import argparse import glob import json import os.path import time import gradio as gr import numpy as np import onnxruntime as rt import tqdm from huggingface_hub import hf_hub_download import MIDI from midi_synthesizer import MidiSynthesizer from midi_tokenizer import MIDITokenizer MAX_SEED = np.iinfo(np.int32).max 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, generator=None): if generator is None: generator = np.random 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([generator.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, generator=None): tokenizer = model[2] if disable_channels is not None: disable_channels = [tokenizer.parameter_ids["channel"][c] for c in disable_channels] else: disable_channels = [] if generator is None: generator = np.random 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, generator) 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} def send_msgs(msgs): return json.dumps(msgs) def run(model_name, tab, mid_seq, instruments, drum_kit, bpm, time_sig, key_sig, mid, midi_events, reduce_cc_st, remap_track_channel, add_default_instr, remove_empty_channels, seed, seed_rand, gen_events, temp, top_p, top_k, allow_cc): model = models[model_name] tokenizer = model[2] bpm = int(bpm) if time_sig == "auto": time_sig = None time_sig_nn = 4 time_sig_dd = 2 else: time_sig_nn, time_sig_dd = time_sig.split('/') time_sig_nn = int(time_sig_nn) time_sig_dd = {2: 1, 4: 2, 8: 3}[int(time_sig_dd)] if key_sig == 0: key_sig = None key_sig_sf = 0 key_sig_mi = 0 else: key_sig = (key_sig - 1) key_sig_sf = key_sig // 2 - 7 key_sig_mi = key_sig % 2 gen_events = int(gen_events) max_len = gen_events if seed_rand: seed = np.random.randint(0, MAX_SEED) generator = np.random.RandomState(seed) 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 tokenizer.version == "v2": if time_sig is not None: mid.append(tokenizer.event2tokens(["time_signature", 0, 0, 0, time_sig_nn - 1, time_sig_dd - 1])) if key_sig is not None: mid.append(tokenizer.event2tokens(["key_signature", 0, 0, 0, key_sig_sf + 7, key_sig_mi])) 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 + 1, 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 tab == 1 and mid is not None: eps = 4 if reduce_cc_st else 0 mid = tokenizer.tokenize(MIDI.midi2score(mid), cc_eps=eps, tempo_eps=eps, remap_track_channel=remap_track_channel, add_default_instr=add_default_instr, remove_empty_channels=remove_empty_channels) mid = np.asarray(mid, dtype=np.int64) mid = mid[:int(midi_events)] mid_seq = [] for token_seq in mid: mid_seq.append(token_seq.tolist()) elif tab == 2 and mid_seq is not None: mid = np.asarray(mid_seq, dtype=np.int64) else: mid_seq = [] mid = None if mid is not None: max_len += len(mid) events = [tokenizer.tokens2event(tokens) for tokens in mid_seq] if tab == 2: init_msgs = [create_msg("visualizer_continue", tokenizer.version)] else: init_msgs = [create_msg("visualizer_clear", tokenizer.version), create_msg("visualizer_append", events)] yield mid_seq, None, None, seed, send_msgs(init_msgs) midi_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, generator=generator) events = [] t = time.time() + 1 for i, token_seq in enumerate(midi_generator): token_seq = token_seq.tolist() mid_seq.append(token_seq) events.append(tokenizer.tokens2event(token_seq)) ct = time.time() if ct - t > 0.5: yield mid_seq, None, None, seed, send_msgs( [create_msg("visualizer_append", events), create_msg("progress", [i + 1, gen_events])]) t = ct events = [] mid = tokenizer.detokenize(mid_seq) with open(f"output.mid", 'wb') as f: f.write(MIDI.score2midi(mid)) audio = synthesizer.synthesis(MIDI.score2opus(mid)) events = [tokenizer.tokens2event(tokens) for tokens in mid_seq] yield mid_seq, "output.mid", (44100, audio), seed, send_msgs([create_msg("visualizer_end", events)]) def cancel_run(model_name, mid_seq): if mid_seq is None: return None, None, [] tokenizer = models[model_name][2] mid = tokenizer.detokenize(mid_seq) with open(f"output.mid", 'wb') as f: f.write(MIDI.score2midi(mid)) audio = synthesizer.synthesis(MIDI.score2opus(mid)) events = [tokenizer.tokens2event(tokens) for tokens in mid_seq] return "output.mid", (44100, audio), send_msgs([create_msg("visualizer_end", events)]) 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" template_response_ori = gr.routes.templates.TemplateResponse def template_response(*args, **kwargs): res = template_response_ori(*args, **kwargs) res.body = res.body.replace( b'', f'{javascript}'.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 def get_tokenizer(config_name): tv, size = config_name.split("-") tv = tv[1:] if tv[-1] == "o": o = True tv = tv[:-1] else: o = False if tv not in ["v1", "v2"]: raise ValueError(f"Unknown tokenizer version {tv}") tokenizer = MIDITokenizer(tv) tokenizer.set_optimise_midi(o) return tokenizer 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()} key_signatures = ['C♭', 'A♭m', 'G♭', 'E♭m', 'D♭', 'B♭m', 'A♭', 'Fm', 'E♭', 'Cm', 'B♭', 'Gm', 'F', 'Dm', 'C', 'Am', 'G', 'Em', 'D', 'Bm', 'A', 'F♯m', 'E', 'C♯m', 'B', 'G♯m', 'F♯', 'D♯m', 'C♯', 'A♯m'] 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") synthesizer = MidiSynthesizer(soundfont_path) models_info = {"generic pretrain model (tv2o-large) by asigalov61": ["asigalov61/Music-Llama", "", "tv2o-large"], "generic pretrain model (tv2o-medium) by asigalov61": ["asigalov61/Music-Llama-Medium", "", "tv2o-medium"], "generic pretrain model (tv1-medium) by skytnt": ["skytnt/midi-model", "", "tv1-medium"], "j-pop finetune model (tv1-medium) by skytnt": ["skytnt/midi-model-ft", "jpop/", "tv1-medium"], "touhou finetune model (tv1-medium) by skytnt": ["skytnt/midi-model-ft", "touhou/", "tv1-medium"], } models = {} providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] for name, (repo_id, path, config) 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) tokenizer = get_tokenizer(config) models[name] = [model_base, model_token, tokenizer] load_javascript() app = gr.Blocks() with app: gr.Markdown("