Switch to image::imageops::resize for resizing in the imagenet demo

The `imageops::resize` function produces better quality outputs when the output
is small, and has been significantly downscaled from the input, because it does
antialiasing. The the implementation in the `Resize` operator does not.

This improves the alignment between prediction outputs of the
`tools/export-timm-model.py` script and the imagenet demo. The results are still
not perfectly identical because the transforms don't exactly match.

examples/imagenet.rs

@@ -58,9 +58,27 @@ fn read_image<N: Fn(usize, f32) -> f32>(
     normalize_pixel: N,
     out_chan_order: ChannelOrder,
     out_dim_order: DimOrder,
+    out_height: u32,
+    out_width: u32,
 ) -> Result<Tensor<f32>, Box<dyn Error>> {
     let input_img = image::open(path)?;
     let input_img = input_img.into_rgb8();
+
+    // Resize the image using the `imageops::resize` function from the `image`
+    // crate rather than using Wasnn's `resize` operator because
+    // `imageops::resize` supports antialiasing. This significantly improves
+    // output image quality and thus prediction accuracy when the output is
+    // small (eg. 224 or 256px).
+    //
+    // The outputs of `imageops::resize` still don't match PyTorch exactly
+    // though, which can lead to small differences in prediction outputs.
+    let input_img = image::imageops::resize(
+        &input_img,
+        out_width,
+        out_height,
+        image::imageops::FilterType::Triangle,
+    );
+
     let (width, height) = input_img.dimensions();
 
     // Map input channel index, in RGB order, to output channel index
@@ -197,18 +215,6 @@ fn main() -> Result<(), Box<dyn Error>> {
         },
     };
 
-    let img_tensor = read_image(
-        &args.image,
-        normalize_pixel,
-        args.config.chan_order,
-        args.config.dim_order,
-    )?;
-
-    let (height, width) = match args.config.dim_order {
-        DimOrder::Nchw => (img_tensor.size(2), img_tensor.size(3)),
-        DimOrder::Nhwc => (img_tensor.size(1), img_tensor.size(2)),
-    };
-
     let input_id = model
         .input_ids()
         .get(0)
@@ -237,11 +243,14 @@ fn main() -> Result<(), Box<dyn Error>> {
         }
     };
 
-    let img_tensor = if height != in_height as usize || width != in_width as usize {
-        img_tensor.resize_image([in_height, in_width])?
-    } else {
-        img_tensor
-    };
+    let img_tensor = read_image(
+        &args.image,
+        normalize_pixel,
+        args.config.chan_order,
+        args.config.dim_order,
+        in_height as u32,
+        in_width as u32,
+    )?;
 
     let logits: NdTensor<f32, 2> = model.run_one(img_tensor.view().into(), None)?.try_into()?;