Add Resize to onnx ops (#2946)

* added resize to candle-onnx, not currently working

* changed unreachable to bail, and bailed when both scales and sizes are set

* cleanup and added other unused options for this op

* cleanup

* fixed image loading to make output work

* cleanup and removed unused variables

* removed path path creation code, and changed unwrap to ?

candle-examples/examples/onnx/main.rs

@@ -5,12 +5,14 @@ extern crate intel_mkl_src;
 extern crate accelerate_src;
 
 use candle::{IndexOp, D};
+use candle_examples::save_image;
 use clap::{Parser, ValueEnum};
 
 #[derive(Clone, Copy, Debug, ValueEnum)]
 enum Which {
     SqueezeNet,
     EfficientNet,
+    EsrGan,
 }
 
 #[derive(Parser)]
@@ -28,10 +30,21 @@ struct Args {
 
 pub fn main() -> anyhow::Result<()> {
     let args = Args::parse();
-    let image = candle_examples::imagenet::load_image224(args.image)?;
+    let image = match args.which {
+        Which::SqueezeNet | Which::EfficientNet => {
+            candle_examples::imagenet::load_image224(&args.image)?
+        }
+        Which::EsrGan => candle_examples::imagenet::load_image_with_std_mean(
+            &args.image,
+            128,
+            &[0.0f32, 0.0, 0.0],
+            &[1.0f32, 1.0, 1.0],
+        )?,
+    };
     let image = match args.which {
         Which::SqueezeNet => image,
         Which::EfficientNet => image.permute((1, 2, 0))?,
+        Which::EsrGan => image,
     };
 
     println!("loaded image {image:?}");
@@ -45,6 +58,9 @@ pub fn main() -> anyhow::Result<()> {
             Which::EfficientNet => hf_hub::api::sync::Api::new()?
                 .model("onnx/EfficientNet-Lite4".into())
                 .get("efficientnet-lite4-11.onnx")?,
+            Which::EsrGan => hf_hub::api::sync::Api::new()?
+                .model("qualcomm/Real-ESRGAN-x4plus".into())
+                .get("Real-ESRGAN-x4plus.onnx")?,
         },
     };
 
@@ -57,7 +73,11 @@ pub fn main() -> anyhow::Result<()> {
     let prs = match args.which {
         Which::SqueezeNet => candle_nn::ops::softmax(&output, D::Minus1)?,
         Which::EfficientNet => output,
+        Which::EsrGan => output,
     };
+
+    match args.which {
+        Which::EfficientNet | Which::SqueezeNet => {
             let prs = prs.i(0)?.to_vec1::<f32>()?;
 
             // Sort the predictions and take the top 5
@@ -73,6 +93,21 @@ pub fn main() -> anyhow::Result<()> {
                     p * 100.0
                 );
             }
+        }
+        Which::EsrGan => {
+            let max_pixel_val = candle::Tensor::try_from(255.0f32)?
+                .to_device(prs.device())?
+                .broadcast_as(prs.shape())?;
+            let out = (prs * max_pixel_val)?.i(0)?.to_dtype(candle::DType::U8)?;
+
+            let pb = std::path::PathBuf::from(args.image);
+            let input_file_name = pb.file_name().unwrap();
+            let mut output_file_name = std::ffi::OsString::from("super_");
+            output_file_name.push(input_file_name);
+
+            save_image(&out, output_file_name)?;
+        }
+    }
 
     Ok(())
 }

candle-onnx/src/eval.rs

@@ -1960,6 +1960,76 @@ fn simple_eval_(
                 let output = input.sign()?;
                 values.insert(node.output[0].clone(), output);
             }
+            "Resize" => {
+                let input = get(&node.input[0])?;
+
+                if input.rank() != 4 {
+                    bail!("Unsupported rank for nearest resize: {}", input.rank());
+                }
+
+                let scales = if node.input.len() > 2 && !node.input[2].is_empty() {
+                    Some(get(&node.input[2])?)
+                } else {
+                    None
+                };
+
+                let sizes = if node.input.len() > 3 && !node.input[3].is_empty() {
+                    Some(get(&node.input[3])?)
+                } else {
+                    None
+                };
+
+                let output_dims = match (scales, sizes) {
+                    (Some(_), Some(_)) => {
+                        bail!("Scales and sizes cannot both be set for Resize operation")
+                    }
+                    (Some(scales_tensor), None) => {
+                        let scale_values = scales_tensor.to_vec1::<f32>()?;
+                        input
+                            .dims()
+                            .iter()
+                            .enumerate()
+                            .map(|(i, &d)| (d as f32 * scale_values[i]) as usize)
+                            .collect::<Vec<_>>()
+                    }
+                    (None, Some(sizes_tensor)) => sizes_tensor
+                        .to_vec1::<i64>()?
+                        .iter()
+                        .map(|&d| d as usize)
+                        .collect::<Vec<_>>(),
+                    (None, None) => bail!("Either scales or sizes should be present"),
+                };
+
+                let coordinate_transformation_mode =
+                    get_attr_opt::<str>(node, "coordinate_transformation_mode")?
+                        .unwrap_or("half_pixel");
+                // Interpolation mode: nearest, linear, or cubic.
+                let mode = get_attr_opt::<str>(node, "mode")?.unwrap_or("nearest");
+                // How to determine the "nearest" pixel in nearest interpolation mode.
+                let nearest_mode =
+                    get_attr_opt::<str>(node, "nearest_mode")?.unwrap_or("round_prefer_floor");
+
+                if mode != "nearest" {
+                    bail!("Unsupported resize mode: {}", mode);
+                }
+
+                if nearest_mode != "floor" {
+                    bail!("Unsupported nearest_mode for resize: {}", nearest_mode);
+                }
+
+                if coordinate_transformation_mode != "asymmetric" {
+                    bail!(
+                        "Unsupported coordinate_transformation_mode for resize: {}",
+                        coordinate_transformation_mode
+                    );
+                }
+
+                let h = output_dims[2];
+                let w = output_dims[3];
+                let output = input.upsample_nearest2d(h, w)?;
+
+                values.insert(node.output[0].clone(), output);
+            }
             op_type => bail!("unsupported op_type {op_type} for op {node:?}"),
         }
     }