Files
usls/src/models/yolo/impl.rs
Jamjamjon 475a680703 Add moondream2
* Add moondream2

* Update README.md
2025-01-20 21:37:54 +08:00

632 lines
26 KiB
Rust

use aksr::Builder;
use anyhow::Result;
use image::DynamicImage;
use log::{error, info};
use ndarray::{s, Array, Axis};
use rayon::prelude::*;
use regex::Regex;
use crate::{
elapsed,
models::{BoxType, YOLOPredsFormat},
Bbox, DynConf, Engine, Keypoint, Mask, Mbr, Ops, Options, Polygon, Prob, Processor, Task, Ts,
Version, Xs, Ys, Y,
};
#[derive(Debug, Builder)]
pub struct YOLO {
engine: Engine,
height: usize,
width: usize,
batch: usize,
layout: YOLOPredsFormat,
task: Task,
version: Option<Version>,
names: Vec<String>,
names_kpt: Vec<String>,
nc: usize,
nk: usize,
confs: DynConf,
kconfs: DynConf,
iou: f32,
find_contours: bool,
processor: Processor,
ts: Ts,
spec: String,
classes_excluded: Vec<usize>,
classes_retained: Vec<usize>,
}
impl TryFrom<Options> for YOLO {
type Error = anyhow::Error;
fn try_from(options: Options) -> Result<Self, Self::Error> {
Self::new(options)
}
}
impl YOLO {
pub fn new(options: Options) -> Result<Self> {
let engine = options.to_engine()?;
let (batch, height, width, ts, spec) = (
engine.batch().opt(),
engine.try_height().unwrap_or(&640.into()).opt(),
engine.try_width().unwrap_or(&640.into()).opt(),
engine.ts.clone(),
engine.spec().to_owned(),
);
let processor = options
.to_processor()?
.with_image_width(width as _)
.with_image_height(height as _);
let task: Option<Task> = match &options.model_task {
Some(task) => Some(task.clone()),
None => match engine.try_fetch("task") {
Some(x) => match x.as_str() {
"classify" => Some(Task::ImageClassification),
"detect" => Some(Task::ObjectDetection),
"pose" => Some(Task::KeypointsDetection),
"segment" => Some(Task::InstanceSegmentation),
"obb" => Some(Task::OrientedObjectDetection),
x => {
error!("Unsupported YOLO Task: {}", x);
None
}
},
None => None,
},
};
// Task & layout
let version = options.model_version;
let (layout, task) = match &options.yolo_preds_format {
// customized
Some(layout) => {
// check task
let task_parsed = layout.task();
let task = match task {
Some(task) => {
if task_parsed != task {
anyhow::bail!(
"Task specified: {:?} is inconsistent with parsed from yolo_preds_format: {:?}",
task,
task_parsed
);
}
task_parsed
}
None => task_parsed,
};
(layout.clone(), task)
}
// version + task
None => match (task, version) {
(Some(task), Some(version)) => {
let layout = match (task.clone(), version) {
(Task::ImageClassification, Version(5, 0)) => {
YOLOPredsFormat::n_clss().apply_softmax(true)
}
(Task::ImageClassification, Version(8, 0) | Version(11, 0)) => {
YOLOPredsFormat::n_clss()
}
(Task::ObjectDetection, Version(5, 0) | Version(6, 0) | Version(7, 0)) => {
YOLOPredsFormat::n_a_cxcywh_confclss()
}
(Task::ObjectDetection, Version(8, 0) | Version(9, 0) | Version(11, 0)) => {
YOLOPredsFormat::n_cxcywh_clss_a()
}
(Task::ObjectDetection, Version(10, 0)) => {
YOLOPredsFormat::n_a_xyxy_confcls().apply_nms(false)
}
(Task::KeypointsDetection, Version(8, 0) | Version(11, 0)) => {
YOLOPredsFormat::n_cxcywh_clss_xycs_a()
}
(Task::InstanceSegmentation, Version(5, 0)) => {
YOLOPredsFormat::n_a_cxcywh_confclss_coefs()
}
(Task::InstanceSegmentation, Version(8, 0) | Version(11, 0)) => {
YOLOPredsFormat::n_cxcywh_clss_coefs_a()
}
(Task::OrientedObjectDetection, Version(8, 0) | Version(11, 0)) => {
YOLOPredsFormat::n_cxcywh_clss_r_a()
}
(task, version) => {
anyhow::bail!("Task: {:?} is unsupported for Version: {:?}. Try using `.with_yolo_preds()` for customization.", task, version)
}
};
(layout, task)
}
(None, Some(version)) => {
let layout = match version {
// single task, no need to specified task
Version(6, 0) | Version(7, 0) => YOLOPredsFormat::n_a_cxcywh_confclss(),
Version(9, 0) => YOLOPredsFormat::n_cxcywh_clss_a(),
Version(10, 0) => YOLOPredsFormat::n_a_xyxy_confcls().apply_nms(false),
_ => {
anyhow::bail!(
"No clear YOLO Task specified for Version: {:?}.",
version
)
}
};
(layout, Task::ObjectDetection)
}
(Some(task), None) => {
anyhow::bail!("No clear YOLO Version specified for Task: {:?}.", task)
}
(None, None) => {
anyhow::bail!("No clear YOLO Task and Version specified.")
