app.py

import os.path

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
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 GenerateAccompaniment(input_midi, input_num_tokens, input_conditioning_type, input_strip_notes):
    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 = 8192 # Models seq len
    PAD_IDX = 707 # Models pad index
    DEVICE = 'cuda' # 'cuda'

    # instantiate the model

    model = TransformerWrapper(
        num_tokens = PAD_IDX+1,
        max_seq_len = SEQ_LEN,
        attn_layers = Decoder(dim = 2048, depth = 4, 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('Chords_Progressions_Transformer_Small_2048_Trained_Model_12947_steps_0.9316_loss_0.7386_acc.pth',
                   map_location=DEVICE))
    print('=' * 70)

    model.eval()

    if DEVICE == 'cpu':
        dtype = torch.bfloat16
    else:
        dtype = torch.float16

    ctx = torch.amp.autocast(device_type=DEVICE, dtype=dtype)

    print('Done!')
    print('=' * 70)

    fn = os.path.basename(input_midi.name)
    fn1 = fn.split('.')[0]

    input_num_tokens = max(4, min(128, input_num_tokens))

    print('-' * 70)
    print('Input file name:', fn)
    print('Req num toks:', input_num_tokens)
    print('Conditioning type:', input_conditioning_type)
    print('Strip notes:', input_strip_notes)
    print('-' * 70)

    #===============================================================================
    raw_score = TMIDIX.midi2single_track_ms_score(input_midi.name)
    
    #===============================================================================
    # Enhanced score notes
    
    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
    
    no_drums_escore_notes = [e for e in escore_notes if e[6] < 80]
    
    if len(no_drums_escore_notes) > 0:
    
        #=======================================================
        # PRE-PROCESSING
        
        #===============================================================================
        # Augmented enhanced score notes
        
        no_drums_escore_notes = TMIDIX.augment_enhanced_score_notes(no_drums_escore_notes)
        
        cscore = TMIDIX.chordify_score([1000, no_drums_escore_notes])
        
        clean_cscore = []
        
        for c in cscore:
            pitches = []
            cho = []
            for cc in c:
                if cc[4] not in pitches:
                    cho.append(cc)
                    pitches.append(cc[4])
        
            clean_cscore.append(cho)
        
        #=======================================================
        # FINAL PROCESSING
        
        melody_chords = []
        chords = []
        times = [0]
        durs = []
        
        #=======================================================
        # MAIN PROCESSING CYCLE
        #=======================================================
        
        pe = clean_cscore[0][0]
        
        first_chord = True
        
        for c in clean_cscore:
        
            # Chords
        
            c.sort(key=lambda x: x[4], reverse=True)
        
            tones_chord = sorted(set([cc[4] % 12 for cc in c]))
        
            try:
                chord_token = TMIDIX.ALL_CHORDS_SORTED.index(tones_chord)
            except:
                checked_tones_chord = TMIDIX.check_and_fix_tones_chord(tones_chord)
                chord_token = TMIDIX.ALL_CHORDS_SORTED.index(checked_tones_chord)
        
            melody_chords.extend([chord_token+384])
        
            if input_strip_notes:
              if len(tones_chord) > 1:
                chords.extend([chord_token+384])
        
            else:
              chords.extend([chord_token+384])
        
            if first_chord:
                    melody_chords.extend([0])
                    first_chord = False
        
            for e in c:
        
                #=======================================================
                # Timings...
        
                time = e[1]-pe[1]
        
                dur = e[2]
        
                if time != 0 and time % 2 != 0:
                    time += 1
                if dur % 2 != 0:
                    dur += 1
        
                delta_time = int(max(0, min(255, time)) / 2)
        
                # Durations
        
                dur = int(max(0, min(255, dur)) / 2)
        
                # Pitches
        
                ptc = max(1, min(127, e[4]))
        
                #=======================================================
                # FINAL NOTE SEQ
        
                # Writing final note asynchronously
        
                if delta_time != 0:
                    melody_chords.extend([delta_time, dur+128, ptc+256])
                    if input_strip_notes:
                      if len(c) > 1:
                        times.append(delta_time)
                        durs.append(dur+128)
                    else:
                        times.append(delta_time)
                        durs.append(dur+128)
                else:
                    melody_chords.extend([dur+128, ptc+256])
        
                pe = e
        
    #==================================================================

    print('=' * 70)
    
    print('Sample output events', melody_chords[:5])
    print('=' * 70)
    print('Generating...')

    output = []
    
    max_chords_limit = 8
    temperature=0.9
    num_memory_tokens=4096

    output = []
    
    idx = 0
    
    for c in chords[:input_num_tokens]:
     
        output.append(c)
    
        if input_conditioning_type == 'Chords-Times' or input_conditioning_type == 'Chords-Times-Durations':
          output.append(times[idx])
    
        if input_conditioning_type == 'Chords-Times-Durations':
          output.append(durs[idx])

        x = torch.tensor([output] * 1, dtype=torch.long, device='cuda')
        
        o = 0
        
        ncount = 0
        
        while o < 384 and ncount < max_chords_limit:
          with ctx:
            out = model.generate(x[-num_memory_tokens:],
                                1,
                                temperature=temperature,
                                return_prime=False,
                                verbose=False)
        
          o = out.tolist()[0][0]
        
          if 256 <= o < 384:
            ncount += 1
        
          if o < 384:
            x = torch.cat((x, out), 1)
        
        outy =  x.tolist()[0][len(output):]
          
        output.extend(outy)
    
        idx += 1

        if idx == len(chords[:input_num_tokens])-1:
            break

    print('=' * 70)
    print('Done!')
    print('=' * 70)
    
    #===============================================================================
    print('Rendering results...')
    
    print('=' * 70)
    print('Sample INTs', output[:12])
    print('=' * 70)
    
    out1 = output

    if len(out1) != 0:
    
        song = out1
        song_f = []
    
        time = 0
        dur = 0
        vel = 90
        pitch = 0
        channel = 0
        
        patches = [0] * 16
        
        channel = 0
        
        for ss in song:
        
          if 0 <= ss < 128:
        
              time += ss * 32
        
          if 128 <= ss < 256:
        
              dur = (ss-128) * 32
        
          if 256 <= ss < 384:
        
              pitch = (ss-256)
        
              vel = max(40, pitch)
        
              song_f.append(['note', time, dur, channel, pitch, vel, 0])

    fn1 = "Chords-Progressions-Transformer-Composition"
    
    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
                                                              output_signature = 'Chords Progressions 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,
                        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:', '')
    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"
   
    app = gr.Blocks()
    with app:
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Melody2Song Seq2Seq Music Transformer</h1>")
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Generate unique songs from melodies with se2seq music transformer</h1>")
        gr.Markdown(
            "![Visitors](https://api.visitorbadge.io/api/visitors?path=asigalov61.Melody2Song-Seq2Seq-Music-Transformer&style=flat)\n\n")
        
        input_midi = gr.File(label="Input MIDI", file_types=[".midi", ".mid", ".kar"])
        input_num_tokens = gr.Slider(4, 128, value=32, step=1, label="Number of composition chords to generate progression for")
        input_conditioning_type = gr.Radio(["Chords", "Chords-Times", "Chords-Times-Durations"], label="Conditioning type")
        input_strip_notes = gr.Checkbox(label="Strip notes from the composition")
        
        run_btn = gr.Button("generate", variant="primary")

        gr.Markdown("## Generation 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(GenerateAccompaniment, [input_midi, input_num_tokens, input_conditioning_type, input_strip_notes],
                                  [output_midi_title, output_midi_summary, output_midi, output_audio, output_plot])

        app.queue().launch()