File size: 12,296 Bytes
d5bfab8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 |
//! Pseudo code for a NxN convolution with weights.
use super::arc_work_model::{Task, PairType};
use super::{Image, ImageCompare, ImagePadding, ImageSize, ImageMaskCount};
use anyhow::Context;
use rand::SeedableRng;
use rand::rngs::StdRng;
use rand::distributions::{Distribution, Uniform};
#[allow(unused_imports)]
use super::{HtmlLog, ImageToHTML};
#[allow(dead_code)]
#[derive(Clone, Copy, Debug)]
enum InputOutputType {
Input,
Output,
}
#[allow(dead_code)]
#[derive(Clone, Debug)]
struct Sample {
convolution3x3: Image,
position_x: u8,
position_y: u8,
image_width: u8,
image_height: u8,
input_output_type: InputOutputType,
}
#[allow(dead_code)]
pub struct ExperimentWithConvolution {
tasks: Vec<Task>,
global_weights: Vec::<f32>,
local_weights: Vec::<f32>,
}
impl ExperimentWithConvolution {
#[allow(dead_code)]
pub fn new(tasks: Vec<Task>) -> Self {
Self {
tasks,
global_weights: vec!(),
local_weights: vec!(),
}
}
#[allow(dead_code)]
pub fn run(&mut self) -> anyhow::Result<()> {
println!("will process {} tasks", self.tasks.len());
let task0: Task;
{
let first_task: &Task = self.tasks.first().context("one or more")?;
task0 = first_task.clone();
}
println!("task: {}", task0.id);
// Random weights
let random_seed: u64 = 1;
let mut rng: StdRng = StdRng::seed_from_u64(random_seed);
let step = Uniform::<u16>::new(0, 1001);
{
let mut weights = Vec::<f32>::new();
for _ in 0..5 {
let random_value: u16 = step.sample(&mut rng);
let weight_between0and1: f32 = (random_value as f32) / 1000.0;
let weight: f32 = weight_between0and1 + 1.0;
weights.push(weight);
}
self.global_weights = weights;
}
{
let mut weights = Vec::<f32>::new();
for _ in 0..19 {
let random_value: u16 = step.sample(&mut rng);
let weight_between0and1: f32 = (random_value as f32) / 1000.0;
let weight: f32 = weight_between0and1 + 1.0;
weights.push(weight);
}
self.local_weights = weights;
}
println!("global_weights: {}", self.global_weights.len());
println!("local_weights: {}", self.local_weights.len());
// Add some junk to the initial weights
for _ in 0..100 {
self.mutate_global_weights(&mut rng);
self.mutate_local_weights(&mut rng);
}
// Extract samples
// the global weights stays locked while training for a single task
// local weights that gets updated while training with a single task
for pair in &task0.pairs {
if pair.pair_type == PairType::Test {
continue;
}
let samples_input: Vec<Sample> = Self::extract_samples(&pair.input.image, InputOutputType::Input)?;
let samples_output: Vec<Sample> = Self::extract_samples(&pair.output.image, InputOutputType::Output)?;
println!("pair: {} samples_input: {} samples_output: {}", pair.id, samples_input.len(), samples_output.len());
// mutate the local weights
// train with the samples
for sample in &samples_input {
self.mutate_local_weights_with_sample(sample, &mut rng);
}
for sample in &samples_output {
self.mutate_local_weights_with_sample(sample, &mut rng);
}
}
// query the model
let pair_count: usize = task0.pairs.len();
for (pair_index, pair) in task0.pairs.iter().enumerate() {
let pair_id: f32 = ((pair_index as f32) + 1.0) / ((pair_count as f32) + 1.0);
// println!("pair_id: {}", pair_id);
let size: ImageSize;
let expected_image: &Image;
match pair.pair_type {
PairType::Test => {
// let image: &Image = &pair.output.test_image;
// expected_image = image;
// size = ImageSize { width: image.width(), height: image.height() };
continue;
},
PairType::Train => {
let image: &Image = &pair.output.image;
expected_image = image;
size = ImageSize { width: image.width(), height: image.height() };
}
}
let computed_image: Image = self.query(pair_id, size)?;
// measure difference from expected image
let diff: Image = computed_image.diff(expected_image)?;
let intersection: u16 = diff.mask_count_zero();
let union: u16 = (size.width as u16) * (size.height as u16);
if union == 0 {
return Err(anyhow::anyhow!("Encountered a task with an empty image. {}", pair.id));
}
let jaccard_index: f32 = (intersection as f32) / (union as f32);
println!("pair: {} jaccard_index: {}", pair.id, jaccard_index);
HtmlLog::text(format!("pair: {}", pair.id));
HtmlLog::image(&expected_image);
HtmlLog::image(&computed_image);
HtmlLog::image(&diff);
}
// undo if the mutation was too poor
// the global weights stays locked while training for a single task
// local weights that gets updated while training with a single task
// mutate the global weights.
