Total same code but in totally different behavior

[ElinLiu0]

Hi:
I'm using below Numpy.NET code to interpolate an NDarray:

public NDarray interpolate_f0(NDarray f0)
{
    int N = f0.shape[0];
    if (np.all(f0 <= 0))
    {
        return np.zeros_like(f0);
    }
    // 获取所有F0点位置
    NDarray nz = np.nonzero(f0 > 0)[0];
    // 如果只有一个非零点，返回该点的值
    if (nz.size == 1)
    {
        return np.full_like(f0, f0[nz[0]].GetData<float>()[0]);
    }
    // 每个位置在nz中的插入点索引
    NDarray pos = np.arange(N);
    NDarray idx = np.searchsorted(nz,pos,side: "left");

    // 限制idx范围
    int max_i = nz.size - 1;
    NDarray idx_low = np.clip(idx - 1, np.array(0), np.array(max_i));
    NDarray idx_high = np.clip(idx,np.array(0), np.array(max_i));

    // 左右端点在原数组中的索引
    NDarray left_idx = nz[idx_low];
    NDarray right_idx = nz[idx_high];

    // 两段的值
    NDarray left_val = f0[left_idx];
    NDarray right_val = f0[right_idx];

    // 计算分母避免除0
    NDarray demon = (right_idx - left_idx).astype(np.float64);
    demon = np.where(np.array(demon == np.array(0)),np.array(1),demon);

    // 插值比例
    NDarray t = (pos - left_idx).astype(np.float64) / demon;
    // 计算插值结果
    NDarray output = left_val + t * (right_val - left_val);

    // 原非零位置保持不变
    output = np.where(np.array(f0 > 0), f0, output);
    return output;
}

This code has the same Python version like:

def interpolate_numpy(f0: np.ndarray) -> np.ndarray:
    """
    输入 f0：1D numpy array
    输出：对 f0 中零值区间按照两端非零值做线性插值后的新数组
    """
    N = f0.shape[0]

    # 全部<=0，直接返回全零
    if np.all(f0 <= 0):
        return np.zeros_like(f0)

    # 找到所有>0的位置
    nz = np.nonzero(f0 > 0)[0]
    # 只有一个非零点，全部填该值
    if nz.size == 1:
        return np.full_like(f0, f0[nz[0]])

    # 每个位置在 nz 中的插入点索引
    pos = np.arange(N)
    idx = np.searchsorted(nz, pos, side='left')

    # 限制 idx_low, idx_high 到合法范围 [0, nz.size-1]
    max_i = nz.size - 1
    idx_low  = np.clip(idx - 1, 0, max_i)
    idx_high = np.clip(idx,     0, max_i)

    # 左右端点在原数组中的索引
    left_idx  = nz[idx_low]
    right_idx = nz[idx_high]

    # 两端的值
    left_val  = f0[left_idx]
    right_val = f0[right_idx]

    # 计算分母并避免除0
    denom = (right_idx - left_idx).astype(float)
    denom[denom == 0] = 1.0

    # 插值比例
    t = (pos - left_idx).astype(float) / denom

    # 线性插值
    out = left_val + t * (right_val - left_val)

    # 原非零位置保留原值
    out[f0 > 0] = f0[f0 > 0]

    return out

It should like: array([363. , 363. , 363. , ..., 76.3, 76.3, 76.3])
But it's actually got: Interpolated F0: [1.87585069e-309 1.17119999e+171 5.22741680e-037 ... nan
nan nan]

To make reproduce,please use library pyworld with below code:

f0,t = pyworld.dio(
    audio.astype(np.double),
    fs = 16000,
    f0_floor = 30,
    f0_ceil = 8000,
    frame_period = 1000 * 160 / 16000
)
f0:np.ndarray = pyworld.stonemask(audio.astype(np.double), f0, t, 16000)
for index, pitch in enumerate(f0):
    f0[index] = round(pitch, 1)

You can use any audio that you have and readed in librosa with sample rate 16000.

[ElinLiu0]

Either below too:
C# Version:

private NDarray resize_f0(NDarray x,int target_len)
{
    NDarray source = np.array(x);
    source = np.where(source < 0.001,np.array(np.nan),source);
    var xp = np.arange(start: 0, stop: source.size).GetData<float>();
    var fp = source.GetData<float>();
    IReadOnlyCollection<float> xp_collection = xp;
    IReadOnlyCollection<float> fp_collection = fp;
    NDarray target = np.interp(x: np.arange(start: 0, stop: source.size * target_len,step: source.size / target_len),xp:xp_collection,fp:fp_collection);
    NDarray res = np.nan_to_num(target);
    return res;
}

Python Version:

def resize_f0(
    x : np.ndarray,
    target_len : int
) -> np.ndarray:
    """
    对F0进行重塑
    Args:
        x (np.ndarray): 输入F0
        target_len (int): 目标长度
    Returns:
        np.ndarray: 重塑后的F0
    """
    source:np.ndarray = np.array(x)
    source[source < 0.001] = np.nan
    target:np.ndarray = np.interp(
        np.arange(0, len(source) * target_len, len(source)) / target_len,
        np.arange(0, len(source)),
        source,
    )
    res:np.ndarray = np.nan_to_num(target)
    return res

[ElinLiu0]

Looks like different input let it be i guess,close first : (