[Python] polars arr.get(i) replacement

I have a code

df.select([
pl.all().exclude("elapsed_time_linreg"),
pl.col("elapsed_time_linreg").arr.get(0).suffix("_slope"),
pl.col("elapsed_time_linreg").arr.get(1).suffix("_intercept"),
pl.col("elapsed_time_linreg").arr.get(2).suffix("_resid_std"),
])


which unpacked the result of a function

@jit
def linear_regression(session_np: np.ndarray) -> np.ndarray:
w = len(session_np)
x = np.arange(w)
sx = w ** 2 / 2
sy = np.sum(session_np)
sx2 = (w * (w + 1) * (2 * w + 1)) / 6
sxy = np.sum(x * session_np)
slope = (w * sxy - sx * sy) / (w * sx2 - sx**2)
intercept = (sy - slope * sx) / w
resids = session_np - (x * slope + intercept)
return slope, intercept, resids.std()


def get_linreg_aggs(session) -> np.ndarray:
return linear_regression(np.array(session))


df.select(
pl.col("elapsed_time")
.apply(utils.get_linreg_aggs)
.cast(pl.Float32)
.alias("elapsed_time_linreg")
)


which worked perfectly on polars 17.15, but since I upgraded it to the latest (0.18.4), it stopped working. Now I get the following error and have no idea how to fix this:

AttributeError: 'ExprArrayNameSpace' object has no attribute 'get'


Is there a way of fixing it without downgrading the version? Or even a more efficient way of creating a bunch of columns from this function?

I thought of the caching the computational result with lru-cache, but since polars uses multiprocessing, idk if it helps.

Continue reading...