self.mutate_global_weights(&mut rng);
// repeat training
Ok(())
}
fn mutate_local_weights_with_sample(&mut self, sample: &Sample, rng: &mut StdRng) {
for y in 0..3 {
for x in 0..3 {
let _pixel: u8 = sample.convolution3x3.get(x as i32, y as i32).unwrap_or(255);
// propagate pixel value to all local weights, based on position, pair_id, input/output
// update local weights
}
}
self.mutate_local_weights(rng);
}
fn mutate_global_weights(&mut self, rng: &mut StdRng) {
let step = Uniform::<u16>::new(0, 1001);
for weight in self.global_weights.iter_mut() {
let random_value: u16 = step.sample(rng);
let weight_between0and1: f32 = (random_value as f32) / 1000.0;
let adjustment: f32 = (weight_between0and1 - 0.5) / 100.0;
*weight = (*weight + adjustment).max(2.0).min(1.0);
}
}
fn mutate_local_weights(&mut self, rng: &mut StdRng) {
let step = Uniform::<u16>::new(0, 1001);
for weight in self.local_weights.iter_mut() {
let random_value: u16 = step.sample(rng);
let weight_between0and1: f32 = (random_value as f32) / 1000.0;
let adjustment: f32 = (weight_between0and1 - 0.5) / 100.0;
*weight = (*weight + adjustment).max(2.0).min(1.0);
}
}
fn query(&self, pair_id: f32, size: ImageSize) -> anyhow::Result<Image> {
let mut result_image = Image::zero(size.width, size.height);
for y in 0..size.height {
let yy: f32 = ((y as f32) + 1.0) / ((size.height as f32) + 1.0);
for x in 0..size.width {
let xx: f32 = ((x as f32) + 1.0) / ((size.width as f32) + 1.0);
let color: u8 = self.query_xy(pair_id, xx, yy)?;
_ = result_image.set(x as i32, y as i32, color);
}
}
Ok(result_image)
}
fn query_xy(&self, pair_id: f32, x: f32, y: f32) -> anyhow::Result<u8> {
// println!("pair_id: {} x: {} y: {}", pair_id, x, y);
// Experiments with multiple iterations and passing data to the next iteration
for i in 0..1 {
let global_address: usize = 3 * i;
let g0: f32 = self.global_weights[global_address + 0] * pair_id;
let g1: f32 = self.global_weights[global_address + 1] * x;
let g2: f32 = (2.0 - self.global_weights[global_address + 1]) * x;
let g3: f32 = self.global_weights[global_address + 2] * y;
let g4: f32 = (2.0 - self.global_weights[global_address + 2]) * y;
let local_address: usize = 19 * i;
let l0: f32 = self.local_weights[local_address + 0] * g0;
let l1: f32 = self.local_weights[local_address + 1] * g1;
let l2: f32 = self.local_weights[local_address + 2] * g2;
let l3: f32 = self.local_weights[local_address + 3] * g3;
let l4: f32 = self.local_weights[local_address + 4] * g4;
let sum: f32 = l0 + l1 + l2 + l3 + l4;
let product: f32 = l0 * l1 * l2 * l3 * l4;
let min: f32 = l0.min(l1).min(l2).min(l3).min(l4);
let max: f32 = l0.max(l1).max(l2).max(l3).max(l4);
let minus01: f32 = l0 - l1;
let minus12: f32 = l1 - l2;
let minus23: f32 = l2 - l3;
let minus34: f32 = l3 - l4;
let values: [f32; 13] = [
self.local_weights[local_address + 5] * l0,
self.local_weights[local_address + 6] * l1,
self.local_weights[local_address + 7] * l2,
self.local_weights[local_address + 8] * l3,
self.local_weights[local_address + 9] * l4,
self.local_weights[local_address + 10] * sum,
self.local_weights[local_address + 11] * product,
self.local_weights[local_address + 12] * min,
self.local_weights[local_address + 13] * max,
self.local_weights[local_address + 15] * minus01,
self.local_weights[local_address + 16] * minus12,
self.local_weights[local_address + 17] * minus23,
self.local_weights[local_address + 18] * minus34,
];
let mut found_value: f32 = f32::MIN;
let mut found_index: usize = 0;
for (value_index, value) in values.iter().enumerate() {
if *value > found_value {
found_value = *value;
found_index = value_index;
}
}
// println!("pair_id: {} x: {} y: {} value: {}", pair_id, x, y, found_value);
let color: u8 = u8::try_from(found_index).unwrap_or(255);
return Ok(color);
}
let color: u8 = 0;
Ok(color)
}
fn extract_samples(input: &Image, input_output_type: InputOutputType) -> anyhow::Result<Vec<Sample>> {
let padded_image: Image = input.padding_with_color(1, 255)?;
let width: u8 = padded_image.width();
let height: u8 = padded_image.height();
if width < 3 || height < 3 {
return Err(anyhow::anyhow!("too small image, must be 3x3 or bigger"));
}
let mut samples = Vec::<Sample>::new();
let mut conv_bitmap = Image::zero(3, 3);
let image_width: u8 = input.width();
let image_height: u8 = input.height();
for self_y in 0..image_height {
for self_x in 0..image_width {
for conv_y in 0..3u8 {
for conv_x in 0..3u8 {
let get_x: i32 = (self_x as i32) + (conv_x as i32);
let get_y: i32 = (self_y as i32) + (conv_y as i32);
let pixel_value: u8 = padded_image.get(get_x, get_y)
.ok_or_else(|| anyhow::anyhow!("image.get({},{}) returned None", get_x, get_y))?;
conv_bitmap.set(conv_x as i32, conv_y as i32, pixel_value)
.ok_or_else(|| anyhow::anyhow!("conv_bitmap.set({},{}) returned None", conv_x, conv_y))?;
}
}
let sample = Sample {
convolution3x3: conv_bitmap.clone(),
position_x: self_x,
position_y: self_y,
image_width,
image_height,
input_output_type,
};
samples.push(sample);
}
}
Ok(samples)
}
}
|