csv - 使用索引选择是 Polars : How to parse and transform (select/filter? 中的反模式)似乎需要这样的 CSV？

Question

我想阅读来自 Rigol MSO5000 示波器的带有 Pola-rs 的以下(非常破损，IMO)CSV:

D7-D0,D15-D8,t0 = -25.01s, tInc = 2e-06,
+2.470000E+02,2.000000E+00,,
+1.590000E+02,1.600000E+01,,
+2.400000E+02,2.000000E+00,,
+2.470000E+02,+1.300000E+02,,
+1.590000E+02,1.800000E+01,,
+2.470000E+02,+1.300000E+02,,
9.500000E+01,1.800000E+01,,
9.500000E+01,1.800000E+01,,
+2.400000E+02,0.000000E+00,,
(...)

在发现 Selecting with indexing is discouraged in Pola-rs 之前，这是我当前的 Jupyter Notebook 迭代/尝试:

import polars as pl

df = pl.read_csv("normal0.csv")

# Grab initial condition and increments
t0 = df.columns[2]; assert "t0" in t0; t0 = float(t0.split('=')[1].replace('s', '').strip())
tinc = df.columns[3]; assert "tInc" in tinc; tinc = float(tinc.split('=')[1].strip())

# TODO: Generate Series() from t0+tinc and include it in the final DataFrame

# Reshape and cleanup
probes = df.with_column(df["D7-D0"].cast(pl.Float32).alias("D0-D7 floats")) \
             .with_column(df["D15-D8"].cast(pl.Float32).alias("D15-D8 floats")) \
             .drop(df.columns[2]) \
             .drop(df.columns[3])
(...)

def split_probes(probes: pl.Series):
    ''' Splits bundles of probe cables such as D0-D7 or D15-D8 into individual dataframe columns
    '''
    out = pl.DataFrame(columns=["D"+str(line) for line in range(0,16)])
#    for row in range(probes.height):
#        for probe in range(0, 7):
#            out["D"+str(probe)].with_columns(probes["D0-D7 floats"][row % (probe + 1)])
#         for probe in reversed(range(9, 16)): # TODO: Fix future captures or parametrise this
#             outprobes["D15-D8 floats"][row % probe]

这是我在 I was told on Polars Discord that this problem might not be optimally solvable with dataframe libraries 时的低级 CSV 解析 Rust 伪代码方法:

use csv;
use std::error::Error;
use std::io;
use std::process;
use serde::Deserialize;

#[derive(Debug, Deserialize)]
struct OrigOscilloscope {
    #[serde(rename = "D7-D0")]
    d7_d0: String, // TODO: Unfortunately those fields are "user-flippable" in order from the scope, i.e: d0_d7 vs d7_d0
    #[serde(rename = "D15-D8")]
    d15_d8: String,
    // Do not even register those on Serde as they are empty rows anyway
    // t0: Option<String>,
    // t_inc: Option<String>
}

#[derive(Debug, Deserialize)]
struct LAProbesOscilloscopeState {
    //Vec<d0...d15>: Vec<Vec<16*f32>>, // TODO: More appropriate struct representation for the target dataframe
    d0: f32,
    d1: f32,
    d2: f32,
    d3: f32,
    d4: f32,
    d5: f32,
    d6: f32,
    d7: f32,
    d8: f32,
    d9: f32,
    d10: f32,
    d11: f32,
    d12: f32,
    d13: f32,
    d14: f32,
    d15: f32,
    timestamp: f32
}

fn run() -> Result<(), Box<dyn Error>> {
    let mut rdr = csv::ReaderBuilder::new()
        .has_headers(false)
        .flexible(true) // ignore broken header
        .from_reader(io::stdin());

    // Get timeseries information from header... :facepalm: rigol
    // Initial timestamp...
    let header = rdr.headers()?.clone();
    let t0_header: Vec<&str> = header[2].split('=').collect();
    let t0 = t0_header[1].trim_start().replace('s', "").parse::<f32>()?;
    // ...and increments
    let tinc_header: Vec<&str> = header[3].split('=').collect();
    let tinc = tinc_header[1].trim_start().parse::<f32>()?;
    println!("Initial timestamp {t0} with increments of {tinc} seconds");

    // Now do the splitting of wires from its Dx-Dy "bundles"
    let mut timestamp = t0;
    for result in rdr.deserialize().skip(1) {
        let row: OrigOscilloscope = result?;

        // if rdr.position().line().rem_euclid(8) {
        // // Read D0-D15 field(s), expanding them into the current row, matching its column
        // }
        // println!("{:#?}", row.d7_d0.parse::<f32>()?);

        // update timestamp for this row
        timestamp = timestamp + tinc;
    }
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

我希望很明显，所需的目标数据框看起来像这样:

d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10, d11, d12, d13, d14, d15, timestamp
95, 95, 247, 159, 247, 240, 159, 247 (...)                            -25.01+000000.2
(...)

