allegro_music_transformer_composer.py

# -*- coding: utf-8 -*-
"""Allegro_Music_Transformer_Composer.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/github/asigalov61/Allegro-Music-Transformer/blob/main/Allegro_Music_Transformer_Composer.ipynb

# Allegro Music Transformer Composer (ver. 5.0)

***

Powered by tegridy-tools: https://github.com/asigalov61/tegridy-tools

***

WARNING: This complete implementation is a functioning model of the Artificial Intelligence. Please excercise great humility, care, and respect. https://www.nscai.gov/

***

#### Project Los Angeles

#### Tegridy Code 2023

***

# (GPU CHECK)
"""

#@title NVIDIA GPU check
!nvidia-smi

"""# (SETUP ENVIRONMENT)"""

#@title Install dependencies
!git clone --depth 1 https://github.com/asigalov61/Allegro-Music-Transformer
!pip install huggingface_hub
!pip install torch
!pip install einops
!pip install torch-summary
!pip install tqdm
!pip install matplotlib
!apt install fluidsynth #Pip does not work for some reason. Only apt works
!pip install midi2audio

# Commented out IPython magic to ensure Python compatibility.
#@title Import modules

print('=' * 70)
print('Loading core Allegro Music Transformer modules...')

import os
import pickle
import random
import secrets
import statistics
import tqdm

print('=' * 70)
print('Loading main Allegro Music Transformer modules...')
import torch

# %cd /content/Allegro-Music-Transformer

import TMIDIX
from x_transformer import *

# %cd /content/
print('=' * 70)
print('Loading aux Allegro Music Transformer modules...')

import matplotlib.pyplot as plt

from torchsummary import summary
from sklearn import metrics

from midi2audio import FluidSynth
from IPython.display import Audio, display

from huggingface_hub import hf_hub_download

from google.colab import files

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

"""# (LOAD MODEL)

# Choose one and do not forget to restart colab runtime if switching between models
"""

#@title Load Allegro Music Transformer Tiny Model (FAST)

#@markdown Very fast model, 16 layers, 225k MIDIs training corpus

full_path_to_model_checkpoint = "/content/Allegro-Music-Transformer/Models/Tiny/Allegro_Music_Transformer_Tiny_Trained_Model_80000_steps_0.9457_loss_0.7443_acc.pth" #@param {type:"string"}

#@markdown Model precision option

model_precision = "bfloat16" # @param ["bfloat16", "float16", "float32"]

#@markdown bfloat16 == Third precision/triple speed (if supported, otherwise the model will default to float16)

#@markdown float16 == Half precision/double speed

#@markdown float32 == Full precision/normal speed

plot_tokens_embeddings = True # @param {type:"boolean"}

print('=' * 70)
print('Loading Allegro Music Transformer Tiny Pre-Trained Model...')
print('Please wait...')
print('=' * 70)

if os.path.isfile(full_path_to_model_checkpoint):
  print('Model already exists...')

else:
  hf_hub_download(repo_id='asigalov61/Allegro-Music-Transformer',
                  filename='Allegro_Music_Transformer_Tiny_Trained_Model_80000_steps_0.9457_loss_0.7443_acc.pth',
                  local_dir='/content/Allegro-Music-Transformer/Models/Tiny/',
                  local_dir_use_symlinks=False)
print('=' * 70)
print('Instantiating model...')

torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda'

if model_precision == 'bfloat16' and torch.cuda.is_bf16_supported():
  dtype = 'bfloat16'
else:
  dtype = 'float16'

if model_precision == 'float16':
  dtype = 'float16'

if model_precision == 'float32':
  dtype = 'float32'

ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)

SEQ_LEN = 2048

# instantiate the model

model = TransformerWrapper(
    num_tokens = 3088,
    max_seq_len = SEQ_LEN,
    attn_layers = Decoder(dim = 1024, depth = 16, heads = 8, attn_flash=True)
)

model = AutoregressiveWrapper(model)

model = torch.nn.DataParallel(model)

model.cuda()
print('=' * 70)

print('Loading model checkpoint...')

