inference/infer.py

import os
import sys
sys.path.append(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'xcodec_mini_infer'))
sys.path.append(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'xcodec_mini_infer', 'descriptaudiocodec'))
import argparse
import torch
import numpy as np
import json
from omegaconf import OmegaConf
import torchaudio
from torchaudio.transforms import Resample
import soundfile as sf

import uuid
from tqdm import tqdm
from einops import rearrange
from codecmanipulator import CodecManipulator
from mmtokenizer import _MMSentencePieceTokenizer
from transformers import AutoTokenizer, AutoModelForCausalLM, LogitsProcessor, LogitsProcessorList
import glob
import time
import copy
from collections import Counter
from models.soundstream_hubert_new import SoundStream
from vocoder import build_codec_model, process_audio
from post_process_audio import replace_low_freq_with_energy_matched
import re


parser = argparse.ArgumentParser()
# Model Configuration:
parser.add_argument("--stage1_model", type=str, default="m-a-p/YuE-s1-7B-anneal-en-cot", help="The model checkpoint path or identifier for the Stage 1 model.")
parser.add_argument("--stage2_model", type=str, default="m-a-p/YuE-s2-1B-general", help="The model checkpoint path or identifier for the Stage 2 model.")
parser.add_argument("--max_new_tokens", type=int, default=3000, help="The maximum number of new tokens to generate in one pass during text generation.")
parser.add_argument("--run_n_segments", type=int, default=2, help="The number of segments to process during the generation.")
parser.add_argument("--stage2_batch_size", type=int, default=4, help="The batch size used in Stage 2 inference.")
# Prompt
parser.add_argument("--genre_txt", type=str, required=True, help="The file path to a text file containing genre tags that describe the musical style or characteristics (e.g., instrumental, genre, mood, vocal timbre, vocal gender). This is used as part of the generation prompt.")
parser.add_argument("--lyrics_txt", type=str, required=True, help="The file path to a text file containing the lyrics for the music generation. These lyrics will be processed and split into structured segments to guide the generation process.")
parser.add_argument("--use_audio_prompt", action="store_true", help="If set, the model will use an audio file as a prompt during generation. The audio file should be specified using --audio_prompt_path.")
parser.add_argument("--audio_prompt_path", type=str, default="", help="The file path to an audio file to use as a reference prompt when --use_audio_prompt is enabled.")
parser.add_argument("--prompt_start_time", type=float, default=0.0, help="The start time in seconds to extract the audio prompt from the given audio file.")
parser.add_argument("--prompt_end_time", type=float, default=30.0, help="The end time in seconds to extract the audio prompt from the given audio file.")
# Output 
parser.add_argument("--output_dir", type=str, default="./output", help="The directory where generated outputs will be saved.")
parser.add_argument("--keep_intermediate", action="store_true", help="If set, intermediate outputs will be saved during processing.")
parser.add_argument("--disable_offload_model", action="store_true", help="If set, the model will not be offloaded from the GPU to CPU after Stage 1 inference.")
parser.add_argument("--cuda_idx", type=int, default=0)
# Config for xcodec and upsampler
parser.add_argument('--basic_model_config', default='./xcodec_mini_infer/final_ckpt/config.yaml', help='YAML files for xcodec configurations.')
parser.add_argument('--resume_path', default='./xcodec_mini_infer/final_ckpt/ckpt_00360000.pth', help='Path to the xcodec checkpoint.')
parser.add_argument('--config_path', type=str, default='./xcodec_mini_infer/decoders/config.yaml', help='Path to Vocos config file.')
parser.add_argument('--vocal_decoder_path', type=str, default='./xcodec_mini_infer/decoders/decoder_131000.pth', help='Path to Vocos decoder weights.')
parser.add_argument('--inst_decoder_path', type=str, default='./xcodec_mini_infer/decoders/decoder_151000.pth', help='Path to Vocos decoder weights.')
parser.add_argument('-r', '--rescale', action='store_true', help='Rescale output to avoid clipping.')


args = parser.parse_args()
if args.use_audio_prompt and not args.audio_prompt_path:
    raise FileNotFoundError("Please offer audio prompt filepath using '--audio_prompt_path', when you enable 'use_audio_prompt'!")
stage1_model = args.stage1_model
stage2_model = args.stage2_model
cuda_idx = args.cuda_idx
max_new_tokens = args.max_new_tokens
stage1_output_dir = os.path.join(args.output_dir, f"stage1")
stage2_output_dir = stage1_output_dir.replace('stage1', 'stage2')
os.makedirs(stage1_output_dir, exist_ok=True)
os.makedirs(stage2_output_dir, exist_ok=True)

