convert strings to others fast

def parse_dtypes(df, exclude=[]):
    str_cols = [x for x, y in df.schema.items() if y == pl.String and x not in exclude]
    try_casts = df.select(
        pl.struct(pl.all()).alias("original"),
        pl.struct(
            pl.coalesce(
                pl.col(col).str.strptime(pl.Datetime, x, strict=False)
                for x in ["%Y-%m-%dT%H:%M:%S", "%Y-%m-%d %H:%M:%S"]
            )
            for col in str_cols
        ).alias("Datetime"),
        pl.struct(
            pl.coalesce(
                pl.col(col).str.strptime(pl.Date, x, strict=False) for x in ["%Y-%m-%d"]
            )
            for col in str_cols
        ).alias("Date"),
        *[
            pl.struct(pl.col(col).cast(x, strict=False) for col in str_cols).alias(
                str(x)
            )
            for x in [pl.Int64, pl.Float64]
        ],
    )

    good_cast = (
        pl.concat(
            [
                try_casts.select(
                    pl.lit(str(x)).alias("dtype"),
                    *(
                        pl.col(str(x)).struct.field(col).null_count()
                        - pl.col("original").struct.field(col).null_count()
                        for col in str_cols
                    ),
                )
                for x in [pl.Datetime, pl.Date, pl.Float64, pl.Int64]
            ]
        )
        .melt("dtype")
        .filter(pl.col("value") == 0)
        .sort("dtype")
        .unique("variable", keep="last", maintain_order=True)
        .drop("value")
    )

    map_cols = pl.concat(
        [
            good_cast,
            pl.Series("variable", df.columns)
            .to_frame()
            .with_columns(pl.lit("original").alias("dtype")),
        ],
        how="diagonal_relaxed",
    ).unique("variable", keep="first")
    return try_casts.select(
        pl.col(x).struct.field(y) for x, y in map_cols.iter_rows()
    ).select(df.columns)