model.load_state_dict(torch.load(full_path_to_model_checkpoint))
print('=' * 70)

model.eval()

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

print('Model will use', dtype, 'precision...')
print('=' * 70)

# Model stats
print('Model summary...')
summary(model)

# Plot Token Embeddings

if plot_tokens_embeddings:

  tok_emb = model.module.net.token_emb.emb.weight.detach().cpu().tolist()

  cos_sim = metrics.pairwise_distances(
    tok_emb, metric='cosine'
  )
  plt.figure(figsize=(7, 7))
  plt.imshow(cos_sim, cmap="inferno", interpolation="nearest")
  im_ratio = cos_sim.shape[0] / cos_sim.shape[1]
  plt.colorbar(fraction=0.046 * im_ratio, pad=0.04)
  plt.xlabel("Position")
  plt.ylabel("Position")
  plt.tight_layout()
  plt.plot()
  plt.savefig("/content/Allegro-Music-Transformer-Tiny-Tokens-Embeddings-Plot.png", bbox_inches="tight")

#@title Load Allegro Music Transformer Small Model (BEST)

#@markdown Fast model, 32 layers, 225k MIDIs training corpus

full_path_to_model_checkpoint = "/content/Allegro-Music-Transformer/Models/Small/Allegro_Music_Transformer_Small_Trained_Model_56000_steps_0.9399_loss_0.7374_acc.pth" #@param {type:"string"}

#@markdown Model precision option

model_precision = "bfloat16" # @param ["bfloat16", "float16", "float32"]

#@markdown bfloat16 == Third precision/triple speed (if supported, otherwise the model will default to float16)

#@markdown float16 == Half precision/double speed

#@markdown float32 == Full precision/normal speed

plot_tokens_embeddings = True # @param {type:"boolean"}

print('=' * 70)
print('Loading Allegro Music Transformer Small Pre-Trained Model...')
print('Please wait...')
print('=' * 70)

if os.path.isfile(full_path_to_model_checkpoint):
  print('Model already exists...')

else:
  hf_hub_download(repo_id='asigalov61/Allegro-Music-Transformer',
                  filename='Allegro_Music_Transformer_Small_Trained_Model_56000_steps_0.9399_loss_0.7374_acc.pth',
                  local_dir='/content/Allegro-Music-Transformer/Models/Small/',
                  local_dir_use_symlinks=False)
print('=' * 70)
print('Instantiating model...')

torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
device_type = 'cuda'

if model_precision == 'bfloat16' and torch.cuda.is_bf16_supported():
  dtype = 'bfloat16'
else:
  dtype = 'float16'

if model_precision == 'float16':
  dtype = 'float16'

if model_precision == 'float32':
  dtype = 'float32'

ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)

SEQ_LEN = 2048

# instantiate the model

model = TransformerWrapper(
    num_tokens = 3088,
    max_seq_len = SEQ_LEN,
    attn_layers = Decoder(dim = 1024, depth = 32, heads = 8, attn_flash=True)
)

model = AutoregressiveWrapper(model)

model = torch.nn.DataParallel(model)

model.cuda()
print('=' * 70)

print('Loading model checkpoint...')

model.load_state_dict(torch.load(full_path_to_model_checkpoint))
print('=' * 70)

model.eval()

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

print('Model will use', dtype, 'precision...')
print('=' * 70)

# Model stats
print('Model summary...')
summary(model)

# Plot Token Embeddings

if plot_tokens_embeddings:

  tok_emb = model.module.net.token_emb.emb.weight.detach().cpu().tolist()

  cos_sim = metrics.pairwise_distances(
    tok_emb, metric='cosine'
  )
  plt.figure(figsize=(7, 7))
  plt.imshow(cos_sim, cmap="inferno", interpolation="nearest")
  im_ratio = cos_sim.shape[0] / cos_sim.shape[1]
  plt.colorbar(fraction=0.046 * im_ratio, pad=0.04)
  plt.xlabel("Position")
  plt.ylabel("Position")
  plt.tight_layout()
  plt.plot()
  plt.savefig("/content/Allegro-Music-Transformer-Small-Tokens-Embeddings-Plot.png", bbox_inches="tight")

