Spaces:
Running
Running
| import re | |
| import os | |
| import time | |
| import torch | |
| import shutil | |
| import argparse | |
| import warnings | |
| import gradio as gr | |
| from transformers import GPT2Config | |
| from model import Patchilizer, TunesFormer | |
| from convert import abc2xml, xml2, xml2img | |
| from utils import ( | |
| PATCH_NUM_LAYERS, | |
| PATCH_LENGTH, | |
| CHAR_NUM_LAYERS, | |
| PATCH_SIZE, | |
| SHARE_WEIGHTS, | |
| WEIGHTS_PATH, | |
| TEMP_DIR, | |
| TEYVAT, | |
| DEVICE, | |
| ) | |
| def get_args(parser: argparse.ArgumentParser): | |
| parser.add_argument( | |
| "-num_tunes", | |
| type=int, | |
| default=1, | |
| help="the number of independently computed returned tunes", | |
| ) | |
| parser.add_argument( | |
| "-max_patch", | |
| type=int, | |
| default=128, | |
| help="integer to define the maximum length in tokens of each tune", | |
| ) | |
| parser.add_argument( | |
| "-top_p", | |
| type=float, | |
| default=0.8, | |
| help="float to define the tokens that are within the sample operation of text generation", | |
| ) | |
| parser.add_argument( | |
| "-top_k", | |
| type=int, | |
| default=8, | |
| help="integer to define the tokens that are within the sample operation of text generation", | |
| ) | |
| parser.add_argument( | |
| "-temperature", | |
| type=float, | |
| default=1.2, | |
| help="the temperature of the sampling operation", | |
| ) | |
| parser.add_argument("-seed", type=int, default=None, help="seed for randomstate") | |
| parser.add_argument( | |
| "-show_control_code", | |
| type=bool, | |
| default=False, | |
| help="whether to show control code", | |
| ) | |
| return parser.parse_args() | |
| def generate_music(args, region: str): | |
| patchilizer = Patchilizer() | |
| patch_config = GPT2Config( | |
| num_hidden_layers=PATCH_NUM_LAYERS, | |
| max_length=PATCH_LENGTH, | |
| max_position_embeddings=PATCH_LENGTH, | |
| vocab_size=1, | |
| ) | |
| char_config = GPT2Config( | |
| num_hidden_layers=CHAR_NUM_LAYERS, | |
| max_length=PATCH_SIZE, | |
| max_position_embeddings=PATCH_SIZE, | |
| vocab_size=128, | |
| ) | |
| model = TunesFormer(patch_config, char_config, share_weights=SHARE_WEIGHTS) | |
| checkpoint = torch.load(WEIGHTS_PATH, map_location=torch.device("cpu")) | |
| model.load_state_dict(checkpoint["model"]) | |
| model = model.to(DEVICE) | |
| model.eval() | |
| prompt = f"A:{region}\n" | |
| tunes = "" | |
| num_tunes = args.num_tunes | |
| max_patch = args.max_patch | |
| top_p = args.top_p | |
| top_k = args.top_k | |
| temperature = args.temperature | |
| seed = args.seed | |
| show_control_code = args.show_control_code | |
| print(" Hyper parms ".center(60, "#"), "\n") | |
| arg_dict: dict = vars(args) | |
| for key in arg_dict.keys(): | |
| print(f"{key}: {str(arg_dict[key])}") | |
| print("\n", " Output tunes ".center(60, "#")) | |
| start_time = time.time() | |
| for i in range(num_tunes): | |
| title_artist = f"T:{region} Style Fragment\nC:Generated by AI\n" | |
| tune = f"X:{str(i + 1)}\n{title_artist + prompt}" | |
| lines = re.split(r"(\n)", tune) | |
| tune = "" | |
| skip = False | |
| for line in lines: | |
| if show_control_code or line[:2] not in ["S:", "B:", "E:"]: | |
| if not skip: | |
| print(line, end="") | |
| tune += line | |
| skip = False | |
| else: | |
| skip = True | |
| input_patches = torch.tensor( | |
| [patchilizer.encode(prompt, add_special_patches=True)[:-1]], device=DEVICE | |
| ) | |
| if tune == "": | |
| tokens = None | |
| else: | |
| prefix = patchilizer.decode(input_patches[0]) | |
| remaining_tokens = prompt[len(prefix) :] | |
| tokens = torch.tensor( | |
| [patchilizer.bos_token_id] + [ord(c) for c in remaining_tokens], | |
