v2rmp 0.5.0

rmpca — Route Optimization TUI
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98

use anyhow::{Context, Result};
use ort::session::Session;
use ort::value::Value;
use serde::{Deserialize, Serialize};
use std::path::Path;

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct NeuralRouteRequest {
    pub locations: Vec<[f64; 2]>,
    pub demands: Vec<f64>,
    pub capacity: f64,
    pub model_path: String,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct NeuralRouteResponse {
    pub visit_sequence: Vec<usize>,
    pub total_cost: f64,
    pub solve_time_ms: u64,
}

pub struct NeuralInferenceEngine {
    session: Session,
}

impl NeuralInferenceEngine {
    /// Initialize the engine by loading an ONNX model.
    pub fn new<P: AsRef<Path>>(model_path: P) -> Result<Self> {
        let session = Session::builder()?
            .commit_from_file(model_path)
            .context("Failed to load ONNX model")?;
        
        Ok(Self { session })
    }

    /// Solve a VRP instance using the loaded neural model.
    pub fn solve(&mut self, req: &NeuralRouteRequest) -> Result<NeuralRouteResponse> {
        let start_time = std::time::Instant::now();

        let n = req.locations.len();
        
        // 1. Prepare raw vectors for inputs
        // locs: [1, n, 2]
        let mut locs_data = Vec::with_capacity(n * 2);
        for loc in &req.locations {
            locs_data.push(loc[0] as f32);
            locs_data.push(loc[1] as f32);
        }

        // demands: [1, n, 1]
        let mut demands_data = Vec::with_capacity(n);
        for &demand in &req.demands {
            demands_data.push(demand as f32);
        }

        // capacity: [1, 1]
        let capacity_data = vec![req.capacity as f32];

        // 2. Create Values using (shape, vec) which is more version-resilient
        let locs_value = Value::from_array(([1, n, 2], locs_data))?;
        let demands_value = Value::from_array(([1, n, 1], demands_data))?;
        let capacity_value = Value::from_array(([1, 1], capacity_data))?;

        // 3. Run Inference
        let inputs = ort::inputs![
            "locs" => locs_value,
            "demand" => demands_value,
            "capacity" => capacity_value,
        ];

        let outputs = self.session.run(inputs)?;

        // 4. Post-process: Extract the route from the output tensor
        let output_tensor_value = outputs
            .get("actions")
            .context("Model output 'actions' not found")?;
        
        let (_shape, data) = output_tensor_value.try_extract_tensor::<i64>()?;

        let visit_sequence: Vec<usize> = data
            .iter()
            .map(|&id| id as usize)
            .filter(|&id| id != 0) // Usually skip depot (0) in actions
            .collect();

        Ok(NeuralRouteResponse {
            visit_sequence,
            total_cost: 0.0,
            solve_time_ms: start_time.elapsed().as_millis() as u64,
        })
    }
}

/// Convenience function for one-off neural routing
pub fn solve_neural(req: &NeuralRouteRequest) -> Result<NeuralRouteResponse> {
    let mut engine = NeuralInferenceEngine::new(&req.model_path)?;
    engine.solve(req)
}