Spaces:
Running
on
Zero
Running
on
Zero
| import argparse | |
| import glob | |
| import os.path | |
| import uuid | |
| import gradio as gr | |
| import numpy as np | |
| import onnxruntime as rt | |
| import tqdm | |
| import json | |
| from huggingface_hub import hf_hub_download | |
| import MIDI | |
| from midi_synthesizer import synthesis | |
| from midi_tokenizer import MIDITokenizer | |
| in_space = os.getenv("SYSTEM") == "spaces" | |
| def softmax(x, axis): | |
| x_max = np.amax(x, axis=axis, keepdims=True) | |
| exp_x_shifted = np.exp(x - x_max) | |
| return exp_x_shifted / np.sum(exp_x_shifted, axis=axis, keepdims=True) | |
| def sample_top_p_k(probs, p, k): | |
| probs_idx = np.argsort(-probs, axis=-1) | |
| probs_sort = np.take_along_axis(probs, probs_idx, -1) | |
| probs_sum = np.cumsum(probs_sort, axis=-1) | |
| mask = probs_sum - probs_sort > p | |
| probs_sort[mask] = 0.0 | |
| mask = np.zeros(probs_sort.shape[-1]) | |
| mask[:k] = 1 | |
| probs_sort = probs_sort * mask | |
| probs_sort /= np.sum(probs_sort, axis=-1, keepdims=True) | |
| shape = probs_sort.shape | |
| probs_sort_flat = probs_sort.reshape(-1, shape[-1]) | |
| probs_idx_flat = probs_idx.reshape(-1, shape[-1]) | |
| next_token = np.stack([np.random.choice(idxs, p=pvals) for pvals, idxs in zip(probs_sort_flat, probs_idx_flat)]) | |
| next_token = next_token.reshape(*shape[:-1]) | |
| return next_token | |
| def generate(model, prompt=None, max_len=512, temp=1.0, top_p=0.98, top_k=20, | |
| disable_patch_change=False, disable_control_change=False, disable_channels=None): | |
| if disable_channels is not None: | |
| disable_channels = [tokenizer.parameter_ids["channel"][c] for c in disable_channels] | |
| else: | |
| disable_channels = [] | |
| max_token_seq = tokenizer.max_token_seq | |
| if prompt is None: | |
| input_tensor = np.full((1, max_token_seq), tokenizer.pad_id, dtype=np.int64) | |
| input_tensor[0, 0] = tokenizer.bos_id # bos | |
| else: | |
| prompt = prompt[:, :max_token_seq] | |
| if prompt.shape[-1] < max_token_seq: | |
| prompt = np.pad(prompt, ((0, 0), (0, max_token_seq - prompt.shape[-1])), | |
| mode="constant", constant_values=tokenizer.pad_id) | |
| input_tensor = prompt | |
| input_tensor = input_tensor[None, :, :] | |
| cur_len = input_tensor.shape[1] | |
| bar = tqdm.tqdm(desc="generating", total=max_len - cur_len, disable=in_space) | |
| with bar: | |
| while cur_len < max_len: | |
| end = False | |
| hidden = model[0].run(None, {'x': input_tensor})[0][:, -1] | |
| next_token_seq = np.empty((1, 0), dtype=np.int64) | |
| event_name = "" | |
| for i in range(max_token_seq): | |
| mask = np.zeros(tokenizer.vocab_size, dtype=np.int64) | |
| if i == 0: | |
| mask_ids = list(tokenizer.event_ids.values()) + [tokenizer.eos_id] | |
| if disable_patch_change: | |
| mask_ids.remove(tokenizer.event_ids["patch_change"]) | |
| if disable_control_change: | |
| mask_ids.remove(tokenizer.event_ids["control_change"]) | |
| mask[mask_ids] = 1 | |
| else: | |
| param_name = tokenizer.events[event_name][i - 1] | |
| mask_ids = tokenizer.parameter_ids[param_name] | |
| if param_name == "channel": | |
| mask_ids = [i for i in mask_ids if i not in disable_channels] | |
| mask[mask_ids] = 1 | |
| logits = model[1].run(None, {'x': next_token_seq, "hidden": hidden})[0][:, -1:] | |
| scores = softmax(logits / temp, -1) * mask | |
| sample = sample_top_p_k(scores, top_p, top_k) | |
| if i == 0: | |
| next_token_seq = sample | |
| eid = sample.item() | |
| if eid == tokenizer.eos_id: | |
| end = True | |
| break | |
| event_name = tokenizer.id_events[eid] | |
