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;
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);
let theoretical_grads = crate::runtime::eval(gradients, feeds.clone());
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 i in 0..v_arr.len() as isize {
let evacuated;
unsafe {
evacuated = *head_ptr.offset(i);
*head_ptr.offset(i) = evacuated + eps;
}
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())
};
unsafe {
*head_ptr.offset(i) = evacuated - eps;
}
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())
};
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) };
let diff = (g_num - g_th).abs();
if diff > tol {
panic!(
"Gradient checking failed with too large error: numerical={}, theoretical={}",
g_num, g_th
);
}
}
}
}
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;
} else {
visited.insert(t);
}
f(&t);
for child in t.inputs.iter() {
visit_once_internal(child, f, visited)
}
}
impl<'a, T: Float> Ord for &'a Tensor<T> {
#[inline]
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]
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
}};
}