Last commit not found
| #================================================================== | |
| # https://huggingface.co/spaces/asigalov61/Popular-Hook-Transformer | |
| #================================================================== | |
| import time as reqtime | |
| import datetime | |
| from pytz import timezone | |
| import statistics | |
| import re | |
| import tqdm | |
| import gradio as gr | |
| from x_transformer_1_23_2 import * | |
| import random | |
| from midi_to_colab_audio import midi_to_colab_audio | |
| import TMIDIX | |
| import matplotlib.pyplot as plt | |
| #===================================================================================== | |
| print('=' * 70) | |
| print('Popular Hook Transformer') | |
| print('=' * 70) | |
| print('Loading Popular Hook Transformer training data...') | |
| #==================================================================================== | |
| SEQ_LEN = 512 | |
| PAD_IDX = 918 | |
| DEVICE = 'cpu' | |
| #==================================================================================== | |
| def str_strip(string): | |
| return re.sub(r'[^A-Za-z-]+', '', string).rstrip('-') | |
| def mode_time(seq): | |
| return statistics.mode([t for t in seq if 0 < t < 128]) | |
| def mode_dur(seq): | |
| return statistics.mode([t-128 for t in seq if 128 < t < 256]) | |
| def mode_pitch(seq): | |
| return statistics.mode([t % 128 for t in seq if 256 < t < 512]) | |
| parts_dict = sorted(set([str_strip(s[2]).rstrip('-') for s in melody_chords_f])) | |
| train_data = [] | |
| for m in tqdm.tqdm(melody_chords_f): | |
| if 64 < len(m[5]) < 506: | |
| for tv in range(-3, 3): | |
| part = str_strip(m[2]) | |
| part_tok = parts_dict.index(part) | |
| score = [t+tv if 256 < t < 512 else t for t in m[5]] | |
| seq = [916] + [part_tok+512, mode_time(score)+532, mode_dur(score)+660, mode_pitch(score)+tv+788] | |
| seq += score | |
| seq += [917] | |
| seq = seq + [PAD_IDX] * (SEQ_LEN - len(seq)) | |
| train_data.append(seq) | |
| #==================================================================================== | |
| print('Done!') | |
| print('=' * 70) | |
| print('All data is good:', len(max(train_data, key=len)) == len(min(train_data, key=len))) | |
| print('=' * 70) | |
| print('Randomizing training data...') | |
| random.shuffle(train_data) | |
| print('Done!') | |
| print('=' * 70) | |
| print('Total length of training data:', len(train_data)) | |
| print('=' * 70) | |
| #==================================================================================== | |
| print('Loading Popular Hook Transformer pre-trained model...') | |
| print('=' * 70) | |
| print('Instantiating model...') | |
| model = TransformerWrapper( | |
| num_tokens = PAD_IDX+1, | |
| max_seq_len = SEQ_LEN, | |
| attn_layers = Decoder(dim = 1024, | |
| depth = 4, | |
| heads = 32, | |
| rotary_pos_emb = True, | |
| attn_flash = True | |
| ) | |
| ) | |
| model = AutoregressiveWrapper(model, ignore_index = PAD_IDX, pad_value=PAD_IDX) | |
| print('=' * 70) | |
| print('Loading model checkpoint...') | |
| model_path = 'Popular_Hook_Transformer_Small_Trained_Model_10869_steps_0.2308_loss_0.9252_acc.pth' | |
| model.load_state_dict(torch.load(model_path)) | |
| print('=' * 70) | |
| model.to(DEVICE) | |
| model.eval() | |
| ctx = torch.amp.autocast(device_type=DEVICE, dtype=torch.bfloat16) | |
| print('Done!') | |
| print('=' * 70) | |
| #==================================================================================== | |
| def Generate_POP_Section(input_parsons_code, | |
| input_first_note_duration, | |
| iinput_first_note_MIDI_pitch | |
| ): | |
| print('=' * 70) | |
| print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT))) | |
| start_time = reqtime.time() | |
| print('=' * 70) | |
| print('Requested settings:') | |
| print('-' * 70) | |
| print('Parsons code:', input_parsons_code) | |
| print('First note duration:', input_first_note_duration) | |
| print('First note MIDI pitch:', iinput_first_note_MIDI_pitch) | |
| print('=' * 70) | |
| #=============================================================================== | |
| print('Instantiating Parsons Code Melody Transformer model...') | |
| SEQ_LEN = 322 | |
| PAD_IDX = 392 | |
| model = TransformerWrapper( | |
| num_tokens = PAD_IDX+1, | |
| max_seq_len = SEQ_LEN, | |
| attn_layers = Decoder(dim = 1024, | |
| depth = 4, | |
| heads = 8, | |
| rotary_pos_emb = True, | |
| attn_flash = True | |
| ) | |
| ) | |
| model = AutoregressiveWrapper(model, ignore_index = PAD_IDX, pad_value=PAD_IDX) | |
| print('=' * 70) | |
| print('Loading model checkpoint...') | |
| model_path = 'Parsons_Code_Melody_Transformer_Trained_Model_13786_steps_0.3058_loss_0.8819_acc.pth' | |