| else: | |
| next_token_seq = np.concatenate([next_token_seq, sample], axis=1) | |
| if len(tokenizer.events[event_name]) == i: | |
| break | |
| if next_token_seq.shape[1] < max_token_seq: | |
| next_token_seq = np.pad(next_token_seq, ((0, 0), (0, max_token_seq - next_token_seq.shape[-1])), | |
| mode="constant", constant_values=tokenizer.pad_id) | |
| next_token_seq = next_token_seq[None, :, :] | |
| input_tensor = np.concatenate([input_tensor, next_token_seq], axis=1) | |
| cur_len += 1 | |
| bar.update(1) | |
| yield next_token_seq.reshape(-1) | |
| if end: | |
| break | |
| def create_msg(name, data): | |
| return {"name": name, "data": data, "uuid": uuid.uuid4().hex} | |
| def send_msgs(msgs, msgs_history=None): | |
| if msgs_history is None: | |
| msgs_history = [] | |
| msgs_history.append(msgs) | |
| if len(msgs_history) > 25: | |
| msgs_history= msgs_history[1:] | |
| return json.dumps(msgs_history) | |
| def run(model_name, tab, instruments, drum_kit, bpm, mid, midi_events, gen_events, temp, top_p, top_k, allow_cc): | |
| msgs_history = [] | |
| mid_seq = [] | |
| bpm = int(bpm) | |
| gen_events = int(gen_events) | |
| max_len = gen_events | |
| disable_patch_change = False | |
| disable_channels = None | |
| if tab == 0: | |
| i = 0 | |
| mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)] | |
| if bpm != 0: | |
| mid.append(tokenizer.event2tokens(["set_tempo",0,0,0, bpm])) | |
| patches = {} | |
| if instruments is None: | |
| instruments = [] | |
| for instr in instruments: | |
| patches[i] = patch2number[instr] | |
| i = (i + 1) if i != 8 else 10 | |
| if drum_kit != "None": | |
| patches[9] = drum_kits2number[drum_kit] | |
| for i, (c, p) in enumerate(patches.items()): | |
| mid.append(tokenizer.event2tokens(["patch_change", 0, 0, i, c, p])) | |
| mid_seq = mid | |
| mid = np.asarray(mid, dtype=np.int64) | |
| if len(instruments) > 0: | |
| disable_patch_change = True | |
| disable_channels = [i for i in range(16) if i not in patches] | |
| elif mid is not None: | |
| mid = tokenizer.tokenize(MIDI.midi2score(mid)) | |
| mid = np.asarray(mid, dtype=np.int64) | |
| mid = mid[:int(midi_events)] | |
| max_len += len(mid) | |
| for token_seq in mid: | |
| mid_seq.append(token_seq.tolist()) | |
| init_msgs = [create_msg("visualizer_clear", False)] | |
| for tokens in mid_seq: | |
| init_msgs.append(create_msg("visualizer_append", tokenizer.tokens2event(tokens))) | |
| yield mid_seq, None, None, send_msgs(init_msgs, msgs_history) | |
| model = models[model_name] | |
| generator = generate(model, mid, max_len=max_len, temp=temp, top_p=top_p, top_k=top_k, | |
| disable_patch_change=disable_patch_change, disable_control_change=not allow_cc, | |
| disable_channels=disable_channels) | |
| for i, token_seq in enumerate(generator): | |
| token_seq = token_seq.tolist() | |
| mid_seq.append(token_seq) | |
| event = tokenizer.tokens2event(token_seq) | |
| yield mid_seq, None, None, send_msgs([create_msg("visualizer_append", event), create_msg("progress", [i + 1, gen_events])], msgs_history) | |
| mid = tokenizer.detokenize(mid_seq) | |
| with open(f"output.mid", 'wb') as f: | |
| f.write(MIDI.score2midi(mid)) | |
| audio = synthesis(MIDI.score2opus(mid), soundfont_path) | |
| # resend all msgs | |
| msgs = [create_msg("visualizer_end", None), create_msg("visualizer_clear", True)] | |
| for tokens in mid_seq: | |
| msgs.append(create_msg("visualizer_append", tokenizer.tokens2event(tokens))) | |
| msgs.append(create_msg("visualizer_end", None)) | |
| yield mid_seq, "output.mid", (44100, audio), send_msgs(msgs) | |
