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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
|
#![feature(iterator_try_collect)]
use anyhow::Result;
use clap::Parser;
use priority_queue::PriorityQueue;
use sha2::{Sha512_256, Digest};
use std::{cmp::Ordering, collections::HashMap, fs, io, path::PathBuf};
use embedders::*;
use pure_embedders::*;
use ai_embedders::*;
mod embedders;
mod pure_embedders;
mod ai_embedders;
#[derive(Debug, Clone, Copy, clap::ValueEnum)]
enum Embedder {
Brightness,
Hue,
Color,
Content,
}
#[derive(Debug, Parser)]
struct Args {
#[arg(short, long, default_value = "hue")]
embedder: Embedder,
images: Vec<PathBuf>,
}
#[derive(Debug)]
struct Config {
base_dirs: xdg::BaseDirectories,
}
fn get_config() -> Result<Config> {
let dirs = xdg::BaseDirectories::with_prefix("embeddings-sort")?;
Ok(Config { base_dirs: dirs })
}
fn get_mst<M>(embeds: &Vec<M>) -> HashMap<usize, Vec<usize>>
where M: MetricElem
{
// wrapper struct to
// - reverse the ordering
// - implement Ord, even though the type is backed by an f64
#[repr(transparent)]
#[derive(Debug, PartialEq)]
struct DownOrd (f64);
impl Eq for DownOrd {}
impl PartialOrd for DownOrd {
fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
Some(self.cmp(other))
}
}
impl Ord for DownOrd {
fn cmp(&self, other: &Self) -> Ordering {
self.0.partial_cmp(&other.0).unwrap().reverse()
}
}
let num_embeds = embeds.len();
let mut possible_edges =
PriorityQueue::with_capacity((num_embeds * num_embeds - num_embeds) / 2);
let mut mst = HashMap::with_capacity(num_embeds);
// here, we start at 0.
// we might get a better result in the end if we started with a vertex next
// to the lowest-cost edge, but we don't know which one that is (though we
// could compute that without changing our asymptotic complexity)
mst.insert(0, Vec::new());
for i in 1..num_embeds {
possible_edges.push((0, i), DownOrd(embeds[0].dist(&embeds[i])));
}
// prims algorithm or something like that
while mst.len() < num_embeds {
// find the edge with the least cost that connects us to a new vertex
let (new, old) = loop {
let ((a, b), _) = possible_edges.pop().unwrap();
if !mst.contains_key(&a) {
break (a, b);
}
else if !mst.contains_key(&b) {
break (b, a);
}
};
mst.insert(new, Vec::new());
// insert all the new edges we could take
mst.entry(old).and_modify(|v|v.push(new));
for i in 0..num_embeds {
// don't consider edges taking us to nodes we already visited
if mst.contains_key(&i) {
continue;
}
possible_edges.push((new, i), DownOrd(embeds[new].dist(&embeds[i])));
}
}
mst
}
fn tsp_from_mst(mst: HashMap<usize, Vec<usize>>) -> Vec<usize> {
fn dfs(cur: usize, t: &HashMap<usize, Vec<usize>>, into: &mut Vec<usize>) {
into.push(cur);
t.get(&cur).unwrap().iter().for_each(|c| dfs(*c, t, into));
}
let mut tsp_path = Vec::with_capacity(mst.len());
dfs(0, &mst, &mut tsp_path);
tsp_path
}
fn hash_file(p: &PathBuf) -> Result<[u8; 32]> {
let mut f = fs::File::open(p)?;
let mut hasher = Sha512_256::new();
io::copy(&mut f, &mut hasher)?;
Ok(hasher.finalize().into_iter().collect::<Vec<u8>>().try_into().unwrap())
}
fn process_embedder<E>(mut e: E, args: Args, cfg: &Config) -> Result<Vec<PathBuf>>
where E: BatchEmbedder
{
if args.images.is_empty() {
return Ok(Vec::new());
}
let db = sled::open(cfg.base_dirs.place_cache_file("embeddings.db")?)?;
let tree = typed_sled::Tree::<[u8; 32], E::Embedding>::open(&db, E::NAME);
let mut embeds: Vec<Option<_>> = args.images
.iter()
.map(|p| {
let h = hash_file(p)?;
let r: Result<Option<E::Embedding>> = tree.get(&h).map_err(|e| e.into());
r
})
.try_collect()?;
let missing_embeds_indices: Vec<_> = embeds
.iter()
.enumerate()
.filter_map(|(i, v)| match v {
None => Some(i),
Some(_) => None,
}).collect();
// TODO only run e.embeds if !missing_embeds_indices.is_empty(); this allows
// for optimizations in the ai embedde (move pip to ::embeds() instead of ::new())
let missing_embeds = e.embeds(&missing_embeds_indices
.iter()
.map(|i| args.images[*i].clone())
.collect::<Vec<_>>())?;
for (idx, emb) in missing_embeds_indices
.into_iter().zip(missing_embeds.into_iter())
{
tree.insert(&hash_file(&args.images[idx])?, &emb)?;
embeds[idx] = Some(emb);
}
let embeds: Vec<_> = embeds.into_iter().map(|e| e.unwrap()).collect();
let tsp_path = tsp_from_mst(get_mst(&embeds));
Ok(tsp_path.iter().map(|i| args.images[*i].clone()).collect())
}
fn main() -> Result<()> {
let cfg = get_config()?;
let args = Args::parse();
let tsp_path = match args.embedder {
Embedder::Brightness => process_embedder(BrightnessEmbedder, args, &cfg),
Embedder::Hue => process_embedder(HueEmbedder, args, &cfg),
Embedder::Color => process_embedder(ColorEmbedder, args, &cfg),
Embedder::Content => process_embedder(ContentEmbedder::new(&cfg), args, &cfg),
}?;
for p in tsp_path {
println!("{:?}", p);
}
Ok(())
}
|
