|
| 1 | + |
| 2 | +# 11. ARRAY, captain! |
| 3 | + |
| 4 | +We've talked about lists, structs, but what about arrays? |
| 5 | + |
| 6 | +In this section we're gonna cover how to deal with fixed sized arrays, e.g., x and y coordinates of 2d points *in the same column*: |
| 7 | + |
| 8 | +```python |
| 9 | +points = pl.Series( |
| 10 | + "points", |
| 11 | + [ |
| 12 | + [6.63, 8.35], |
| 13 | + [7.19, 4.85], |
| 14 | + [2.1, 4.21], |
| 15 | + [3.4, 6.13], |
| 16 | + ], |
| 17 | + dtype=pl.Array(pl.Float64, 2), |
| 18 | +) |
| 19 | +df = pl.DataFrame(points) |
| 20 | + |
| 21 | +print(df) |
| 22 | +``` |
| 23 | + |
| 24 | +``` |
| 25 | +shape: (4, 1) |
| 26 | +┌───────────────┐ |
| 27 | +│ points │ |
| 28 | +│ --- │ |
| 29 | +│ array[f64, 2] │ |
| 30 | +╞═══════════════╡ |
| 31 | +│ [6.63, 8.35] │ |
| 32 | +│ [7.19, 4.85] │ |
| 33 | +│ [2.1, 4.21] │ |
| 34 | +│ [3.4, 6.13] │ |
| 35 | +└───────────────┘ |
| 36 | +``` |
| 37 | + |
| 38 | +Let's get to work - what if we wanted to make a plugin that takes a Series like `points` above, and, likewise, returned a Series of arrays? |
| 39 | +Turns out we _can_ do it! But it's a little bit tricky. |
| 40 | + |
| 41 | +__First of all__, we need to include `features = ["dtype-array"]` in both `pyo3-polars` and `polars-core` in our `Cargo.toml`. |
| 42 | + |
| 43 | +Now let's create a plugin that calculates the midpoint between a reference point and each point in a Series like the one above. |
| 44 | +This should illustrate both how to unpack an array inside our Rust code and also return a Series of the same type. |
| 45 | + |
| 46 | +We'll start by registering our plugin: |
| 47 | + |
| 48 | +```python |
| 49 | +def midpoint_2d(expr: IntoExpr, ref_point: tuple[float, float]) -> pl.Expr: |
| 50 | + return register_plugin_function( |
| 51 | + args=[expr], |
| 52 | + plugin_path=Path(__file__).parent, |
| 53 | + function_name="midpoint_2d", |
| 54 | + is_elementwise=True, |
| 55 | + kwargs={"ref_point": ref_point}, |
| 56 | + ) |
| 57 | +``` |
| 58 | + |
| 59 | +As you can see, we included an additional kwarg: `ref_point`, which we annotated with the type `tuple: [float, float]`. |
| 60 | +In our Rust code, we won't receive it as a tuple, though, it'll also be an array. |
| 61 | +This isn't crucial for this example, so just accept it for now. |
| 62 | +As you saw in the __arguments__ chapter, we take kwargs by defining a struct for them: |
| 63 | + |
| 64 | +```rust |
| 65 | +#[derive(Deserialize)] |
| 66 | +struct MidPoint2DKwargs { |
| 67 | + ref_point: [f64; 2], |
| 68 | +} |
| 69 | +``` |
| 70 | + |
| 71 | +And we can finally move to the actual plugin code: |
| 72 | + |
| 73 | +```rust |
| 74 | +// We need this to ensure the output is of dtype array. |
| 75 | +// Unfortunately, polars plugins do not support something similar to: |
| 76 | +// #[polars_expr(output_type=Array)] |
| 77 | +pub fn point_2d_output(_: &[Field]) -> PolarsResult<Field> { |
| 78 | + Ok(Field::new( |
| 79 | + "point_2d", |
| 80 | + DataType::Array(Box::new(DataType::Float64), 2), |
| 81 | + )) |
| 82 | +} |
| 83 | + |
| 84 | +#[polars_expr(output_type_func=point_2d_output)] |
| 85 | +fn midpoint_2d(inputs: &[Series], kwargs: MidPoint2DKwargs) -> PolarsResult<Series> { |
| 86 | + let ca: &ArrayChunked = inputs[0].array()?; |
| 87 | + let ref_point = kwargs.ref_point; |
| 88 | + |
| 89 | + let out: ArrayChunked = unsafe { |
| 90 | + ca.try_apply_amortized_same_type(|row| { |
| 91 | + let s = row.as_ref(); |
| 92 | + let ca = s.f64()?; |
| 93 | + let out_inner: Float64Chunked = ca |
| 94 | + .iter() |
| 95 | + .enumerate() |
| 96 | + .map(|(idx, opt_val)| { |
| 97 | + opt_val.map(|val| { |
| 98 | + (val + ref_point[idx]) / 2.0f64 |
| 99 | + }) |
| 100 | + }).collect_trusted(); |
| 101 | + Ok(out_inner.into_series()) |
| 102 | + })}?; |