| def cancel_run(mid_seq): | |
| if mid_seq is None: | |
| return None, None, [] | |
| mid = tokenizer.detokenize(mid_seq) | |
| with open(f"output.mid", 'wb') as f: | |
| f.write(MIDI.score2midi(mid)) | |
| audio = synthesis(MIDI.score2opus(mid), soundfont_path) | |
| msgs = [create_msg("visualizer_end", None), create_msg("visualizer_clear", True)] | |
| for tokens in mid_seq: | |
| msgs.append(create_msg("visualizer_append", tokenizer.tokens2event(tokens))) | |
| msgs.append(create_msg("visualizer_end", None)) | |
| return "output.mid", (44100, audio), send_msgs(msgs) | |
| def load_javascript(dir="javascript"): | |
| scripts_list = glob.glob(f"{dir}/*.js") | |
| javascript = "" | |
| for path in scripts_list: | |
| with open(path, "r", encoding="utf8") as jsfile: | |
| javascript += f"\n<!-- {path} --><script>{jsfile.read()}</script>" | |
| template_response_ori = gr.routes.templates.TemplateResponse | |
| def template_response(*args, **kwargs): | |
| res = template_response_ori(*args, **kwargs) | |
| res.body = res.body.replace( | |
| b'</head>', f'{javascript}</head>'.encode("utf8")) | |
| res.init_headers() | |
| return res | |
| gr.routes.templates.TemplateResponse = template_response | |
| def hf_hub_download_retry(repo_id, filename): | |
| print(f"downloading {repo_id} {filename}") | |
| retry = 0 | |
| err = None | |
| while retry < 30: | |
| try: | |
| return hf_hub_download(repo_id=repo_id, filename=filename) | |
| except Exception as e: | |
| err = e | |
| retry += 1 | |
| if err: | |
| raise err | |
| number2drum_kits = {-1: "None", 0: "Standard", 8: "Room", 16: "Power", 24: "Electric", 25: "TR-808", 32: "Jazz", | |
| 40: "Blush", 48: "Orchestra"} | |
| patch2number = {v: k for k, v in MIDI.Number2patch.items()} | |
| drum_kits2number = {v: k for k, v in number2drum_kits.items()} | |
| 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 = hf_hub_download_retry(repo_id="skytnt/midi-model", filename="soundfont.sf2") | |
| models_info = {"generic pretrain model": ["skytnt/midi-model", ""], | |
| "j-pop finetune model": ["skytnt/midi-model-ft", "jpop/"], | |
| "touhou finetune model": ["skytnt/midi-model-ft", "touhou/"], | |
| } | |
| models = {} | |
| tokenizer = MIDITokenizer() | |
| providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] | |
| for name, (repo_id, path) in models_info.items(): | |
| model_base_path = hf_hub_download_retry(repo_id=repo_id, filename=f"{path}onnx/model_base.onnx") | |
| model_token_path = hf_hub_download_retry(repo_id=repo_id, filename=f"{path}onnx/model_token.onnx") | |
| model_base = rt.InferenceSession(model_base_path, providers=providers) | |
| model_token = rt.InferenceSession(model_token_path, providers=providers) | |
| models[name] = [model_base, model_token] | |
| load_javascript() | |
| app = gr.Blocks() | |
| with app: | |
| gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Midi Composer</h1>") | |
| gr.Markdown("\n\n" | |
| "Midi event transformer for music generation\n\n" | |
| "Demo for [SkyTNT/midi-model](https://github.com/SkyTNT/midi-model)\n\n" | |
| "[Open In Colab]" | |
| "(https://colab.research.google.com/github/SkyTNT/midi-model/blob/main/demo.ipynb)" | |
| " for faster running and longer generation\n\n" | |
| "**Update v1.2**: Optimise the tokenizer and dataset" | |
| ) | |
| js_msg = gr.Textbox(elem_id="msg_receiver", visible=False) | |
| js_msg.change(None, [js_msg], [], js=""" | |
| (msg_json) =>{ | |
| let msgs = JSON.parse(msg_json); | |
| executeCallbacks(msgReceiveCallbacks, msgs); | |
| return []; | |
| } | |
| """) | |
