unconditional_sample.py

"""
Modified from https://github.com/karpathy/nanoGPT/blob/master/sample.py
"""
from __future__ import annotations

import argparse
from pathlib import Path

import torch
import torch.nn.functional as F
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "3"

from train import MusicLLM  # 直接导入 MusicLLM 类
from utils import parse_yaml
from piano_transcription.utils import write_midi

def sample(args):
    # Arguments
    config_path = args.config
    ckpt_path = args.ckpt_path

    # Configs
    configs = parse_yaml(config_path)

    num_samples = 5  # Number of samples to draw
    max_new_codes = 1000  # Number of codes generated in each sample
    temperature = 1.0  # 控制随机性
    top_k = 200  # 保留 top-k 个概率最高的 tokens
    
    device = "cuda"
    sr = configs["sample_rate"]

    # 加载 MusicLLM 模型
    model = MusicLLM(configs)
    checkpoint = torch.load(ckpt_path, map_location=device)
    
    # 提取 state_dict 并加载 llm 参数
    state_dict = checkpoint["state_dict"]
    llm_state_dict = {k.replace("llm.", ""): v for k, v in state_dict.items() if k.startswith("llm.")}
    model.llm.load_state_dict(llm_state_dict)
    
    model.to(device)
    model.eval()

    # 初始音频 codes
    B = 1
    start_codes = torch.zeros(size=(B, 1), dtype=torch.long, device=device)

    # Sample
    for n in range(num_samples):
        with torch.no_grad():
            current_codes = start_codes.clone()
            for _ in range(max_new_codes):
                logits = model(current_codes)
                logits = logits[:, -1, :]
                logits = logits / temperature
                if top_k is not None:
                    v, idx = torch.topk(logits, min(top_k, logits.size(-1)))
                    logits[logits < v[:, [-1]]] = -float('Inf')
                probs = F.softmax(logits, dim=-1)
                next_code = torch.multinomial(probs, num_samples=1)
                current_codes = torch.cat([current_codes, next_code], dim=1)

            audio_codes = current_codes.unsqueeze(dim=2)
            events = model.codec.decode(audio_codes)

            results_dir = Path("results", Path(config_path).stem)
            results_dir.mkdir(parents=True, exist_ok=True)
            midi_path = Path(results_dir, f"sample_{n}.midi")
            Path(midi_path).parent.mkdir(parents=True, exist_ok=True)
            write_midi(events, midi_path)
            print(f"Generated sample {n} saved to {midi_path}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True, help="Path of config yaml.")
    parser.add_argument('--ckpt_path', type=str, required=True)
    args = parser.parse_args()

    sample(args)