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

use crate::ops::prelude::*;
use ndarray::*;

#[derive(Debug, Clone, new, Default)]
pub struct ArgMaxMin {
    max: bool,
    axis: usize,
    keepdims: bool,
}

impl ArgMaxMin {
    fn eval_t<T: Datum + PartialOrd>(&self, input: SharedTensor) -> TractResult<SharedTensor> {
        use std::cmp::Ordering;
        let array = input.to_array_view::<T>()?;
        let f: fn(&(usize, &T), &(usize, &T)) -> Ordering = if self.max {
            |a, b| a.1.partial_cmp(&b.1).unwrap_or(a.0.cmp(&b.0))
        } else {
            |a, b| b.1.partial_cmp(&a.1).unwrap_or(a.0.cmp(&b.0))
        };
        let mut values = array.map_axis(Axis(self.axis), |row| {
            row.iter().enumerate().max_by(f).unwrap().0 as i64
        });
        if self.keepdims {
            values = values.insert_axis(Axis(self.axis));
        }
        Ok(Tensor::from(values).into())
    }
}

impl Op for ArgMaxMin {
    fn name(&self) -> Cow<str> {
        "ArgMaxMin".into()
    }
}

impl StatelessOp for ArgMaxMin {
    fn eval(&self, mut inputs: TVec<SharedTensor>) -> TractResult<TVec<SharedTensor>> {
        let input = args_1!(inputs);
        Ok(tvec!(dispatch_numbers!(Self::eval_t(input.datum_type())(
            self, input
        ))?))
    }
}

impl InferenceRulesOp for ArgMaxMin {
    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        s: &mut Solver<'r>,
        inputs: &'p SharedTensorsProxy,
        outputs: &'p SharedTensorsProxy,
    ) -> InferenceResult {
        s.equals(&inputs.len, 1)?;
        s.equals(&outputs.len, 1)?;
        s.equals(&outputs[0].datum_type, DatumType::I64)?;
        if self.keepdims {
            s.equals(&outputs[0].rank, &inputs[0].rank)?;
            for i in 0..self.axis {
                s.equals(&outputs[0].shape[i], &inputs[0].shape[i])?;
            }
            s.equals(&outputs[0].shape[self.axis], 1.to_dim())?;
            s.given(&inputs[0].rank, move |s, rank| {
                for i in (self.axis + 1)..(rank as usize) {
                    s.equals(&outputs[0].shape[i], &inputs[0].shape[i])?;
                }
                Ok(())
            })?;
        } else {
            s.equals(&outputs[0].rank, inputs[0].rank.bex() - 1)?;
            for i in 0..self.axis {
                s.equals(&outputs[0].shape[i], &inputs[0].shape[i])?;
            }
            s.given(&inputs[0].rank, move |s, rank| {
                for i in (self.axis + 1)..(rank as usize - 1) {
                    s.equals(&outputs[0].shape[i], &inputs[0].shape[i + 1])?;
                }
                Ok(())
            })?;
        };
        Ok(())
    }
}