# load tokenizer and model
device = torch.device(f"cuda:{cuda_idx}" if torch.cuda.is_available() else "cpu")

# Now you can use `device` to move your tensors or models to the GPU (if available)
print(f"Using device: {device}")

mmtokenizer = _MMSentencePieceTokenizer("./mm_tokenizer_v0.2_hf/tokenizer.model")
model = AutoModelForCausalLM.from_pretrained(
    stage1_model, 
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2", # To enable flashattn, you have to install flash-attn
    )
model.to(device)
model.eval()

codectool = CodecManipulator("xcodec", 0, 1)
codectool_stage2 = CodecManipulator("xcodec", 0, 8)
model_config = OmegaConf.load(args.basic_model_config)
codec_model = eval(model_config.generator.name)(**model_config.generator.config).to(device)
parameter_dict = torch.load(args.resume_path, map_location='cpu')
codec_model.load_state_dict(parameter_dict['codec_model'])
codec_model.to(device)
codec_model.eval()

class BlockTokenRangeProcessor(LogitsProcessor):
    def __init__(self, start_id, end_id):
        self.blocked_token_ids = list(range(start_id, end_id))

    def __call__(self, input_ids, scores):
        scores[:, self.blocked_token_ids] = -float("inf")
        return scores

def load_audio_mono(filepath, sampling_rate=16000):
    audio, sr = torchaudio.load(filepath)
    # Convert to mono
    audio = torch.mean(audio, dim=0, keepdim=True)
    # Resample if needed
    if sr != sampling_rate:
        resampler = Resample(orig_freq=sr, new_freq=sampling_rate)
        audio = resampler(audio)
    return audio

def split_lyrics(lyrics):
    pattern = r"\[(\w+)\](.*?)\n(?=\[|\Z)"
    segments = re.findall(pattern, lyrics, re.DOTALL)
    structured_lyrics = [f"[{seg[0]}]\n{seg[1].strip()}\n\n" for seg in segments]
    return structured_lyrics

# Call the function and print the result
stage1_output_set = []
# Tips:
# genre tags support instrumental，genre，mood，vocal timbr and vocal gender
# all kinds of tags are needed
with open(args.genre_txt) as f:
    genres = f.read().strip()
with open(args.lyrics_txt) as f:
    lyrics = split_lyrics(f.read())
# intruction
full_lyrics = "\n".join(lyrics)
prompt_texts = [f"Generate music from the given lyrics segment by segment.\n[Genre] {genres}\n{full_lyrics}"]
prompt_texts += lyrics


random_id = uuid.uuid4()
output_seq = None
# Here is suggested decoding config
top_p = 0.93
temperature = 1.0
repetition_penalty = 1.2
# special tokens
start_of_segment = mmtokenizer.tokenize('[start_of_segment]')
end_of_segment = mmtokenizer.tokenize('[end_of_segment]')
# Format text prompt
run_n_segments = min(args.run_n_segments+1, len(lyrics))
for i, p in enumerate(tqdm(prompt_texts[:run_n_segments])):
    section_text = p.replace('[start_of_segment]', '').replace('[end_of_segment]', '')
    guidance_scale = 1.5 if i <=1 else 1.2
    if i==0:
        continue
    if i==1:
        if args.use_audio_prompt:
            audio_prompt = load_audio_mono(args.audio_prompt_path)
            audio_prompt.unsqueeze_(0)
            with torch.no_grad():
                raw_codes = codec_model.encode(audio_prompt.to(device), target_bw=0.5)
            raw_codes = raw_codes.transpose(0, 1)
            raw_codes = raw_codes.cpu().numpy().astype(np.int16)
            # Format audio prompt
            code_ids = codectool.npy2ids(raw_codes[0])
            audio_prompt_codec = code_ids[int(args.prompt_start_time *50): int(args.prompt_end_time *50)] # 50 is tps of xcodec
            audio_prompt_codec_ids = [mmtokenizer.soa] + codectool.sep_ids + audio_prompt_codec + [mmtokenizer.eoa]
            sentence_ids = mmtokenizer.tokenize("[start_of_reference]") +  audio_prompt_codec_ids + mmtokenizer.tokenize("[end_of_reference]")
            head_id = mmtokenizer.tokenize(prompt_texts[0]) + sentence_ids
        else:
            head_id = mmtokenizer.tokenize(prompt_texts[0])
        prompt_ids = head_id + start_of_segment + mmtokenizer.tokenize(section_text) + [mmtokenizer.soa] + codectool.sep_ids
    else:
        prompt_ids = end_of_segment + start_of_segment + mmtokenizer.tokenize(section_text) + [mmtokenizer.soa] + codectool.sep_ids