}
},
};
// Class names
let names: Option<Vec<String>> = match Self::fetch_names_from_onnx(&engine) {
Some(names_parsed) => match &options.class_names {
Some(names) => {
if names.len() == names_parsed.len() {
// prioritize user-defined
Some(names.clone())
} else {
// Fail to override
anyhow::bail!(
"The lengths of parsed class names: {} and user-defined class names: {} do not match.",
names_parsed.len(),
names.len(),
)
}
}
None => Some(names_parsed),
},
None => options.class_names.clone(),
};
// Class names & Number of class
let (nc, names) = match (options.nc(), names) {
(_, Some(names)) => (names.len(), names.to_vec()),
(Some(nc), None) => (nc, Self::n2s(nc)),
(None, None) => {
anyhow::bail!(
"Neither class names nor the number of classes were specified. \
\nConsider specify them with `Options::default().with_nc()` or `Options::default().with_class_names()`"
);
}
};
// Keypoint names & Number of keypoints
let (nk, names_kpt) = if let Task::KeypointsDetection = task {
let nk = Self::fetch_nk_from_onnx(&engine).or(options.nk());
match (&options.keypoint_names, nk) {
(Some(names), Some(nk)) => {
if names.len() != nk {
anyhow::bail!(
"The lengths of user-defined keypoint names: {} and nk parsed: {} do not match.",
names.len(),
nk,
);
}
(nk, names.clone())
}
(Some(names), None) => (names.len(), names.clone()),
(None, Some(nk)) => (nk, Self::n2s(nk)),
(None, None) => anyhow::bail!(
"Neither keypoint names nor the number of keypoints were specified when doing `KeypointsDetection` task. \
\nConsider specify them with `Options::default().with_nk()` or `Options::default().with_keypoint_names()`"
),
}
} else {
(0, vec![])
};
// Attributes
let confs = DynConf::new(options.class_confs(), nc);
let kconfs = DynConf::new(options.keypoint_confs(), nk);
let iou = options.iou().unwrap_or(0.45);
let classes_excluded = options.classes_excluded().to_vec();
let classes_retained = options.classes_retained().to_vec();
let find_contours = options.find_contours();
let mut info = format!(
"YOLO Version: {}, Task: {:?}, Category Count: {}, Keypoint Count: {}",
version.map_or("Unknown".into(), |x| x.to_string()),
task,
nc,
nk,
);
if !classes_excluded.is_empty() {
info = format!("{}, classes_excluded: {:?}", info, classes_excluded);
}
if !classes_retained.is_empty() {
info = format!("{}, classes_retained: {:?}", info, classes_retained);
}
info!("{}", info);
Ok(Self {
engine,
height,
width,
batch,
task,
version,
spec,
layout,
names,
names_kpt,
confs,
kconfs,
iou,
nc,
nk,
find_contours,
classes_excluded,
classes_retained,
processor,
ts,
})
}
fn preprocess(&mut self, xs: &[DynamicImage]) -> Result<Xs> {
let x = self.processor.process_images(xs)?;
Ok(x.into())
}
fn inference(&mut self, xs: Xs) -> Result<Xs> {
self.engine.run(xs)
}
pub fn forward(&mut self, xs: &[DynamicImage]) -> Result<Ys> {
let ys = elapsed!("preprocess", self.ts, { self.preprocess(xs)? });
let ys = elapsed!("inference", self.ts, { self.inference(ys)? });
let ys = elapsed!("postprocess", self.ts, { self.postprocess(ys)? });
Ok(ys)
}
pub fn summary(&mut self) {
self.ts.summary();
}
fn postprocess(&self, xs: Xs) -> Result<Ys> {
let protos = if xs.len() == 2 { Some(&xs[1]) } else { None };
let ys: Vec<Y> = xs[0]
.axis_iter(Axis(0))
.into_par_iter()
.enumerate()
.filter_map(|(idx, preds)| {
let mut y = Y::default();
// Parse predictions
let (
slice_bboxes,
slice_id,
slice_clss,
slice_confs,
slice_kpts,
slice_coefs,
slice_radians,
) = self.layout.parse_preds(preds, self.nc);
// ImageClassifcation
if let Task::ImageClassification = self.task {
let x = if self.layout.apply_softmax {
let exps = slice_clss.mapv(|x| x.exp());
let stds = exps.sum_axis(Axis(0));
exps / stds
} else {
slice_clss.into_owned()
};
let probs = Prob::default()
.with_probs(&x.into_raw_vec_and_offset().0)
.with_names(&self.names.iter().map(|x| x.as_str()).collect::<Vec<_>>());
return Some(y.with_probs(probs));
}
// Original image size
let (image_height, image_width) = self.processor.image0s_size[idx];
let ratio = self.processor.scale_factors_hw[idx][0];
// Other tasks
let (y_bboxes, y_mbrs) = slice_bboxes?
