feat(yolo): integrate slsl::Tensor to accelerate preprocessing, inference, and postprocessing (#143)

* Attempt to introduce `slsl`

* update: hub download and system font

* Accelerate image pre-processing

Author: jamjamjon

.github/dependabot.yml

@@ -4,4 +4,4 @@ updates:
     directory: "/"
     schedule:
       interval: "monthly"
-    open-pull-requests-limit: 3
+    open-pull-requests-limit: 5

Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "usls"
 edition = "2021"
-version = "0.1.0"
+version = "0.1.1"
 rust-version = "1.85"
 description = "A Rust library integrated with ONNXRuntime, providing a collection of ML models."
 repository = "https://github.com/jamjamjon/usls"
@@ -18,7 +18,7 @@ ab_glyph = { version = "0.2.29" }
 image = { version = "0.25" }
 imageproc = { version = "0.25" }
 ndarray = { version = "0.16.1", features = ["rayon", "serde"] }
-indicatif = { version = "0.17.11" }
+indicatif = { version = "0.18" }
 log = "0.4.26"
 minifb = { version = "0.28.0", optional = true }
 rand = { version = "0.9" }
@@ -30,12 +30,12 @@ rayon = { version = "1.10.0" }
 glob = "0.3.2"
 dirs = "6.0.0"
 natord = "1.0.9"
-geo = "0.30.0"
+geo = "0.31.0"
 chrono = "0.4.40"
 regex = "1.11.1"
 tempfile = "3.19.1"
 video-rs = { version = "0.10.3", features = ["ndarray"], optional = true }
-fast_image_resize = { version = "=5.1.2", features = ["image"] }
+fast_image_resize = { version = "=5.1.4", features = ["image", "rayon"] }
 ndarray-npy = "0.9.1"
 half = { version = "2.6.0", features = ["bytemuck", "num-traits"] }
 prost = "0.14.1"
@@ -47,6 +47,7 @@ ort = { version = "=2.0.0-rc.10", default-features = false, optional = true, fea
 hf-hub = { version = "0.4.3", default-features = false, features = ["ureq", "native-tls"] }
 tokenizers = { version = "0.21.1" }
 paste = "1.0.15"
+slsl = { version = "0.0.5", features = ["rayon"] }
 
 
 [dev-dependencies]

examples/clip/main.rs

@@ -1,9 +1,9 @@
 use anyhow::Result;
-use ndarray::Axis;
+use slsl::Tensor;
 use usls::{models::Clip, Config, DataLoader};
 
-#[derive(argh::FromArgs)]
 /// CLIP Example
+#[derive(argh::FromArgs)]
 struct Args {
     /// device
     #[argh(option, default = "String::from(\"cpu:0\")")]
@@ -44,22 +44,29 @@ fn main() -> Result<()> {
         "There is a doll with red hair and a clock on a table.",
         "Some people holding wine glasses in a restaurant.",
     ];
-    let feats_text = model.encode_texts(&texts)?.norm(1)?;
+
+    // encode texts
+    let feats_text = model.encode_texts(&texts)?;
+    let feats_text_norm = feats_text.norm_l2_keepdim(-1)?.to_dtype::<f32>()?;
+    let feats_text = (feats_text / feats_text_norm).t()?;
 
     // load images
     let dl = DataLoader::new("./examples/clip/images")?.build()?;
 
     // run
     for images in &dl {
-        let feats_image = model.encode_images(&images)?.norm(1)?;
+        // encode image
+        let feats_image: Tensor = model.encode_images(&images)?;
+        let feats_image_norm = feats_image.norm_l2_keepdim(-1)?.to_dtype::<f32>()?;
+        let feats_image = feats_image / feats_image_norm;
 
         // use image to query texts
-        let matrix = (feats_image * 100.).dot2(&feats_text)?.softmax(1)?;
+        let matrix = (feats_image * 100.0f32).matmul(&feats_text)?.softmax(-1)?;
 
         // Process each image's matching scores
-        for (i, row) in matrix.axis_iter(Axis(0)).enumerate() {
+        for (i, row) in matrix.iter_dim(0).enumerate() {
             let (id, &score) = row
-                .iter()
+                .iter::<f32>()
                 .enumerate()
                 .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
                 .unwrap();
@@ -72,6 +79,7 @@ fn main() -> Result<()> {
             );
         }
     }
+
     usls::perf(false);
 
     Ok(())

examples/yolo/main.rs

@@ -171,7 +171,7 @@ fn main() -> Result<()> {
         .with_topk(args.topk)
         .retain_classes(&args.retain_classes)
         .exclude_classes(&args.exclude_classes)
-        .with_model_num_dry_run(2);
+        .with_model_num_dry_run(5);
     if args.use_coco_80_classes {
         config = config.with_class_names(&NAMES_COCO_80);
     }
@@ -234,7 +234,7 @@ fn main() -> Result<()> {
     // run & annotate
     for xs in &dl {
         let ys = model.forward(&xs)?;
-        // println!("ys: {:?}", ys);
+        println!("ys: {:?}", ys);
 
         for (x, y) in xs.iter().zip(ys.iter()) {
             annotator.annotate(x, y)?.save(format!(