    prompt_ids = torch.as_tensor(prompt_ids).unsqueeze(0).to(device) 
    input_ids = torch.cat([raw_output, prompt_ids], dim=1) if i > 1 else prompt_ids
    # Use window slicing in case output sequence exceeds the context of model
    max_context = 16384-max_new_tokens-1
    if input_ids.shape[-1] > max_context:
        print(f'Section {i}: output length {input_ids.shape[-1]} exceeding context length {max_context}, now using the last {max_context} tokens.')
        input_ids = input_ids[:, -(max_context):]
    with torch.no_grad():
        output_seq = model.generate(
            input_ids=input_ids, 
            max_new_tokens=max_new_tokens, 
            min_new_tokens=100, 
            do_sample=True, 
            top_p=top_p,
            temperature=temperature, 
            repetition_penalty=repetition_penalty, 
            eos_token_id=mmtokenizer.eoa,
            pad_token_id=mmtokenizer.eoa,
            logits_processor=LogitsProcessorList([BlockTokenRangeProcessor(0, 32002), BlockTokenRangeProcessor(32016, 32016)]),
            guidance_scale=guidance_scale,
            )
        if output_seq[0][-1].item() != mmtokenizer.eoa:
            tensor_eoa = torch.as_tensor([[mmtokenizer.eoa]]).to(model.device)
            output_seq = torch.cat((output_seq, tensor_eoa), dim=1)
    if i > 1:
        raw_output = torch.cat([raw_output, prompt_ids, output_seq[:, input_ids.shape[-1]:]], dim=1)
    else:
        raw_output = output_seq

# save raw output and check sanity
ids = raw_output[0].cpu().numpy()
soa_idx = np.where(ids == mmtokenizer.soa)[0].tolist()
eoa_idx = np.where(ids == mmtokenizer.eoa)[0].tolist()
if len(soa_idx)!=len(eoa_idx):
    raise ValueError(f'invalid pairs of soa and eoa, Num of soa: {len(soa_idx)}, Num of eoa: {len(eoa_idx)}')

vocals = []
instrumentals = []
range_begin = 1 if args.use_audio_prompt else 0
for i in range(range_begin, len(soa_idx)):
    codec_ids = ids[soa_idx[i]+1:eoa_idx[i]]
    if codec_ids[0] == 32016:
        codec_ids = codec_ids[1:]
    codec_ids = codec_ids[:2 * (codec_ids.shape[0] // 2)]
    vocals_ids = codectool.ids2npy(rearrange(codec_ids,"(n b) -> b n", b=2)[0])
    vocals.append(vocals_ids)
    instrumentals_ids = codectool.ids2npy(rearrange(codec_ids,"(n b) -> b n", b=2)[1])
    instrumentals.append(instrumentals_ids)
vocals = np.concatenate(vocals, axis=1)
instrumentals = np.concatenate(instrumentals, axis=1)
vocal_save_path = os.path.join(stage1_output_dir, f"cot_{genres.replace(' ', '-')}_tp{top_p}_T{temperature}_rp{repetition_penalty}_maxtk{max_new_tokens}_vocal_{random_id}".replace('.', '@')+'.npy')
inst_save_path = os.path.join(stage1_output_dir, f"cot_{genres.replace(' ', '-')}_tp{top_p}_T{temperature}_rp{repetition_penalty}_maxtk{max_new_tokens}_instrumental_{random_id}".replace('.', '@')+'.npy')
np.save(vocal_save_path, vocals)
np.save(inst_save_path, instrumentals)
stage1_output_set.append(vocal_save_path)
stage1_output_set.append(inst_save_path)


# offload model
if not args.disable_offload_model:
    model.cpu()
    del model
    torch.cuda.empty_cache()

print("Stage 2 inference...")
model_stage2 = AutoModelForCausalLM.from_pretrained(
    stage2_model, 
    torch_dtype=torch.float16,
    attn_implementation="flash_attention_2"
    )
model_stage2.to(device)
model_stage2.eval()

def stage2_generate(model, prompt, batch_size=16):
    codec_ids = codectool.unflatten(prompt, n_quantizer=1)
    codec_ids = codectool.offset_tok_ids(
                    codec_ids, 
                    global_offset=codectool.global_offset, 
                    codebook_size=codectool.codebook_size, 
                    num_codebooks=codectool.num_codebooks, 
                ).astype(np.int32)
    