| device=DEVICE, | |
| ) | |
| while input_patches.shape[1] < max_patch: | |
| predicted_patch, seed = model.generate( | |
| input_patches, | |
| tokens, | |
| top_p=top_p, | |
| top_k=top_k, | |
| temperature=temperature, | |
| seed=seed, | |
| ) | |
| tokens = None | |
| if predicted_patch[0] != patchilizer.eos_token_id: | |
| next_bar = patchilizer.decode([predicted_patch]) | |
| if show_control_code or next_bar[:2] not in ["S:", "B:", "E:"]: | |
| print(next_bar, end="") | |
| tune += next_bar | |
| if next_bar == "": | |
| break | |
| next_bar = remaining_tokens + next_bar | |
| remaining_tokens = "" | |
| predicted_patch = torch.tensor( | |
| patchilizer.bar2patch(next_bar), device=DEVICE | |
| ).unsqueeze(0) | |
| input_patches = torch.cat( | |
| [input_patches, predicted_patch.unsqueeze(0)], dim=1 | |
| ) | |
| else: | |
| break | |
| tunes += f"{tune}\n\n" | |
| print("\n") | |
| print("Generation time: {:.2f} seconds".format(time.time() - start_time)) | |
| timestamp = time.strftime("%a_%d_%b_%Y_%H_%M_%S", time.localtime()) | |
| try: | |
| xml = abc2xml(tunes, f"{TEMP_DIR}/[{region}]{timestamp}.musicxml") | |
| midi = xml2(xml, "mid") | |
| audio = xml2(xml, "wav") | |
| pdf, jpg = xml2img(xml) | |
| mxl = xml2(xml, "mxl") | |
| return tunes, midi, pdf, xml, mxl, jpg, audio | |
| except Exception as e: | |
| print(f"Invalid abc generated: {e}, retrying...") | |
| return generate_music(args, region) | |
| def infer(region: str): | |
| if os.path.exists(TEMP_DIR): | |
| shutil.rmtree(TEMP_DIR) | |
| os.makedirs(TEMP_DIR, exist_ok=True) | |
| parser = argparse.ArgumentParser() | |
| args = get_args(parser) | |
| if region == "Natlan": | |
| region = "Teyvat" | |
| return generate_music(args, region) | |
| if __name__ == "__main__": | |
| warnings.filterwarnings("ignore") | |
| with gr.Blocks() as demo: | |
| with gr.Row(): | |
| with gr.Column(): | |
| region_opt = gr.Dropdown( | |
| choices=TEYVAT, | |
| value="Mondstadt", | |
| label="Region", | |
| ) | |
| gen_btn = gr.Button("Generate") | |
| gr.Markdown( | |
| """ | |
| Welcome to this space based on the Tunesformer open source project, which is totally free! | |
| The current model is still in debugging, the plan is in the Genshin Impact after the main line is killed, all countries and regions after all the characters are open, the second creation of the concert will be complete and the sample is balanced, at that time to re-fine-tune the model and add the reality of the style of screening to assist in the game of each country's output to strengthen the learning in order to enhance the output differentiation and quality. | |
| Data source: <a href="https://musescore.org">MuseScore</a><br> | |
| Tags source: <a href="https://genshin-impact.fandom.com/wiki/Genshin_Impact_Wiki">Genshin Impact Wiki | Fandom</a><br> | |
| Model base: <a href="https://github.com/sander-wood/tunesformer">Tunesformer</a> | |
| Note: Data engineering on the Star Rail is in operation, and will hopefully be baselined in the future as well with the mainline kill.""" | |
| ) | |
| with gr.Column(): | |
| wav_output = gr.Audio(label="Audio", type="filepath") | |
| dld_midi = gr.File(label="Download MIDI") | |
| pdf_score = gr.File(label="Download PDF") | |
| dld_xml = gr.File(label="Download MusicXML") | |
| dld_mxl = gr.File(label="Download MXL") | |
| abc_output = gr.Textbox(label="ABC notation", show_copy_button=True) | |
| img_score = gr.Image(label="Staff", type="filepath") | |
| gen_btn.click( | |
| infer, | |
| inputs=region_opt, | |
| outputs=[ | |
| abc_output, | |
| dld_midi, | |
| pdf_score, | |
| dld_xml, | |
| dld_mxl, | |
| img_score, | |
| wav_output, | |
| ], | |
| ) | |
| demo.launch() | |