| 103 | + |
| 104 | + Ok(out.into_series()) |
| 105 | +} |
| 106 | +``` |
| 107 | + |
| 108 | +Uh-oh, unsafe, we're doomed! |
| 109 | + |
| 110 | +Hold on a moment - it's true that we need unsafe here, but let's not freak out. |
| 111 | +If we read the docs of `try_apply_amortized_same_type`, we see the following: |
| 112 | + |
| 113 | +> ```rust |
| 114 | +> /// Try apply a closure `F` to each array. |
| 115 | +> /// |
| 116 | +> /// # Safety |
| 117 | +> /// Return series of `F` must has the same dtype and number of elements as input if it is Ok. |
| 118 | +> pub unsafe fn try_apply_amortized_same_type<F>(&self, mut f: F) -> PolarsResult<Self> |
| 119 | +> where |
| 120 | +> F: FnMut(AmortSeries) -> PolarsResult<Series>, |
| 121 | +> ``` |
| 122 | +
|
| 123 | +
|
| 124 | +In this example, we can uphold that contract - we know we're returning a Series with the same number of elements and same dtype as the input! |
| 125 | +
|
| 126 | +Still, the code looks a bit scary, doesn't it? So let's break it down: |
| 127 | +
|
| 128 | +```rust |
| 129 | +let out: ArrayChunked = unsafe { |
| 130 | +
|
| 131 | + // This is similar to apply_values, but it's amortized and made specifically |
| 132 | + // for arrays. |
| 133 | + ca.try_apply_amortized_same_type(|row| { |
| 134 | + let s = row.as_ref(); |
| 135 | + // `s` is a Series which contains two elements. |
| 136 | + // We unpack it similarly to the way we've been unpacking Series in the |
| 137 | + // previous chapters: |
| 138 | + // |
| 139 | + // Previously we've been doing this to unpack a column we had behind a |
| 140 | + // Series - this time, inside this closure, the Series contains the two |
| 141 | + // elements composing the "row" (x and y): |
| 142 | + let ca = s.f64()?; |
| 143 | +
|
| 144 | + // There are many ways to extract the x and y coordinates from ca. |
| 145 | + // Here, we remain idiomatic and consistent with what we've been doing |
| 146 | + // in the past - iterate, enumerate and map: |
| 147 | + let out_inner: Float64Chunked = ca |
| 148 | + .iter() |
| 149 | + .enumerate() |
| 150 | + .map(|(idx, opt_val)| { |
| 151 | +
|
| 152 | + // We only use map here because opt_val is an Option |
| 153 | + opt_val.map(|val| { |
| 154 | +
|
| 155 | + // Here's where the simple logic of calculating a |
| 156 | + // midpoint happens. We take the coordinate (`val`) at |
| 157 | + // index `idx`, add it to the `idx-th` entry of our |
| 158 | + // reference point (which is a coordinate of our point), |
| 159 | + // then divide it by two, since we're dealing with 2d |
| 160 | + // points only. |
| 161 | + (val + ref_point[idx]) / 2.0f64 |
| 162 | + }) |
| 163 | + // Our map already returns Some or None, so we don't have to |
| 164 | + // worry about wrapping the result in, e.g., Some() |
| 165 | + }).collect_trusted(); |
| 166 | +
|
| 167 | + // At last, we convert out_inner (which is a Float64Chunked) back to a |
| 168 | + // Series |
| 169 | + Ok(out_inner.into_series()) |
| 170 | + })}?; |
| 171 | +
|
| 172 | +// And finally, we convert our ArrayChunked into a Series, ready to ship to |
| 173 | +// Python-land: |
| 174 | +Ok(out.into_series()) |
| 175 | +``` |
| 176 | +
|
| 177 | +That's it. What does the result look like? |
| 178 | +In `run.py`, we have: |
| 179 | + |
| 180 | +```python |
| 181 | +import polars as pl |
| 182 | +from minimal_plugin import midpoint_2d |
| 183 | + |
| 184 | +points = pl.Series( |
| 185 | + "points", |
| 186 | + [ |
| 187 | + [6.63, 8.35], |
| 188 | + [7.19, 4.85], |
| 189 | + [2.1, 4.21], |
| 190 | + [3.4, 6.13], |
| 191 | + [2.48, 9.26], |
| 192 | + [9.41, 7.26], |
| 193 | + [7.45, 8.85], |
| 194 | + [6.58, 5.22], |
| 195 | + [6.05, 5.77], |
| 196 | + [8.57, 4.16], |
| 197 | + [3.22, 4.98], |
| 198 | + [6.62, 6.62], |
| 199 | + [9.36, 7.44], |