    # Prepare prompt_ids based on batch size or single input
    if batch_size > 1:
        codec_list = []
        for i in range(batch_size):
            idx_begin = i * 300
            idx_end = (i + 1) * 300
            codec_list.append(codec_ids[:, idx_begin:idx_end])

        codec_ids = np.concatenate(codec_list, axis=0)
        prompt_ids = np.concatenate(
            [
                np.tile([mmtokenizer.soa, mmtokenizer.stage_1], (batch_size, 1)),
                codec_ids,
                np.tile([mmtokenizer.stage_2], (batch_size, 1)),
            ],
            axis=1
        )
    else:
        prompt_ids = np.concatenate([
            np.array([mmtokenizer.soa, mmtokenizer.stage_1]),
            codec_ids.flatten(),  # Flatten the 2D array to 1D
            np.array([mmtokenizer.stage_2])
        ]).astype(np.int32)
        prompt_ids = prompt_ids[np.newaxis, ...]

    codec_ids = torch.as_tensor(codec_ids).to(device)
    prompt_ids = torch.as_tensor(prompt_ids).to(device)
    len_prompt = prompt_ids.shape[-1]
    
    block_list = LogitsProcessorList([BlockTokenRangeProcessor(0, 46358), BlockTokenRangeProcessor(53526, mmtokenizer.vocab_size)])

    # Teacher forcing generate loop
    for frames_idx in range(codec_ids.shape[1]):
        cb0 = codec_ids[:, frames_idx:frames_idx+1]
        prompt_ids = torch.cat([prompt_ids, cb0], dim=1)
        input_ids = prompt_ids

        with torch.no_grad():
            stage2_output = model.generate(input_ids=input_ids, 
                min_new_tokens=7,
                max_new_tokens=7,
                eos_token_id=mmtokenizer.eoa,
                pad_token_id=mmtokenizer.eoa,
                logits_processor=block_list,
            )
        
        assert stage2_output.shape[1] - prompt_ids.shape[1] == 7, f"output new tokens={stage2_output.shape[1]-prompt_ids.shape[1]}"
        prompt_ids = stage2_output

    # Return output based on batch size
    if batch_size > 1:
        output = prompt_ids.cpu().numpy()[:, len_prompt:]
        output_list = [output[i] for i in range(batch_size)]
        output = np.concatenate(output_list, axis=0)
    else:
        output = prompt_ids[0].cpu().numpy()[len_prompt:]

    return output

def stage2_inference(model, stage1_output_set, stage2_output_dir, batch_size=4):
    stage2_result = []
    for i in tqdm(range(len(stage1_output_set))):
        output_filename = os.path.join(stage2_output_dir, os.path.basename(stage1_output_set[i]))
        
        if os.path.exists(output_filename):
            print(f'{output_filename} stage2 has done.')
            continue
        
        # Load the prompt
        prompt = np.load(stage1_output_set[i]).astype(np.int32)
        
        # Only accept 6s segments
        output_duration = prompt.shape[-1] // 50 // 6 * 6
        num_batch = output_duration // 6
        
        if num_batch <= batch_size:
            # If num_batch is less than or equal to batch_size, we can infer the entire prompt at once
            output = stage2_generate(model, prompt[:, :output_duration*50], batch_size=num_batch)
        else:
            # If num_batch is greater than batch_size, process in chunks of batch_size
            segments = []
            num_segments = (num_batch // batch_size) + (1 if num_batch % batch_size != 0 else 0)

            for seg in range(num_segments):
                start_idx = seg * batch_size * 300
                # Ensure the end_idx does not exceed the available length
                end_idx = min((seg + 1) * batch_size * 300, output_duration*50)  # Adjust the last segment
                current_batch_size = batch_size if seg != num_segments-1 or num_batch % batch_size == 0 else num_batch % batch_size
                segment = stage2_generate(
                    model,
                    prompt[:, start_idx:end_idx],
                    batch_size=current_batch_size
                )
                segments.append(segment)

            # Concatenate all the segments
            output = np.concatenate(segments, axis=0)
        
        # Process the ending part of the prompt
        if output_duration*50 != prompt.shape[-1]:
            ending = stage2_generate(model, prompt[:, output_duration*50:], batch_size=1)
            output = np.concatenate([output, ending], axis=0)
        output = codectool_stage2.ids2npy(output)