"""# (LOAD SEED MIDI)"""

#@title Load Seed MIDI

#@markdown Press play button to to upload your own seed MIDI or to load one of the provided sample seed MIDIs from the dropdown list below

select_seed_MIDI = "Upload your own custom MIDI" #@param ["Upload your own custom MIDI", "Allegro-Music-Transformer-Piano-Seed-1", "Allegro-Music-Transformer-Piano-Seed-2", "Allegro-Music-Transformer-Piano-Seed-3", "Allegro-Music-Transformer-Piano-Seed-4", "Allegro-Music-Transformer-Piano-Seed-5", "Allegro-Music-Transformer-MI-Seed-1", "Allegro-Music-Transformer-MI-Seed-2", "Allegro-Music-Transformer-MI-Seed-3", "Allegro-Music-Transformer-MI-Seed-4", "Allegro-Music-Transformer-MI-Seed-5"]
number_of_prime_tokens = 300 # @param {type:"slider", min:128, max:5000, step:16}
render_MIDI_to_audio = False # @param {type:"boolean"}

print('=' * 70)
print('Allegro Music Transformer Seed MIDI Loader')
print('=' * 70)

f = ''

if select_seed_MIDI != "Upload your own custom MIDI":
  print('Loading seed MIDI...')
  f = '/content/Allegro-Music-Transformer/Seeds/'+select_seed_MIDI+'.mid'
  score = TMIDIX.midi2single_track_ms_score(open(f, 'rb').read(), recalculate_channels=False)

else:
  print('Upload your own custom MIDI...')
  print('=' * 70)
  uploaded_MIDI = files.upload()
  if list(uploaded_MIDI.keys()):
    score = TMIDIX.midi2single_track_ms_score(list(uploaded_MIDI.values())[0], recalculate_channels=False)
    f = list(uploaded_MIDI.keys())[0]

if f != '':

  print('=' * 70)
  print('File:', f)
  print('=' * 70)

  #=======================================================
  # START PROCESSING

  melody_chords_f = []

  # INSTRUMENTS CONVERSION CYCLE
  events_matrix = []
  itrack = 1
  patches = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

  patch_map = [
              [0, 1, 2, 3, 4, 5, 6, 7], # Piano
              [24, 25, 26, 27, 28, 29, 30], # Guitar
              [32, 33, 34, 35, 36, 37, 38, 39], # Bass
              [40, 41], # Violin
              [42, 43], # Cello
              [46], # Harp
              [56, 57, 58, 59, 60], # Trumpet
              [64, 65, 66, 67, 68, 69, 70, 71], # Sax
              [72, 73, 74, 75, 76, 77, 78], # Flute
              [-1], # Drums
              [52, 53], # Choir
              [16, 17, 18, 19, 20] # Organ
              ]

  while itrack < len(score):
      for event in score[itrack]:
          if event[0] == 'note' or event[0] == 'patch_change':
              events_matrix.append(event)
      itrack += 1

  events_matrix.sort(key=lambda x: x[1])

  events_matrix1 = []

  for event in events_matrix:
          if event[0] == 'patch_change':
              patches[event[2]] = event[3]

          if event[0] == 'note':
              event.extend([patches[event[3]]])
              once = False

              for p in patch_map:
                  if event[6] in p and event[3] != 9: # Except the drums
                      event[3] = patch_map.index(p)
                      once = True

              if not once and event[3] != 9: # Except the drums
                  event[3] = 15 # All other instruments/patches channel
                  event[5] = max(80, event[5])

              if event[3] < 12: # We won't write chans 12-16 for now...
                  events_matrix1.append(event)

  if len(events_matrix1) > 0:


