Segment-anything fixes: avoid normalizing twice. (#767)

* Segment-anything fixes: avoid normalizing twice.

* More fixes for the image aspect ratio.

candle-examples/examples/segment-anything/main.rs

@@ -108,7 +108,8 @@ pub fn main() -> anyhow::Result<()> {
 
     let device = candle_examples::device(args.cpu)?;
 
-    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device)?;
+    let image =
+        candle_examples::load_image(args.image, Some(model_sam::IMAGE_SIZE))?.to_device(&device)?;
     println!("loaded image {image:?}");
 
     let model = match args.model {
@@ -125,7 +126,7 @@ pub fn main() -> anyhow::Result<()> {
     let sam = model_sam::Sam::new(768, 12, 12, &[2, 5, 8, 11], vb)?; // sam_vit_b
 
     let (mask, iou_predictions) = sam.forward(&image, false)?;
-    println!("mask: {mask:?}");
+    println!("mask:\n{mask}");
     println!("iou_predictions: {iou_predictions:?}");
     Ok(())
 }

candle-examples/examples/segment-anything/model_sam.rs

@@ -6,7 +6,7 @@ use crate::model_mask_decoder::MaskDecoder;
 use crate::model_prompt_encoder::PromptEncoder;
 
 const PROMPT_EMBED_DIM: usize = 256;
-const IMAGE_SIZE: usize = 1024;
+pub const IMAGE_SIZE: usize = 1024;
 const VIT_PATCH_SIZE: usize = 16;
 
 #[derive(Debug)]
@@ -90,6 +90,7 @@ impl Sam {
     fn preprocess(&self, img: &Tensor) -> Result<Tensor> {
         let (c, h, w) = img.dims3()?;
         let img = img
+            .to_dtype(DType::F32)?
             .broadcast_sub(&self.pixel_mean)?
             .broadcast_div(&self.pixel_std)?;
         if h > IMAGE_SIZE || w > IMAGE_SIZE {

candle-examples/src/lib.rs

@@ -16,6 +16,34 @@ pub fn device(cpu: bool) -> Result<Device> {
     }
 }
 
+pub fn load_image<P: AsRef<std::path::Path>>(
+    p: P,
+    resize_longest: Option<usize>,
+) -> Result<Tensor> {
+    let img = image::io::Reader::open(p)?
+        .decode()
+        .map_err(candle::Error::wrap)?;
+    let img = match resize_longest {
+        None => img,
+        Some(resize_longest) => {
+            let (height, width) = (img.height(), img.width());
+            let resize_longest = resize_longest as u32;
+            let (height, width) = if height < width {
+                let h = (resize_longest * height) / width;
+                (h, resize_longest)
+            } else {
+                let w = (resize_longest * width) / height;
+                (resize_longest, w)
+            };
+            img.resize_exact(width, height, image::imageops::FilterType::CatmullRom)
+        }
+    };
+    let (height, width) = (img.height() as usize, img.width() as usize);
+    let img = img.to_rgb8();
+    let data = img.into_raw();
+    Tensor::from_vec(data, (height, width, 3), &Device::Cpu)?.permute((2, 0, 1))
+}
+
 pub fn load_image_and_resize<P: AsRef<std::path::Path>>(
     p: P,
     width: usize,