.axis_iter(Axis(0))
.into_par_iter()
.enumerate()
.filter_map(|(i, bbox)| {
// confidence & class_id
let (class_id, confidence) = match &slice_id {
Some(ids) => (ids[[i, 0]] as _, slice_clss[[i, 0]] as _),
None => {
let (class_id, &confidence) = slice_clss
.slice(s![i, ..])
.into_iter()
.enumerate()
.max_by(|a, b| a.1.total_cmp(b.1))?;
match &slice_confs {
None => (class_id, confidence),
Some(slice_confs) => {
(class_id, confidence * slice_confs[[i, 0]])
}
}
}
};
// filter out class id
if !self.classes_excluded.is_empty()
&& self.classes_excluded.contains(&class_id)
{
return None;
}
// filter by class id
if !self.classes_retained.is_empty()
&& !self.classes_retained.contains(&class_id)
{
return None;
}
// filter by conf
if confidence < self.confs[class_id] {
return None;
}
// Bboxes
let bbox = bbox.mapv(|x| x / ratio);
let bbox = if self.layout.is_bbox_normalized {
(
bbox[0] * self.width() as f32,
bbox[1] * self.height() as f32,
bbox[2] * self.width() as f32,
bbox[3] * self.height() as f32,
)
} else {
(bbox[0], bbox[1], bbox[2], bbox[3])
};
let (cx, cy, x, y, w, h) = match self.layout.box_type()? {
BoxType::Cxcywh => {
let (cx, cy, w, h) = bbox;
let x = (cx - w / 2.).max(0.);
let y = (cy - h / 2.).max(0.);
(cx, cy, x, y, w, h)
}
BoxType::Xyxy => {
let (x, y, x2, y2) = bbox;
let (w, h) = (x2 - x, y2 - y);
let (cx, cy) = ((x + x2) / 2., (y + y2) / 2.);
(cx, cy, x, y, w, h)
}
BoxType::Xywh => {
let (x, y, w, h) = bbox;
let (cx, cy) = (x + w / 2., y + h / 2.);
(cx, cy, x, y, w, h)
}
BoxType::Cxcyxy => {
let (cx, cy, x2, y2) = bbox;
let (w, h) = ((x2 - cx) * 2., (y2 - cy) * 2.);
let x = (x2 - w).max(0.);
let y = (y2 - h).max(0.);
(cx, cy, x, y, w, h)
}
BoxType::XyCxcy => {
let (x, y, cx, cy) = bbox;
let (w, h) = ((cx - x) * 2., (cy - y) * 2.);
(cx, cy, x, y, w, h)
}
};
let (y_bbox, y_mbr) = match &slice_radians {
Some(slice_radians) => {
let radians = slice_radians[[i, 0]];
let (w, h, radians) = if w > h {
(w, h, radians)
} else {
(h, w, radians + std::f32::consts::PI / 2.)