      #=======================================================
      # PRE-PROCESSING

      # checking number of instruments in a composition
      instruments_list_without_drums = list(set([y[3] for y in events_matrix1 if y[3] != 9]))

      if len(events_matrix1) > 0 and len(instruments_list_without_drums) > 0:

        # recalculating timings
        for e in events_matrix1:
            e[1] = int(e[1] / 8) # Max 2 seconds for start-times
            e[2] = int(e[2] / 32) # Max 4 seconds for durations

        # Sorting by pitch, then by start-time
        events_matrix1.sort(key=lambda x: x[4], reverse=True)
        events_matrix1.sort(key=lambda x: x[1])

        #=======================================================
        # FINAL PRE-PROCESSING

        melody_chords = []

        pe = events_matrix1[0]

        for e in events_matrix1:

            # Cliping all values...
            time = max(0, min(255, e[1]-pe[1]))
            dur = max(1, min(127, e[2]))
            cha = max(0, min(11, e[3]))
            ptc = max(1, min(127, e[4]))

            # Calculating octo-velocity
            vel = max(8, min(127, e[5]))
            velocity = round(vel / 15)-1

            # Writing final note
            melody_chords.append([time, dur, cha, ptc, velocity])

            pe = e

        times = [y[0] for y in melody_chords[12:]]
        avg_time = sum(times) / len(times)

        times_list = list(set(times))

        mode_dur = statistics.mode([y[1] for y in melody_chords if y[2] != 9])

        instruments_list = list(set([y[2] for y in melody_chords]))
        num_instr = len(instruments_list)

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

        # TOTAL DICTIONARY SIZE 3087+1=3088

        #=======================================================
        # MAIN PROCESSING CYCLE
        #=======================================================

        chords_count = 0

        melody_chords_f.extend([2816]) # Zero chords count

        if melody_chords[0][0] == 0:
          melody_chords_f.extend([0]) # Zero time, if present

        notes_counter = 0
        chords_counter = 0

        for m in melody_chords:

          time = m[0]

          # Chords counter token
          if chords_count % 50 == 0 and chords_count != 0 and time != 0:
            melody_chords_f.extend([2816+min(255, ((chords_count // 50)))])

          if time != 0:
            chords_count += 1

          # WRITING EACH NOTE HERE
          dur_vel = (m[1] * 8) + m[4]
          cha_ptc = (m[2] * 128) + m[3]

          if time != 0:
              melody_chords_f.extend([time, dur_vel+256, cha_ptc+1280])
              chords_counter += 1

          else:
              melody_chords_f.extend([dur_vel+256, cha_ptc+1280])

          notes_counter += 1

  melody_chords_f = melody_chords_f[:number_of_prime_tokens]

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

  song = melody_chords_f

  song_f = []

  time = 0
  dur = 0
  vel = 90
  pitch = 0
  channel = 0

  for ss in song:

    if ss > 0 and ss < 256:

        time += ss * 8

    if ss >= 256 and ss < 1280:

        dur = ((ss-256) // 8) * 32
        vel = (((ss-256) % 8)+1) * 15

    if ss >= 1280 and ss < 2816:
        channel = (ss-1280) // 128
        pitch = (ss-1280) % 128

        song_f.append(['note', time, dur, channel, pitch, vel ])

  detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
                                                            output_signature = 'Allegro Music Transformer',
                                                            output_file_name = '/content/Allegro-Music-Transformer-Seed-Composition',
                                                            track_name='Project Los Angeles',
                                                            list_of_MIDI_patches=[0, 24, 32, 40, 42, 46, 56, 65, 73, 0, 53, 19, 0, 0, 0, 0]
                                                            )