生成的代码应该(理想情况下？)有效地使用 Polars。此外，如上所述，避免“使用索引选择”，因为它可能在未来在 Polars 中完全弃用。

Answer 1

也许您正在寻找 unstack(它的性能令人难以置信)。

让我们从这些数据开始(它只是复制您提供的值)。我还更改了列名，只是为了让事情更容易检查:

import polars as pl
from io import StringIO

normal_csv = """D7-D0,D15-D8,t0 = -25.01s, tInc = 2e-06,
+2.470000E+02,2.000000E+00,,
+1.590000E+02,1.600000E+01,,
+2.400000E+02,2.000000E+00,,
+2.470000E+02,+1.300000E+02,,
+1.590000E+02,1.800000E+01,,
+2.470000E+02,+1.300000E+02,,
9.500000E+01,1.800000E+01,,
9.500000E+01,1.800000E+01,,
+2.470000E+02,2.000000E+00,,
+1.590000E+02,1.600000E+01,,
+2.400000E+02,2.000000E+00,,
+2.470000E+02,+1.300000E+02,,
+1.590000E+02,1.800000E+01,,
+2.470000E+02,+1.300000E+02,,
9.500000E+01,1.800000E+01,,
9.500000E+01,1.800000E+01,,"""


df = pl.read_csv(file=StringIO(normal_csv),
                 new_columns=('D', 'E', 't0', 'tInc'))
df

shape: (16, 4)
┌───────┬───────┬──────┬──────┐
│ D     ┆ E     ┆ t0   ┆ tInc │
│ ---   ┆ ---   ┆ ---  ┆ ---  │
│ f64   ┆ f64   ┆ str  ┆ str  │
╞═══════╪═══════╪══════╪══════╡
│ 247.0 ┆ 2.0   ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 159.0 ┆ 16.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 240.0 ┆ 2.0   ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 247.0 ┆ 130.0 ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 159.0 ┆ 18.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 247.0 ┆ 130.0 ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 95.0  ┆ 18.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 95.0  ┆ 18.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 247.0 ┆ 2.0   ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 159.0 ┆ 16.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 240.0 ┆ 2.0   ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 247.0 ┆ 130.0 ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 159.0 ┆ 18.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 247.0 ┆ 130.0 ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 95.0  ┆ 18.0  ┆ null ┆ null │
├╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┤
│ 95.0  ┆ 18.0  ┆ null ┆ null │
└───────┴───────┴──────┴──────┘

由于我不熟悉您的数据源，我假设您的输入数据具有正则化模式——特别是始终提供 D7-D0(即，csv 文件中没有跳过的行)。

如果是这样，这里有一些高性能代码应该可以启动......

(
    df
    .reverse()
    .unstack(step=8,
             how='horizontal',
             columns=('D', 'E'),
             )
)

shape: (2, 16)
┌──────┬──────┬───────┬───────┬───────┬───────┬───────┬───────┬──────┬──────┬───────┬──────┬───────┬─────┬──────┬─────┐
│ D_0  ┆ D_1  ┆ D_2   ┆ D_3   ┆ D_4   ┆ D_5   ┆ D_6   ┆ D_7   ┆ E_0  ┆ E_1  ┆ E_2   ┆ E_3  ┆ E_4   ┆ E_5 ┆ E_6  ┆ E_7 │
│ ---  ┆ ---  ┆ ---   ┆ ---   ┆ ---   ┆ ---   ┆ ---   ┆ ---   ┆ ---  ┆ ---  ┆ ---   ┆ ---  ┆ ---   ┆ --- ┆ ---  ┆ --- │
│ f64  ┆ f64  ┆ f64   ┆ f64   ┆ f64   ┆ f64   ┆ f64   ┆ f64   ┆ f64  ┆ f64  ┆ f64   ┆ f64  ┆ f64   ┆ f64 ┆ f64  ┆ f64 │
╞══════╪══════╪═══════╪═══════╪═══════╪═══════╪═══════╪═══════╪══════╪══════╪═══════╪══════╪═══════╪═════╪══════╪═════╡
│ 95.0 ┆ 95.0 ┆ 247.0 ┆ 159.0 ┆ 247.0 ┆ 240.0 ┆ 159.0 ┆ 247.0 ┆ 18.0 ┆ 18.0 ┆ 130.0 ┆ 18.0 ┆ 130.0 ┆ 2.0 ┆ 16.0 ┆ 2.0 │
├╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌╌╌┼╌╌╌╌╌┼╌╌╌╌╌╌┼╌╌╌╌╌┤
│ 95.0 ┆ 95.0 ┆ 247.0 ┆ 159.0 ┆ 247.0 ┆ 240.0 ┆ 159.0 ┆ 247.0 ┆ 18.0 ┆ 18.0 ┆ 130.0 ┆ 18.0 ┆ 130.0 ┆ 2.0 ┆ 16.0 ┆ 2.0 │
└──────┴──────┴───────┴───────┴───────┴───────┴───────┴───────┴──────┴──────┴───────┴──────┴───────┴─────┴──────┴─────┘

这没有您想要的列名称(您可以更改)。它也不包含时间信息。但它可能会帮助您朝着正确的方向开始。