Add Hiera vision model. (#2382)

2 files changed, 117 insertions, 0 deletions

diff --git a/candle-examples/examples/hiera/README.md b/candle-examples/examples/hiera/README.md
new file mode 100644
index 00000000..763ce1a5
--- /dev/null
+++ b/candle-examples/examples/hiera/README.md
@@ -0,0 +1,18 @@
+# hiera
+
+[Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles](https://arxiv.org/abs/2306.00989)
+This candle implementation uses pre-trained Hiera models from timm for inference.
+The classification head has been trained on the ImageNet dataset and returns the probabilities for the top-5 classes.
+
+## Running an example
+
+```
+$ cargo run --example  hiera --release  -- --image candle-examples/examples/yolo-v8/assets/bike.jpg --which tiny
+loaded image Tensor[dims 3, 224, 224; f32]
+model built
+mountain bike, all-terrain bike, off-roader: 71.15%
+unicycle, monocycle     : 7.11%
+knee pad                : 4.26%
+crash helmet            : 1.48%
+moped                   : 1.07%
+```
diff --git a/candle-examples/examples/hiera/main.rs b/candle-examples/examples/hiera/main.rs
new file mode 100644
index 00000000..55bb1d54
--- /dev/null
+++ b/candle-examples/examples/hiera/main.rs
@@ -0,0 +1,99 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use clap::{Parser, ValueEnum};
+
+use candle::{DType, IndexOp, D};
+use candle_nn::{Module, VarBuilder};
+use candle_transformers::models::hiera;
+
+#[derive(Clone, Copy, Debug, ValueEnum)]
+enum Which {
+    Tiny,
+    Small,
+    Base,
+    BasePlus,
+    Large,
+    Huge,
+}
+
+impl Which {
+    fn model_filename(&self) -> String {
+        let name = match self {
+            Self::Tiny => "tiny",
+            Self::Small => "small",
+            Self::Base => "base",
+            Self::BasePlus => "base_plus",
+            Self::Large => "large",
+            Self::Huge => "huge",
+        };
+        format!("timm/hiera_{}_224.mae_in1k_ft_in1k", name)
+    }
+
+    fn config(&self) -> hiera::Config {
+        match self {
+            Self::Tiny => hiera::Config::tiny(),
+            Self::Small => hiera::Config::small(),
+            Self::Base => hiera::Config::base(),
+            Self::BasePlus => hiera::Config::base_plus(),
+            Self::Large => hiera::Config::large(),
+            Self::Huge => hiera::Config::huge(),
+        }
+    }
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(long)]
+    model: Option<String>,
+
+    #[arg(long)]
+    image: String,
+
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    #[arg(value_enum, long, default_value_t=Which::Tiny)]
+    which: Which,
+}
+
+pub fn main() -> anyhow::Result<()> {
+    let args = Args::parse();
+
+    let device = candle_examples::device(args.cpu)?;
+
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
+    println!("loaded image {image:?}");
+
+    let model_file = match args.model {
+        None => {
+            let model_name = args.which.model_filename();
+            let api = hf_hub::api::sync::Api::new()?;
+            let api = api.model(model_name);
+            api.get("model.safetensors")?
+        }
+        Some(model) => model.into(),
+    };
+
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
+    let model = hiera::hiera(&args.which.config(), 1000, vb)?;
+    println!("model built");
+    let logits = model.forward(&image.unsqueeze(0)?)?;
+    let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
+        .i(0)?
+        .to_vec1::<f32>()?;
+    let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
+    prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
+    for &(category_idx, pr) in prs.iter().take(5) {
+        println!(
+            "{:24}: {:.2}%",
+            candle_examples::imagenet::CLASSES[category_idx],
+            100. * pr
+        );
+    }
+    Ok(())
+}