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

  print('=' * 70)
  print('Composition stats:')
  print('Composition has', notes_counter, 'notes')
  print('Composition has', chords_counter, 'chords')
  print('Composition has', len(melody_chords_f), 'tokens')
  print('=' * 70)

  fname = '/content/Allegro-Music-Transformer-Seed-Composition'

  x = []
  y =[]
  c = []

  colors = ['red', 'yellow', 'green', 'cyan', 'blue', 'pink', 'orange', 'purple', 'gray', 'white', 'gold', 'silver']

  block_lines = [(song_f[-1][1] / 1000)]
  block_tokens = [min(len(melody_chords_f), number_of_prime_tokens)]

  for s in song_f:
    x.append(s[1] / 1000)
    y.append(s[4])
    c.append(colors[s[3]])

  if render_MIDI_to_audio:
    FluidSynth("/usr/share/sounds/sf2/FluidR3_GM.sf2", 16000).midi_to_audio(str(fname + '.mid'), str(fname + '.wav'))
    display(Audio(str(fname + '.wav'), rate=16000))

  plt.figure(figsize=(14,5))
  ax=plt.axes(title=fname)
  ax.set_facecolor('black')

  plt.scatter(x,y, c=c)
  plt.xlabel("Time")
  plt.ylabel("Pitch")
  plt.show()

else:
  print('=' * 70)

"""# (COMPOSITION LOOP)

## Run the cells below in a loop to generate endless continuation
"""

#@title Standard Continuation Generator

#@markdown Generation settings

number_of_tokens_to_generate = 300 # @param {type:"slider", min:33, max:1023, step:3}
number_of_batches_to_generate = 4 #@param {type:"slider", min:1, max:16, step:1}
preview_length_in_tokens = 120 # @param {type:"slider", min:33, max:240, step:3}
number_of_memory_tokens = 2048 # @param {type:"slider", min:402, max:2048, step:2}
temperature = 0.9 #@param {type:"slider", min:0.1, max:1, step:0.1}

#@markdown Other settings

render_MIDI_to_audio = True # @param {type:"boolean"}

print('=' * 70)
print('Allegro Music Transformer Standard Continuation Model Generator')
print('=' * 70)

torch.cuda.empty_cache()

preview = melody_chords_f[-preview_length_in_tokens:]

inp = [melody_chords_f[-number_of_memory_tokens:]] * number_of_batches_to_generate

inp = torch.LongTensor(inp).cuda()

with ctx:
  out = model.module.generate(inp,
                              number_of_tokens_to_generate,
                              temperature=temperature,
                              return_prime=False,
                              verbose=True)

out0 = out.tolist()

torch.cuda.empty_cache()

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

#======================================================================
print('=' * 70)
print('Rendering results...')

for i in range(number_of_batches_to_generate):

  print('=' * 70)
  print('Batch #', i)
  print('=' * 70)

  out1 = out0[i]

  print('Sample INTs', out1[:12])
  print('=' * 70)

  if len(out) != 0:

      song = preview + out1
      song_f = []
      time = 0
      dur = 0
      vel = 0
      pitch = 0
      channel = 0

      for ss in song:

        if ss > 0 and ss < 256:

            time += ss * 8

        if ss >= 256 and ss < 1280:

            dur = ((ss-256) // 8) * 32
            vel = (((ss-256) % 8)+1) * 15

        if ss >= 1280 and ss < 2816:
            channel = (ss-1280) // 128
            pitch = (ss-1280) % 128

            song_f.append(['note', time, dur, channel, pitch, vel ])

      detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
                                                                output_signature = 'Allegro Music Transformer',
                                                                output_file_name = '/content/Allegro-Music-Transformer-Music-Composition_'+str(i),
                                                                track_name='Project Los Angeles',
                                                                list_of_MIDI_patches=[0, 24, 32, 40, 42, 46, 56, 65, 73, 0, 53, 19, 0, 0, 0, 0]
                                                                )
      print('=' * 70)
      print('Displaying resulting composition...')
      print('=' * 70)

      fname = '/content/Allegro-Music-Transformer-Music-Composition_'+str(i)

      x = []
      y =[]
      c = []

      colors = ['red', 'yellow', 'green', 'cyan', 'blue', 'pink', 'orange', 'purple', 'gray', 'white', 'gold', 'silver']

      for s in song_f:
        x.append(s[1] / 1000)
        y.append(s[4])
        c.append(colors[s[3]])