| model.load_state_dict(torch.load(model_path, map_location='cpu')) | |
| model.cpu() | |
| model.eval() | |
| dtype = torch.bfloat16 | |
| ctx = torch.amp.autocast(device_type='cpu', dtype=dtype) | |
| print('Done!') | |
| print('=' * 70) | |
| #=============================================================================== | |
| print('Prepping Parsons code string...') | |
| td_str = re.sub('[^*DRU]', '', input_parsons_code) | |
| print(len(td_str)) | |
| print('=' * 70) | |
| if '*' in td_str and len(td_str) > 1: | |
| code_mult = (64 // len(td_str[1:]))+1 | |
| mult_code = ('*' + (td_str[1:] * code_mult))[:64] | |
| else: | |
| mult_code = '*UUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUU' | |
| pcode = parsons_code_to_tokens(mult_code) | |
| print('Done!') | |
| print('=' * 70) | |
| #=============================================================================== | |
| print('Generating melody...') | |
| song = [] | |
| song.append(389) | |
| song.extend(pcode) | |
| song.append(390) | |
| song.extend([388, 0, 10+128, 66+256]) | |
| for i in tqdm.tqdm(range(1, len(td_str[:64]))): | |
| song.append(pcode[i]) | |
| x = torch.tensor(song, dtype=torch.long, device='cpu') | |
| with ctx: | |
| out = model.generate(x, | |
| 3, | |
| filter_logits_fn=top_k, | |
| filter_kwargs={'k': 1}, | |
| temperature=1.0, | |
| return_prime=False, | |
| verbose=False) | |
| y = out.tolist()[0] | |
| song.extend(y) | |
| print('Done!') | |
| print('=' * 70) | |
| #=============================================================================== | |
| print('Rendering results...') | |
| print('=' * 70) | |
| print('Sample INTs', song[:5]) | |
| print('=' * 70) | |
| song_f = [] | |
| time = 0 | |
| dur = 4 | |
| vel = 90 | |
| pitch = 60 | |
| 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 | |
| song_f.append(['note', time, dur, channel, pitch, vel, 0]) | |
| fn1 = 'Parsons-Code-Melody-Transformer-Composition' | |
| detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f, | |
| output_signature = 'Parsons Code Melody Transformer', | |
| output_file_name = fn1, | |
| track_name='Project Los Angeles' | |
| ) | |
| 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 = str(new_fn) | |
| output_audio = (16000, audio) | |
| output_plot = TMIDIX.plot_ms_SONG(song_f, plot_title=output_midi_title, return_plt=True) | |
| 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, 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'>Parsons Code Melody Transformer</h1>") | |
| gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Generate unique melodies from Parsons codes</h1>") | |
| gr.Markdown( | |
| "\n\n" | |
| "This is a demo for Clean Melodies subset of Tegridy MIDI Dataset\n\n" | |
| "Check out [Tegridy MIDI Dataset](https://github.com/asigalov61/Tegridy-MIDI-Dataset) on GitHub!\n\n" | |
| ) | |
| gr.Markdown("## Enter Parsons code:") | |
| input_parsons_code = gr.Textbox(label="Parsons code", | |
| info="Make sure your Parsons code starts with *", | |
| lines=1, | |
| value="*" | |
| ) | |
| clr_btn = gr.ClearButton(components=input_parsons_code) | |
| def reset_pcode(): | |
| return '*' | |
| clr_btn.click(reset_pcode, outputs=input_parsons_code) | |
| gr.Markdown("## Select generation options:") | |
| input_first_note_duration = gr.Slider(1, 127, value=15, step=1, label="First note duration value") | |
| iinput_first_note_MIDI_pitch = gr.Slider(1, 127, value=60, step=1, label="First note MIDI pitch") | |
| run_btn = gr.Button("Generate melody", variant="primary") | |
| gr.Markdown("## Output results") | |
| output_midi_title = gr.Textbox(label="Output MIDI title") | |
| output_audio = gr.Audio(label="Output MIDI audio", format="mp3", 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(Generate_Melody, [input_parsons_code, | |
| input_first_note_duration, | |
| iinput_first_note_MIDI_pitch | |
| ], | |
| [output_midi_title, output_midi, output_audio, output_plot]) | |
| gr.Examples( | |
| [["*UUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUU", 15, 60], | |
| ["*UDDDUDDDUDRURUDUUDRDDUDDRUDUDURUDRUDUDDDUDDDRDUURUDUUDDDUDRRUUD", 15, 60], | |
| ["*DUDDDUUDDUUDDUDUDDDUUUUUDDDDUDDDUUDDUUDDUUDUDDUDDDUUDDUUDDUDUDD", 15, 60], | |
| ["*DUUDDRDDUURUDUDDDUDDDDDURDDUDRDURUURUURDDDUURDUURUDUUDURDUDUDRD", 15, 60], | |
| ["*UUUDDUUUDDDDDUDDUUDDDDUUDDUDDDDDUUUDDDDDUDDUUUDDDURDUDUUUDDUUUD", 15, 60], | |
| ["*UDUUDRUDDUDRURUURUUUUUDUDDUDDUDDUDRUDDUDRUDDDUDUUDRUDDUDRURUURU", 15, 60], | |
| ], | |
| [input_parsons_code, | |
| input_first_note_duration, | |
| iinput_first_note_MIDI_pitch | |
| ], | |
| [output_midi_title, output_midi, output_audio, output_plot], | |
| Generate_Melody, | |
| cache_examples=True, | |
| ) | |
| app.queue().launch() |