        ratios_list = []
    
        #===================================================
    
        atrat = [0]
    
        if add_timings_ratios:
    
          source_times = [avg_time,
                          median_time,
                          mode_time]
    
          match_times = meta_data[0][1][3][1]
    
          times_ratios = []
    
          for i in range(len(source_times)):
            maxtratio = max(source_times[i], match_times[i])
            mintratio = min(source_times[i], match_times[i])
            times_ratios.append(mintratio / maxtratio)
    
          avg_times_ratio = sum(times_ratios) / len(times_ratios)
    
          atrat[0] = avg_times_ratio
    
        #===================================================
    
        adrat = [0]
    
        if add_durations_ratios:
    
          source_durs = [avg_dur,
                          median_dur,
                          mode_dur]
    
          match_durs = meta_data[0][1][4][1]
    
          durs_ratios = []
    
          for i in range(len(source_durs)):
            maxtratio = max(source_durs[i], match_durs[i])
            mintratio = min(source_durs[i], match_durs[i])
            durs_ratios.append(mintratio / maxtratio)
    
          avg_durs_ratio = sum(durs_ratios) / len(durs_ratios)
    
          adrat[0] = avg_durs_ratio
    
        #===================================================
    
        for m in search_pitches:
    
          sprat = []
    
          m.sort(reverse = True, key = lambda x: x[1])
          max_pitches_count = m[0][1]
          trimmed_pitches_counts = [y for y in m if y[1] >= (max_pitches_count * pitches_counts_cutoff_threshold_ratio)]
          total_pitches_counts = sum([y[1] for y in trimmed_pitches_counts])
    
          same_pitches = set([T[0] for T in trimmed_p_counts]) & set([m[0] for m in trimmed_pitches_counts])
          num_same_pitches = len(same_pitches)
    
          if num_same_pitches == len(trimmed_pitches_counts):
            same_pitches_ratio = (num_same_pitches / len(trimmed_p_counts))
          else:
            same_pitches_ratio = (num_same_pitches / max(len(trimmed_p_counts), len(trimmed_pitches_counts)))
    
          if skip_exact_matches:
            if same_pitches_ratio == 1:
              same_pitches_ratio = 0
    
          sprat.append(same_pitches_ratio)
    
          #===================================================
    
          spcrat = [0]
    
          if add_pitches_counts_ratios:
    
            same_trimmed_p_counts = sorted([T for T in trimmed_p_counts if T[0] in same_pitches], reverse = True)
            same_trimmed_pitches_counts = sorted([T for T in trimmed_pitches_counts if T[0] in same_pitches], reverse = True)
    
            same_trimmed_p_counts_ratios = [[s[0], s[1] / total_p_counts] for s in same_trimmed_p_counts]
            same_trimmed_pitches_counts_ratios = [[s[0], s[1] / total_pitches_counts] for s in same_trimmed_pitches_counts]
    
            same_pitches_counts_ratios = []
    
            for i in range(len(same_trimmed_p_counts_ratios)):
              mincratio = min(same_trimmed_p_counts_ratios[i][1], same_trimmed_pitches_counts_ratios[i][1])
              maxcratio = max(same_trimmed_p_counts_ratios[i][1], same_trimmed_pitches_counts_ratios[i][1])
              same_pitches_counts_ratios.append([same_trimmed_p_counts_ratios[i][0], mincratio / maxcratio])
    
            same_counts_ratios = [s[1] for s in same_pitches_counts_ratios]
    
            if len(same_counts_ratios) > 0:
              avg_same_pitches_counts_ratio = sum(same_counts_ratios) / len(same_counts_ratios)
            else:
              avg_same_pitches_counts_ratio = 0
    
            spcrat[0] = avg_same_pitches_counts_ratio
    
          #===================================================
    
          r_list = [sprat[0]]
    
          if add_pitches_counts_ratios:
            r_list.append(spcrat[0])
    
          if add_timings_ratios:
            r_list.append(atrat[0])
    
          if add_durations_ratios:
            r_list.append(adrat[0])
    
          ratios_list.append(r_list)
    
        #===================================================
    
        avg_ratios_list = []
    
        for r in ratios_list:
          avg_ratios_list.append(sum(r) / len(r))
    
        #===================================================
    
        final_ratio = max(avg_ratios_list)
    
        if final_ratio > maximum_match_ratio_to_search_for:
            final_ratio = 0
    
        final_ratios.append(final_ratio)
    
        #===================================================
    
    max_ratio = max(final_ratios)
    max_ratio_index = final_ratios.index(max_ratio)
    
    print('FOUND')
    print('=' * 70)
    print('Match ratio', max_ratio)
    print('MIDI file name', meta_data[max_ratio_index][0])
    print('=' * 70)
    
    fn = meta_data[max_ratio_index][0]

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

    md = meta_data[max_ratio_index]

    mid_seq = md[1][17:-1]
    mid_seq_ticks = md[1][16][1]
    mdata = md[1][:16]

    x = []
    y = []
    c = []
    
    colors = ['red', 'yellow', 'green', 'cyan',
            'blue', 'pink', 'orange', 'purple',
            'gray', 'white', 'gold', 'silver',
            'lightgreen', 'indigo', 'maroon', 'turquoise']
    
    for s in [m for m in mid_seq if m[0] == 'note']:
        x.append(s[1])
        y.append(s[4])
        c.append(colors[s[3]])

    plt.close()
    plt.figure(figsize=(14,5))
    ax=plt.axes(title='MIDI Search Plot')
    ax.set_facecolor('black')
    
    plt.scatter(x,y, c=c)
    plt.xlabel("Time")
    plt.ylabel("Pitch")
 
    with open(f"output.mid", 'wb') as f:
        f.write(MIDI.score2midi([mid_seq_ticks, mid_seq]))
    audio = synthesis(MIDI.score2opus([mid_seq_ticks, mid_seq]), soundfont_path)
    yield mdata, "output.mid", (44100, audio), plt

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

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
    parser.add_argument("--port", type=int, default=7860, help="gradio server port")
    parser.add_argument("--max-gen", type=int, default=1024, help="max")
    
    opt = parser.parse_args()
    
    soundfont_path = "SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2"
    meta_data_path = "meta-data/LAMD_META_10000.pickle"

    print('Loading meta-data...')
    with open(meta_data_path, 'rb') as f:
        meta_data = pickle.load(f)
    print('Done!')
    
    app = gr.Blocks()
    with app:
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>MIDI Match</h1>")
        gr.Markdown("![Visitors](https://api.visitorbadge.io/api/visitors?path=asigalov61.MIDI-Match&style=flat)\n\n"
                    "MIDI Match\n\n"
                    "Demo for [MIDI Match](https://github.com/asigalov61)\n\n"
                    "[Open In Colab]"
                    "(https://colab.research.google.com/github/asigalov61/MIDI-Match/blob/main/demo.ipynb)"
                    " for faster running and longer generation"
                    )
        

        input_midi = gr.File(label="input midi", file_types=[".midi", ".mid"], type="binary")
        
        output_audio = gr.Audio(label="output audio", format="mp3", elem_id="midi_audio")
        output_midi = gr.File(label="output midi", file_types=[".mid"])
        output_midi_seq = gr.Textbox(label="output midi metadata")
        output_plot = gr.Plot(label="output midi plot")
        
        run_event = input_midi.upload(match_midi, [input_midi],
                                                  [output_midi_seq, output_midi, output_audio, output_plot])
        
    app.queue(1).launch(server_port=opt.port, share=opt.share, inbrowser=True)