      if render_MIDI_to_audio:
        FluidSynth("/usr/share/sounds/sf2/FluidR3_GM.sf2", 16000).midi_to_audio(str(fname + '.mid'), str(fname + '.wav'))
        display(Audio(str(fname + '.wav'), rate=16000))

      plt.figure(figsize=(14,5))
      ax=plt.axes(title=fname)
      ax.set_facecolor('black')

      plt.scatter(x,y, c=c)

      pbl = song_f[(int(preview_length_in_tokens / 3))][1] / 1000

      ax.axvline(x=pbl, c='w')

      plt.xlabel("Time")
      plt.ylabel("Pitch")
      plt.show()

#@title Choose one generated block to add to the composition
block_action = "add_last_generated_block" #@param ["add_last_generated_block", "remove_last_added_block"]
add_block_with_batch_number = 0 #@param {type:"slider", min:0, max:15, step:1}
render_MIDI_to_audio = False # @param {type:"boolean"}

print('=' * 70)

if block_action == 'add_last_generated_block':
  melody_chords_f.extend(out0[min(len(out0)-1, add_block_with_batch_number)])
  print('Block added!')
else:
  if len(block_tokens) > 1:
    melody_chords_f = melody_chords_f[:(len(melody_chords_f)-block_tokens[-1])]
    print('Block removed!')
  else:
    print('Nothing to remove!!!')

print('=' * 70)
print('Composition now has', len(melody_chords_f), 'tokens')

print('=' * 70)
print('Sample INTs', out1[:12])
print('=' * 70)

if len(melody_chords_f) != 0:

    song = melody_chords_f
    song_f = []
    time = 0
    dur = 0
    vel = 0
    pitch = 0
    channel = 0

    for ss in song:

      if ss > 0 and ss < 256:

          time += ss * 8

      if ss >= 256 and ss < 1280:

          dur = ((ss-256) // 8) * 32
          vel = (((ss-256) % 8)+1) * 15

      if ss >= 1280 and ss < 2816:
          channel = (ss-1280) // 128
          pitch = (ss-1280) % 128

          song_f.append(['note', time, dur, channel, pitch, vel ])

    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
                                                              output_signature = 'Allegro Music Transformer',
                                                              output_file_name = '/content/Allegro-Music-Transformer-Music-Composition',
                                                              track_name='Project Los Angeles',
                                                              list_of_MIDI_patches=[0, 24, 32, 40, 42, 46, 56, 65, 73, 0, 53, 19, 0, 0, 0, 0]
                                                              )
    print('=' * 70)
    print('Displaying resulting composition...')
    print('=' * 70)

    fname = '/content/Allegro-Music-Transformer-Music-Composition'

    x = []
    y =[]
    c = []

    colors = ['red', 'yellow', 'green', 'cyan', 'blue', 'pink', 'orange', 'purple', 'gray', 'white', 'gold', 'silver']

    if block_action == 'add_last_generated_block':
      block_lines.append((song_f[-1][1] / 1000))
      block_tokens.append(len(out0[min(len(out0)-1, add_block_with_batch_number)]))
    else:
      if len(block_tokens) > 1:
        block_lines.pop()
        block_tokens.pop()

    for s in song_f:
      x.append(s[1] / 1000)
      y.append(s[4])
      c.append(colors[s[3]])

    if render_MIDI_to_audio:
      FluidSynth("/usr/share/sounds/sf2/FluidR3_GM.sf2", 16000).midi_to_audio(str(fname + '.mid'), str(fname + '.wav'))
      display(Audio(str(fname + '.wav'), rate=16000))

    plt.figure(figsize=(14,5))
    ax=plt.axes(title=fname)
    ax.set_facecolor('black')

    plt.scatter(x,y, c=c)

    for bl in block_lines:
      ax.axvline(x=bl, c='w')

    plt.xlabel("Time")
    plt.ylabel("Pitch")
    plt.show()

"""# Congrats! You did it! :)"""