};
let radians = radians % std::f32::consts::PI;
let mbr = Mbr::from_cxcywhr(
cx as f64,
cy as f64,
w as f64,
h as f64,
radians as f64,
)
.with_confidence(confidence)
.with_id(class_id as isize)
.with_name(&self.names[class_id]);
(None, Some(mbr))
}
None => {
let bbox = Bbox::default()
.with_xywh(x, y, w, h)
.with_confidence(confidence)
.with_id(class_id as isize)
.with_id_born(i as isize)
.with_name(&self.names[class_id]);
(Some(bbox), None)
}
};
Some((y_bbox, y_mbr))
})
.collect::<(Vec<_>, Vec<_>)>();
let y_bboxes: Vec<Bbox> = y_bboxes.into_iter().flatten().collect();
let y_mbrs: Vec<Mbr> = y_mbrs.into_iter().flatten().collect();
// Mbrs
if !y_mbrs.is_empty() {
y = y.with_mbrs(&y_mbrs);
if self.layout.apply_nms {
y = y.apply_nms(self.iou);
}
return Some(y);
}
// Bboxes
if !y_bboxes.is_empty() {
y = y.with_bboxes(&y_bboxes);
if self.layout.apply_nms {
y = y.apply_nms(self.iou);
}
}
// KeypointsDetection
if let Some(pred_kpts) = slice_kpts {
let kpt_step = self.layout.kpt_step().unwrap_or(3);
if let Some(bboxes) = y.bboxes() {
let y_kpts = bboxes
.into_par_iter()
.filter_map(|bbox| {
let pred = pred_kpts.slice(s![bbox.id_born(), ..]);
let kpts = (0..self.nk)
.into_par_iter()
.map(|i| {
let kx = pred[kpt_step * i] / ratio;
let ky = pred[kpt_step * i + 1] / ratio;
let kconf = pred[kpt_step * i + 2];
if kconf < self.kconfs[i] {
Keypoint::default()
} else {
Keypoint::default()
.with_id(i as isize)
.with_confidence(kconf)
.with_xy(
kx.max(0.0f32).min(image_width as f32),
ky.max(0.0f32).min(image_height as f32),
)
.with_name(&self.names_kpt[i])
}
})
.collect::<Vec<_>>();
Some(kpts)
})
.collect::<Vec<_>>();
y = y.with_keypoints(&y_kpts);
}
}
// InstanceSegmentation
if let Some(coefs) = slice_coefs {
if let Some(bboxes) = y.bboxes() {
let (y_polygons, y_masks) = bboxes
.into_par_iter()
.filter_map(|bbox| {
let coefs = coefs.slice(s![bbox.id_born(), ..]).to_vec();
let proto = protos.as_ref()?.slice(s![idx, .., .., ..]);
let (nm, mh, mw) = proto.dim();
// coefs * proto => mask
let coefs = Array::from_shape_vec((1, nm), coefs).ok()?; // (n, nm)
let proto = proto.to_shape((nm, mh * mw)).ok()?; // (nm, mh * mw)
let mask = coefs.dot(&proto); // (mh, mw, n)
// Mask rescale
let mask = Ops::resize_lumaf32_u8(
&mask.into_raw_vec_and_offset().0,
mw as _,
mh as _,
image_width as _,
image_height as _,
true,
"Bilinear",
)
.ok()?;
let mut mask: image::ImageBuffer<image::Luma<_>, Vec<_>> =
image::ImageBuffer::from_raw(
image_width as _,
image_height as _,
mask,
)?;
let (xmin, ymin, xmax, ymax) =
(bbox.xmin(), bbox.ymin(), bbox.xmax(), bbox.ymax());
// Using bbox to crop the mask
for (y, row) in mask.enumerate_rows_mut() {
for (x, _, pixel) in row {
if x < xmin as _
|| x > xmax as _
|| y < ymin as _
|| y > ymax as _
{
*pixel = image::Luma([0u8]);
}
}
}
// Find contours
let polygons = if self.find_contours {
let contours: Vec<imageproc::contours::Contour<i32>> =
imageproc::contours::find_contours_with_threshold(&mask, 0);
contours
.into_par_iter()
.map(|x| {
let mut polygon = Polygon::default()
.with_id(bbox.id())
.with_points_imageproc(&x.points)
.verify();
if let Some(name) = bbox.name() {
polygon = polygon.with_name(name);
}
polygon
})
.max_by(|x, y| x.area().total_cmp(&y.area()))?
} else {
Polygon::default()
};
let mut mask = Mask::default().with_mask(mask).with_id(bbox.id());
if let Some(name) = bbox.name() {
mask = mask.with_name(name);
}
Some((polygons, mask))
})
.collect::<(Vec<_>, Vec<_>)>();
if !y_polygons.is_empty() {
y = y.with_polygons(&y_polygons);
}
if !y_masks.is_empty() {
y = y.with_masks(&y_masks);
}
}
}
Some(y)
})
.collect();
Ok(ys.into())
}
fn fetch_names_from_onnx(engine: &Engine) -> Option<Vec<String>> {
// fetch class names from onnx metadata
// String format: `{0: 'person', 1: 'bicycle', 2: 'sports ball', ..., 27: "yellow_lady's_slipper"}`
Regex::new(r#"(['"])([-()\w '"]+)(['"])"#)
.ok()?
.captures_iter(&engine.try_fetch("names")?)
.filter_map(|caps| caps.get(2).map(|m| m.as_str().to_string()))
.collect::<Vec<_>>()
.into()
}
fn fetch_nk_from_onnx(engine: &Engine) -> Option<usize> {
Regex::new(r"(\d+), \d+")
.ok()?
.captures(&engine.try_fetch("kpt_shape")?)
.and_then(|caps| caps.get(1))
.and_then(|m| m.as_str().parse::<usize>().ok())
}
fn n2s(n: usize) -> Vec<String> {
(0..n).map(|x| format!("# {}", x)).collect::<Vec<String>>()
}
}