| 200 | + [8.34, 3.43], |
| 201 | + [4.47, 7.61], |
| 202 | + [4.34, 5.05], |
| 203 | + [5.0, 5.05], |
| 204 | + [5.0, 5.0], |
| 205 | + [2.07, 7.8], |
| 206 | + [9.45, 9.6], |
| 207 | + [3.1, 3.26], |
| 208 | + [4.37, 5.72], |
| 209 | + ], |
| 210 | + dtype=pl.Array(pl.Float64, 2), |
| 211 | +) |
| 212 | +df = pl.DataFrame(points) |
| 213 | + |
| 214 | +# Now we call our plugin: |
| 215 | +result = df.with_columns(midpoints=midpoint_2d("points", ref_point=(5.0, 5.0))) |
| 216 | +print(result) |
| 217 | +``` |
| 218 | + |
| 219 | +Let's compile and run it: |
| 220 | +```shell |
| 221 | +maturin develop |
| 222 | + |
| 223 | +python run.py |
| 224 | +``` |
| 225 | + |
| 226 | +🥁: |
| 227 | +``` |
| 228 | +shape: (22, 2) |
| 229 | +┌───────────────┬────────────────┐ |
| 230 | +│ points ┆ midpoints │ |
| 231 | +│ --- ┆ --- │ |
| 232 | +│ array[f64, 2] ┆ array[f64, 2] │ |
| 233 | +╞═══════════════╪════════════════╡ |
| 234 | +│ [6.63, 8.35] ┆ [5.815, 6.675] │ |
| 235 | +│ [7.19, 4.85] ┆ [6.095, 4.925] │ |
| 236 | +│ [2.1, 4.21] ┆ [3.55, 4.605] │ |
| 237 | +│ [3.4, 6.13] ┆ [4.2, 5.565] │ |
| 238 | +│ [2.48, 9.26] ┆ [3.74, 7.13] │ |
| 239 | +│ … ┆ … │ |
| 240 | +│ [5.0, 5.0] ┆ [5.0, 5.0] │ |
| 241 | +│ [2.07, 7.8] ┆ [3.535, 6.4] │ |
| 242 | +│ [9.45, 9.6] ┆ [7.225, 7.3] │ |
| 243 | +│ [3.1, 3.26] ┆ [4.05, 4.13] │ |
| 244 | +│ [4.37, 5.72] ┆ [4.685, 5.36] │ |
| 245 | +└───────────────┴────────────────┘ |
| 246 | +``` |
| 247 | + |
| 248 | + |
| 249 | +!!!note |
| 250 | + Notice how the dtype remains the same. |
| 251 | + As an exercise, try to achieve the same in pure-Python (without Rust plugins) |
| 252 | + without explicitly casting the type of the Series. |
| 253 | + |
| 254 | +Hurray, we did it! |
| 255 | +And why exactly go through all this trouble instead of just doing the same thing in pure Python? |
| 256 | +For performance of course! |
| 257 | + |
| 258 | +_Spoilers ahead if you haven't tried the exercise from the note above_ |
| 259 | + |
| 260 | +With the following implementation in Python, we can take some measurements: |
| 261 | + |
| 262 | +```python |
| 263 | +ref_point = (5.0, 5.0) |
| 264 | + |
| 265 | +def using_plugin(df=df, ref_point=ref_point): |
| 266 | + result = df.with_columns(midpoints=midpoint_2d("points", ref_point=ref_point)) |
| 267 | + return result |
| 268 | + |
| 269 | +def midpoint(points:pl.Series) -> pl.Series: |
| 270 | + result=[] |
| 271 | + for point in points: |
| 272 | + result.append([(point[0]+ref_point[0])/2, (point[1]+ref_point[1])/2]) |
| 273 | + return pl.Series(result, dtype=pl.Array(pl.Float64, 2)) |
| 274 | + |
| 275 | +def using_python(df=df, ref_point=ref_point): |
| 276 | + result = ( |
| 277 | + df.with_columns( |
| 278 | + midpoints=pl.col('points').map_batches(midpoint, return_dtype=pl.Array(pl.Float64, 2)) |
| 279 | + ) |
| 280 | + ) |
| 281 | + return result |
| 282 | +``` |
| 283 | + |
| 284 | +For the sake of brevity, some extra methods to generate and parse an input file were left out of the code above, as well as the `timeit` bits. |
| 285 | +By measuring both versions with 1.000.000 points a few times and taking the average, we got the following result: |
| 286 | + |
| 287 | +``` |
| 288 | +Using plugin: |
| 289 | +min: 0.5307095803339811 |
| 290 | +max: 0.5741689523274545 |
| 291 | +mean +/- stderr: 0.5524565599986263 +/- 0.0064489015434971925 |
| 292 | +
|
| 293 | +Using python: |
| 294 | +min: 6.682447870339577 |
| 295 | +max: 6.99253460233255 |
| 296 | +mean +/- stderr: 6.808615755191394 +/- 0.03757884107880601 |
| 297 | +``` |
| 298 | + |
| 299 | +A speedup of __12x__, that's a __big win__! |
| 300 | + |
| 301 | +!!!note |
| 302 | + When benchmarking Rust code, remember to use `maturin develop --release`, otherwise the timings will be much slower! |
0 commit comments