| input_model = gr.Dropdown(label="select model", choices=list(models.keys()), | |
| type="value", value=list(models.keys())[0]) | |
| tab_select = gr.State(value=0) | |
| with gr.Tabs(): | |
| with gr.TabItem("instrument prompt") as tab1: | |
| input_instruments = gr.Dropdown(label="🪗instruments (auto if empty)", choices=list(patch2number.keys()), | |
| multiselect=True, max_choices=15, type="value") | |
| input_drum_kit = gr.Dropdown(label="🥁drum kit", choices=list(drum_kits2number.keys()), type="value", | |
| value="None") | |
| input_bpm = gr.Slider(label="BPM (beats per minute, auto if 0)", minimum=0, maximum=255, | |
| step=1, | |
| value=0) | |
| example1 = gr.Examples([ | |
| [[], "None"], | |
| [["Acoustic Grand"], "None"], | |
| [['Acoustic Grand', 'SynthStrings 2', 'SynthStrings 1', 'Pizzicato Strings', | |
| 'Pad 2 (warm)', 'Tremolo Strings', 'String Ensemble 1'], "Orchestra"], | |
| [['Trumpet', 'Oboe', 'Trombone', 'String Ensemble 1', 'Clarinet', | |
| 'French Horn', 'Pad 4 (choir)', 'Bassoon', 'Flute'], "None"], | |
| [['Flute', 'French Horn', 'Clarinet', 'String Ensemble 2', 'English Horn', 'Bassoon', | |
| 'Oboe', 'Pizzicato Strings'], "Orchestra"], | |
| [['Electric Piano 2', 'Lead 5 (charang)', 'Electric Bass(pick)', 'Lead 2 (sawtooth)', | |
| 'Pad 1 (new age)', 'Orchestra Hit', 'Cello', 'Electric Guitar(clean)'], "Standard"], | |
| [["Electric Guitar(clean)", "Electric Guitar(muted)", "Overdriven Guitar", "Distortion Guitar", | |
| "Electric Bass(finger)"], "Standard"] | |
| ], [input_instruments, input_drum_kit]) | |
| with gr.TabItem("midi prompt") as tab2: | |
| input_midi = gr.File(label="input midi", file_types=[".midi", ".mid"], type="binary") | |
| input_midi_events = gr.Slider(label="use first n midi events as prompt", minimum=1, maximum=512, | |
| step=1, | |
| value=128) | |
| example2 = gr.Examples([[file, 128] for file in glob.glob("example/*.mid")], | |
| [input_midi, input_midi_events]) | |
| tab1.select(lambda: 0, None, tab_select, queue=False) | |
| tab2.select(lambda: 1, None, tab_select, queue=False) | |
| input_gen_events = gr.Slider(label="generate n midi events", minimum=1, maximum=opt.max_gen, | |
| step=1, value=opt.max_gen // 2) | |
| with gr.Accordion("options", open=False): | |
| input_temp = gr.Slider(label="temperature", minimum=0.1, maximum=1.2, step=0.01, value=1) | |
| input_top_p = gr.Slider(label="top p", minimum=0.1, maximum=1, step=0.01, value=0.98) | |
| input_top_k = gr.Slider(label="top k", minimum=1, maximum=128, step=1, value=20) | |
| input_allow_cc = gr.Checkbox(label="allow midi cc event", value=True) | |
| example3 = gr.Examples([[1, 0.98, 20], [1, 0.98, 12]], [input_temp, input_top_p, input_top_k]) | |
| run_btn = gr.Button("generate", variant="primary") | |
| stop_btn = gr.Button("stop and output") | |
| output_midi_seq = gr.State() | |
| output_midi_visualizer = gr.HTML(elem_id="midi_visualizer_container") | |
| output_audio = gr.Audio(label="output audio", format="mp3", elem_id="midi_audio") | |
| output_midi = gr.File(label="output midi", file_types=[".mid"]) | |
| run_event = run_btn.click(run, [input_model, tab_select, input_instruments, input_drum_kit, input_bpm, | |
| input_midi, input_midi_events, input_gen_events, input_temp, | |
| input_top_p, input_top_k, input_allow_cc], | |
| [output_midi_seq, output_midi, output_audio, js_msg], | |
| concurrency_limit=3) | |
| stop_btn.click(cancel_run, [output_midi_seq], [output_midi, output_audio, js_msg], cancels=run_event, queue=False) | |
| app.launch(server_port=opt.port, share=opt.share, inbrowser=True) | |