fix: align Rust ML service with verified cargo build

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

apps/ml-service/Cargo.toml

@@ -7,8 +7,8 @@ edition = "2021"
 anyhow = "1.0"
 axum = { version = "0.7", features = ["multipart"] }
 image = "0.25"
-ndarray = "0.15"
-ort = { version = "2.0.0-rc.10", features = ["download-binaries"] }
+ndarray = "0.17"
+ort = { version = "2.0.0-rc.10", features = ["download-binaries", "ndarray"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 tokio = { version = "1.0", features = ["macros", "rt-multi-thread", "net"] }

apps/ml-service/src/image.rs

@@ -15,10 +15,9 @@ use std::io::Cursor;
 pub fn preprocess_image(bytes: &[u8], input_size: u32) -> Result<Array4<f32>, ServiceError> {
     // Decode image from bytes
     let cursor = Cursor::new(bytes);
-    let reader = ImageReader::new(cursor)
-        .map_err(|_| ServiceError::BadRequest("Uploaded file is not a valid image".to_string()))?;
-
-    let image = reader
+    let image = ImageReader::new(cursor)
+        .with_guessed_format()
+        .map_err(|_| ServiceError::BadRequest("Uploaded file is not a valid image".to_string()))?
         .decode()
         .map_err(|_| ServiceError::BadRequest("Uploaded file is not a valid image".to_string()))?;
 
@@ -91,18 +90,9 @@ mod tests {
         // After resizing 2x1 to 2x2, we expect interpolation
         // Check that first pixel channel values are present (at least the first row)
         // The exact values depend on interpolation, but we can verify the structure
-        let first_batch = &array.slice(ndarray::s![0, .., .., ..]);
-        assert_eq!(first_batch.shape(), &[2, 2, 3]);
-
-        // Verify first pixel (0, 0) has 3 channels
-        let pixel_0_0 = &first_batch.slice(ndarray::s![0, 0, ..]);
-        assert_eq!(pixel_0_0.len(), 3);
-
-        // The first pixel should be close to [10, 20, 30] (may be interpolated)
-        // We'll just verify it's in a reasonable range
-        assert!(pixel_0_0[0] >= 5.0 && pixel_0_0[0] <= 25.0);
-        assert!(pixel_0_0[1] >= 15.0 && pixel_0_0[1] <= 35.0);
-        assert!(pixel_0_0[2] >= 25.0 && pixel_0_0[2] <= 45.0);
+        assert_eq!(array[[0, 0, 0, 0]], 10.0);
+        assert_eq!(array[[0, 0, 0, 1]], 20.0);
+        assert_eq!(array[[0, 0, 0, 2]], 30.0);
     }
 
     #[test]

apps/ml-service/src/model.rs

@@ -1,7 +1,7 @@
 use crate::config::LABELS;
 use crate::error::ServiceError;
 use ndarray::Array4;
-use ort::Session;
+use ort::{session::Session, value::TensorRef};
 use serde::Serialize;
 use std::collections::BTreeMap;
 use std::path::Path;
@@ -35,7 +35,7 @@ impl ModelService {
     pub fn new(model_path: &Path, input_size: u32) -> Self {
         let session = Session::builder()
             .ok()
-            .and_then(|builder| builder.commit_from_file(model_path).ok())
+            .and_then(|mut builder| builder.commit_from_file(model_path).ok())
             .map(Mutex::new);
 
         Self {
@@ -71,21 +71,21 @@ impl ModelService {
             .ok_or_else(|| ServiceError::ModelUnavailable("Model is not loaded".to_string()))?;
 
         // Lock the session for thread-safe access
-        let session_guard = session
+        let mut session_guard = session
             .lock()
             .map_err(|_| ServiceError::PredictionFailed("Prediction failed".to_string()))?;
 
-        // Run inference
+        let input = TensorRef::from_array_view(&input)
+            .map_err(|_| ServiceError::PredictionFailed("Prediction failed".to_string()))?;
         let outputs = session_guard
-            .run(ort::inputs![input]?)
+            .run(ort::inputs![input])
             .map_err(|_| ServiceError::PredictionFailed("Prediction failed".to_string()))?;
 
-        // Extract output as f32 vector
         let output_tensor = outputs[0]
             .try_extract_tensor::<f32>()
             .map_err(|_| ServiceError::PredictionFailed("Prediction failed".to_string()))?;
 
-        let probabilities: Vec<f32> = output_tensor.iter().copied().collect();
+        let probabilities: Vec<f32> = output_tensor.1.iter().copied().collect();
 
         Self::prediction_from_probabilities(&probabilities)
     }

apps/ml-service/src/routes.rs

@@ -5,7 +5,6 @@ use crate::error::ServiceError;
 use crate::image::preprocess_image;
 use axum::{
     extract::{State, Multipart},
-    http::StatusCode,
     routing::{get, post},
     Json, Router,
 };