3D CNN Forward Pass Implementation

def conv3d_forward(
    input_vol: list[list[list[list[float]]]],
    filters: list[list[list[list[list[float]]]]],
    biases: list[float],
    stride: int = 1,
    padding: int = 0,
) -> list[list[list[list[float]]]]:
    # input_vol: [C_in, D, H, W]
    # filters: [C_out, C_in, Fd, Fh, Fw]
    C_in = len(input_vol)
    D = len(input_vol[0])
    H = len(input_vol[0][0])
    W = len(input_vol[0][0][0])
    C_out = len(filters)
    Fd = len(filters[0][0])
    Fh = len(filters[0][0][0])
    Fw = len(filters[0][0][0][0])

    # Apply zero padding
    if padding > 0:
        D_p = D + 2 * padding
        H_p = H + 2 * padding
        W_p = W + 2 * padding
        padded = [[[[0.0] * W_p for _ in range(H_p)] for _ in range(D_p)] for _ in range(C_in)]
        for c in range(C_in):
            for d in range(D):
                for h in range(H):
                    for w in range(W):
                        padded[c][d + padding][h + padding][w + padding] = input_vol[c][d][h][w]
        input_vol = padded
        D, H, W = D_p, H_p, W_p

    out_D = (D - Fd) // stride + 1
    out_H = (H - Fh) // stride + 1
    out_W = (W - Fw) // stride + 1

    output = [[[[0.0] * out_W for _ in range(out_H)] for _ in range(out_D)] for _ in range(C_out)]

    for co in range(C_out):
        for d in range(out_D):
            for h in range(out_H):
                for w in range(out_W):
                    val = biases[co]
                    for ci in range(C_in):
                        for fd in range(Fd):
                            for fh in range(Fh):
                                for fw in range(Fw):
                                    val += (
                                        input_vol[ci][d * stride + fd][h * stride + fh][w * stride + fw]
                                        * filters[co][ci][fd][fh][fw]
                                    )
                    output[co][d][h][w] = round(val, 4)

    return output