numpy - Numba：已检测到从 nopython 编译路径回退到对象模式编译路径

Question

我正在尝试使用 numba @jit 编译 python 函数。在参数列表中，N1,N2,N3 是整数，FX,FY,FZ 是具有 float64 数据类型的 numpy 矩阵，inx,...,ground 是具有 int64 数据类型的 numpy 数组，而 inb,res 具有 int64 数据类型。在下面的函数中，spdiags、lil_matrix、csx_matrix 来自 scipy.sparse 库。

@jit
def ThreeDChannel_Differentiate1(N1,N2,N3,FX,FY,FZ,
                                 inx,iny,inz,A0,AN,AS,AE,AW,AA,AG,east,
                                 west,north,south,air,ground,inb,res):

    data = np.zeros((7,(N1*N2*(N3-2))))
    diags = np.array([-N1*(N3-2), -N1, -1, 0, 1, N1, N1*(N3-2)])
    M = spdiags(data, diags, N1*N2*(N3-2), N1*N2*(N3-2))
    #M = sp.csc_matrix(M)
    M = sp.lil_matrix(M)

    for k in inz-1:
        for j in inx:
            for i in iny:
                FY0 = FY[i,j,k+1]
                FYN = FY[north[i],j,k+1]
                FYS = FY[south[i],j,k+1]

                FX0 = FX[i,j,k+1]
                FXE = FX[i,east[j],k+1]
                FXW = FX[i,west[j],k+1]

                FZ0 = FZ[i,j,k+1]
                FZA = FZ[i,j,air[k]]
                FZG = FZ[i,j,ground[k]]

                if inb == 1:
                    M[A0[i,j,k],A0[i,j,k]] = 1/FY0*(FYN/(FY0+FYN)-FYS/(FYS+FY0))

                if inb == 2:
                    M[A0[i,j,k],A0[i,j,k]] = 1/FX0*(FXE/(FX0+FXE)-FXW/(FXW+FX0))

                if inb == 3:
                    M[A0[i,j,k],A0[i,j,k]] = 1/FZ0*(FZA/(FZ0+FZA)-FZG/(FZG+FZ0))

                if inb == 1:
                    M[A0[i,j,k],AN.ravel(order='F')[A0[i,j,k]]] = 1/FY0*FY0/(FY0+FYN)
                    M[A0[i,j,k],AS.ravel(order='F')[A0[i,j,k]]] = -1/FY0*FY0/(FY0+FYS)

                if inb == 2:
                    M[A0[i,j,k],AE.ravel(order='F')[A0[i,j,k]]] = 1/FX0*FX0/(FX0+FXE)
                    M[A0[i,j,k],AW.ravel(order='F')[A0[i,j,k]]] = -1/FX0*FX0/(FX0+FXW)

                # Account for wall Dirichlet (no-slip) condition in the wall-normal direction
                if inb == 3:
                    if AG.ravel(order='F')[A0[i,j,k]] >= 0:
                        M[A0[i,j,k],AG.ravel(order='F')[A0[i,j,k]]] = -1/FZ0*FZ0/(FZ0+FZG)
                    else:
                        M[A0[i,j,k],A0[i,j,k]] = 1/FZ0*(FZA/(FZ0+FZA)-FZG/(FZG+FZ0) + \
                                                 FZ0/(FZG+FZ0))

                    if AA.ravel(order='F')[A0[i,j,k]] < N1*N2*(N3-2):
                        M[A0[i,j,k],AA.ravel(order='F')[A0[i,j,k]]] = 1/FZ0*FZ0/(FZ0+FZA)

                    else:
                        M[A0[i,j,k],A0[i,j,k]] = 1/FZ0*(FZA/(FZ0+FZA)-FZG/(FZG+FZ0) - \
                                                 FZ0/(FZA+FZ0))

    M = sp.csc_matrix(M, copy=False)

    return M

如果我调用上面的函数，我会得到以下错误。此外，我收到与 spdiag 对象相关的警告。我认为 Numba 不支持 scipy 对象。但以下警告与 for 循环有关。这个错误与什么有关？

File "<ipython-input-5-66816b729db2>", line 13:
def ThreeDChannel_Differentiate1(N1,N2,N3,FX,FY,FZ,
    <source elided>

    for k in inz-1:
    ^

  self.func_ir.loc))
/home/suraj/anaconda3/lib/python3.7/site-packages/numba/compiler.py:734: NumbaDeprecationWarning: 
Fall-back from the nopython compilation path to the object mode compilation path has been detected, this is deprecated behaviour.

谢谢你。

标签： numpyscipynumba