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
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159

use crate::ndarray_ext;
use crate::tensor::Tensor;
use crate::{Feed, Float};
use std::cmp::Ordering;
use std::collections::btree_set::BTreeSet;

/// Checks the validity of `gradients` with finite difference trick.
/// For this test only, `variables` must be *shared* variables.
pub fn check_theoretical_grads<'k, 'v, A, T>(
    objective: &'k Tensor<T>,
    gradients: &'k [A],
    variables: &[&Tensor<T>],
    feeds: &'v [Feed<'k, 'v, T>],
    eps: T,
    tol: T,
) where
    A: AsRef<Tensor<T>>,
    T: Float,
{
    let objective = crate::ops::reduce_sum_to_scalar(objective);
    // backprop
    let theoretical_grads = crate::runtime::eval(gradients, feeds.clone());

    // for each variable nodes
    for (var_node, th_grad) in variables.iter().zip(theoretical_grads) {
        let th_copied = if th_grad.as_ref().unwrap().is_standard_layout() {
            None
        } else {
            Some(ndarray_ext::deep_copy(&th_grad.as_ref().unwrap().view()))
        };
        let th_ptr = if let Some(ref inner) = th_copied {
            inner.as_ptr()
        } else {
            th_grad.as_ref().unwrap().as_ptr()
        };

        let v_arr = unsafe {
            var_node
                .get_persistent_array_mut()
                .expect("This is not a variable")
        };
        let head_ptr: *mut T = v_arr.as_mut_ptr();

        // for each values
        for i in 0..v_arr.len() as isize {
            let evacuated;

            // perturbation (+)
            unsafe {
                evacuated = *head_ptr.offset(i);
                *head_ptr.offset(i) = evacuated + eps;
            }

            // eval
            let obj_pos_orig = objective.eval(feeds).unwrap();
            let obj_pos = if obj_pos_orig.is_standard_layout() {
                obj_pos_orig
            } else {
                ndarray_ext::deep_copy(&obj_pos_orig.view())
            };

            // perturbation (-)
            unsafe {
                *head_ptr.offset(i) = evacuated - eps;
            }

            // eval
            let obj_neg_orig = objective.eval(feeds).unwrap();
            let obj_neg = if obj_neg_orig.is_standard_layout() {
                obj_neg_orig
            } else {
                ndarray_ext::deep_copy(&obj_neg_orig.view())
            };

            // restore
            unsafe {
                *head_ptr.offset(i) = evacuated;
            }

            let two = T::one() + T::one();
            let g_num = (obj_pos - obj_neg).scalar_sum() / (two * eps);
            let g_th = unsafe { *th_ptr.offset(i) };

            // compare
            let diff = (g_num - g_th).abs();
            if diff > tol {
                panic!(
                    "Gradient checking failed with too large error: numerical={}, theoretical={}",
                    g_num, g_th
                );
            }
        }
    }
}

/// Traverse a graph from endpoint "t".
pub fn visit_once<F, T: Float>(t: &Tensor<T>, f: &mut F)
where
    F: FnMut(&Tensor<T>) -> (),
{
    visit_once_internal(t, f, &mut BTreeSet::new())
}

fn visit_once_internal<'a, F, T: Float>(
    t: &'a Tensor<T>,
    f: &mut F,
    visited: &mut BTreeSet<&'a Tensor<T>>,
) where
    F: FnMut(&'a Tensor<T>) -> (),
{
    if visited.contains(&t) {
        return; // exit early
    } else {
        visited.insert(t); // first visit
    }

    f(&t);

    for child in t.inputs.iter() {
        visit_once_internal(child, f, visited)
    }
}

impl<'a, T: Float> Ord for &'a Tensor<T> {
    #[inline]
    /// Compares the addresses of the two tensors.
    /// This can be used for ordering-based data structures (e.g. BinaryTree).
    fn cmp(&self, other: &&'a Tensor<T>) -> Ordering {
        let a = (*self) as *const Tensor<T>;
        let b = (*other) as *const Tensor<T>;
        a.cmp(&b)
    }
}

impl<'a, T: Float> PartialOrd for &'a Tensor<T> {
    #[inline]
    /// Compares the addresses of the two tensors.
    /// This can be used for ordering-based data structures (e.g. BinaryTree).
    fn partial_cmp(&self, other: &&'a Tensor<T>) -> Option<Ordering> {
        Some(self.cmp(&other))
    }
}

#[doc(hidden)]
#[macro_export]
macro_rules! eval_with_time {
    ($x:expr) => {{
        use std::time::{Duration, Instant};
        let start = Instant::now();
        let result = $x;
        let end = start.elapsed();
        println!(
            "{}.{:03} sec",
            end.as_secs(),
            end.subsec_nanos() / 1_000_000
        );
        result
    }};
}