        # Fix invalid codes (a dirty solution, which may harm the quality of audio)
        # We are trying to find better one
        fixed_output = copy.deepcopy(output)
        for i, line in enumerate(output):
            for j, element in enumerate(line):
                if element < 0 or element > 1023:
                    counter = Counter(line)
                    most_frequant = sorted(counter.items(), key=lambda x: x[1], reverse=True)[0][0]
                    fixed_output[i, j] = most_frequant
        # save output
        np.save(output_filename, fixed_output)
        stage2_result.append(output_filename)
    return stage2_result

stage2_result = stage2_inference(model_stage2, stage1_output_set, stage2_output_dir, batch_size=args.stage2_batch_size)
print(stage2_result)
print('Stage 2 DONE.\n')
# convert audio tokens to audio
def save_audio(wav: torch.Tensor, path, sample_rate: int, rescale: bool = False):
    folder_path = os.path.dirname(path)
    if not os.path.exists(folder_path):
        os.makedirs(folder_path)
    limit = 0.99
    max_val = wav.abs().max()
    wav = wav * min(limit / max_val, 1) if rescale else wav.clamp(-limit, limit)
    torchaudio.save(str(path), wav, sample_rate=sample_rate, encoding='PCM_S', bits_per_sample=16)
# reconstruct tracks
recons_output_dir = os.path.join(args.output_dir, "recons")
recons_mix_dir = os.path.join(recons_output_dir, 'mix')
os.makedirs(recons_mix_dir, exist_ok=True)
tracks = []
for npy in stage2_result:
    codec_result = np.load(npy)
    decodec_rlt=[]
    with torch.no_grad():
        decoded_waveform = codec_model.decode(torch.as_tensor(codec_result.astype(np.int16), dtype=torch.long).unsqueeze(0).permute(1, 0, 2).to(device))
    decoded_waveform = decoded_waveform.cpu().squeeze(0)
    decodec_rlt.append(torch.as_tensor(decoded_waveform))
    decodec_rlt = torch.cat(decodec_rlt, dim=-1)
    save_path = os.path.join(recons_output_dir, os.path.splitext(os.path.basename(npy))[0] + ".mp3")
    tracks.append(save_path)
    save_audio(decodec_rlt, save_path, 16000)
# mix tracks
for inst_path in tracks:
    try:
        if (inst_path.endswith('.wav') or inst_path.endswith('.mp3')) \
            and 'instrumental' in inst_path:
            # find pair
            vocal_path = inst_path.replace('instrumental', 'vocal')
            if not os.path.exists(vocal_path):
                continue
            # mix
            recons_mix = os.path.join(recons_mix_dir, os.path.basename(inst_path).replace('instrumental', 'mixed'))
            vocal_stem, sr = sf.read(inst_path)
            instrumental_stem, _ = sf.read(vocal_path)
            mix_stem = (vocal_stem + instrumental_stem) / 1
            sf.write(recons_mix, mix_stem, sr)
    except Exception as e:
        print(e)

# vocoder to upsample audios
vocal_decoder, inst_decoder = build_codec_model(args.config_path, args.vocal_decoder_path, args.inst_decoder_path)
vocoder_output_dir = os.path.join(args.output_dir, 'vocoder')
vocoder_stems_dir = os.path.join(vocoder_output_dir, 'stems')
vocoder_mix_dir = os.path.join(vocoder_output_dir, 'mix')
os.makedirs(vocoder_mix_dir, exist_ok=True)
os.makedirs(vocoder_stems_dir, exist_ok=True)
for npy in stage2_result:
    if 'instrumental' in npy:
        # Process instrumental
        instrumental_output = process_audio(
            npy,
            os.path.join(vocoder_stems_dir, 'instrumental.mp3'),
            args.rescale,
            args,
            inst_decoder,
            codec_model
        )
    else:
        # Process vocal
        vocal_output = process_audio(
            npy,
            os.path.join(vocoder_stems_dir, 'vocal.mp3'),
            args.rescale,
            args,
            vocal_decoder,
            codec_model
        )
# mix tracks
try:
    mix_output = instrumental_output + vocal_output
    vocoder_mix = os.path.join(vocoder_mix_dir, os.path.basename(recons_mix))
    save_audio(mix_output, vocoder_mix, 44100, args.rescale)
    print(f"Created mix: {vocoder_mix}")
except RuntimeError as e:
    print(e)
    print(f"mix {vocoder_mix} failed! inst: {instrumental_output.shape}, vocal: {vocal_output.shape}")

# Post process
replace_low_freq_with_energy_matched(
    a_file=recons_mix,     # 16kHz
    b_file=vocoder_mix,     # 48kHz
    c_file=os.path.join(args.output_dir, os.path.basename(recons_mix)),
    cutoff_freq=5500.0
)