CLAUDE: Add ONNX-Runtime Java inference for C5P DNN front-end (Phase 1, CPU)

PyTorch-trained, ONNX-exported all-conv FCN (per-pixel Vx,Vy,s) run in Java via
ONNX Runtime 1.20.0 (CPU EP). CuasDnnInfer loads the model with a location resolver
(local path / scp user@host:path / http(s) -> ~/.cache/c5p_dnn/, fetching the
model.onnx + external-data .data pair) and runs a float[N][H][W] patch to raw
det/vel/off output. Verified bit-exact vs PyTorch (max abs diff 2.9e-6) via a fixed
test vector. New config param curt_dnn_model (empty default) selects the model,
mirroring the tile_processor_gpu kernel-source default/override scheme. CPU first
(.224 has no cuDNN; 82k net is microseconds/patch); GPU (CUDA/TensorRT EP) and the
CuasDetectRT integration are the next phase.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

pom.xml

@@ -76,6 +76,14 @@
                 <groupId>org.tensorflow</groupId>
                 <artifactId>libtensorflow_jni_gpu</artifactId>
                 <version>1.10.0</version>
+            </dependency>
+            <!-- ONNX Runtime (CPU) for the C5P DNN front-end inference (PyTorch-trained, ONNX-exported). // By Claude on 06/13/2026
+                 CPU first - the 82k net is microseconds/patch and .224 has no cuDNN; swap to
+                 onnxruntime_gpu once cuDNN 9 is present (CUDA EP) or use the TensorRT EP. -->
+            <dependency>
+                <groupId>com.microsoft.onnxruntime</groupId>
+                <artifactId>onnxruntime</artifactId>
+                <version>1.20.0</version>
             </dependency>
 			<dependency>
 			    <groupId>org.apache.ant</groupId>

src/main/java/com/elphel/imagej/cuas/rt/CuasDnnInfer.java

+package com.elphel.imagej.cuas.rt;
+// C5P DNN front-end inference via ONNX Runtime (CPU). // By Claude on 06/13/2026
+// Loads a PyTorch-trained, ONNX-exported all-conv FCN that maps an N-frame patch stack to a
+// per-pixel (Vx,Vy,s) output. Phase 1: CPU, fed from Java float arrays (the convenient testing
+// path, like the old TensorFlow integration). Phase 2 (later): onnxruntime_gpu CUDA/TensorRT EP,
+// device chosen via the existing GPU-detection logic; eventual zero-copy from JCuda CUdeviceptr.
+//
+// Model location (curt_dnn_model): resolved as a local path, scp "user@host:/path", or http(s)
+// URL; remote specs are fetched to ~/.cache/c5p_dnn/ together with the external-data sibling
+// (model.onnx + model.onnx.data, which torch 2.9's exporter splits).
+
+import java.io.ByteArrayOutputStream;
+import java.io.DataInputStream;
+import java.io.File;
+import java.io.FileInputStream;
+import java.io.InputStream;
+import java.net.URL;
+import java.nio.ByteBuffer;
+import java.nio.ByteOrder;
+import java.nio.FloatBuffer;
+import java.nio.file.Files;
+import java.util.Collections;
+
+import ai.onnxruntime.OnnxTensor;
+import ai.onnxruntime.OrtEnvironment;
+import ai.onnxruntime.OrtSession;
+
+public class CuasDnnInfer implements AutoCloseable {
+    private final OrtEnvironment env;
+    private final OrtSession session;
+    private final String inputName;
+    private final int nframes;
+
+    public CuasDnnInfer(String modelSpec, int nframes) throws Exception {
+        String local = resolveModel(modelSpec);
+        this.nframes = nframes;
+        this.env = OrtEnvironment.getEnvironment();
+        OrtSession.SessionOptions opts = new OrtSession.SessionOptions();
+        // CPU EP by default. GPU later: opts.addCUDA(deviceId) once cuDNN 9 is present.
+        this.session = env.createSession(local, opts);
+        this.inputName = session.getInputNames().iterator().next();  // "frames"
+        System.out.println("CuasDnnInfer: loaded " + local + "  input=" + inputName
+                + " outputs=" + session.getOutputNames());
+    }
+
+    /** Run one patch. frames = [N][H][W] (newest first). Returns the raw output channels
+     *  [outCh] for the center output pixel: [0]=det logit, [1..V*V]=vel logits, last 2 = offset. */
+    public float[] inferRaw(float[][][] frames) throws Exception {
+        int N = frames.length, H = frames[0].length, W = frames[0][0].length;
+        FloatBuffer fb = FloatBuffer.allocate(N * H * W);
+        for (int n = 0; n < N; n++)
+            for (int y = 0; y < H; y++)
+                fb.put(frames[n][y]);
+        fb.rewind();
+        try (OnnxTensor in = OnnxTensor.createTensor(env, fb, new long[] {1, N, H, W});
+             OrtSession.Result res = session.run(Collections.singletonMap(inputName, in))) {
+            float[][][][] out = (float[][][][]) res.get(0).getValue();   // [1, outCh, 1, 1]
+            int outCh = out[0].length;
+            float[] flat = new float[outCh];
+            for (int c = 0; c < outCh; c++) flat[c] = out[0][c][0][0];
+            return flat;
+        }
+    }
+
+    public static float sigmoid(float x) { return (float) (1.0 / (1.0 + Math.exp(-x))); }
+
+    @Override public void close() throws Exception { session.close(); }
+
+    // ---- model location resolver (mirrors the tile_processor_gpu default/override scheme) ----
+    static String resolveModel(String spec) throws Exception {
+        if (spec == null || spec.trim().isEmpty())
+            throw new IllegalArgumentException("CuasDnnInfer: empty model spec (curt_dnn_model)");
+        spec = spec.trim();
+        if (new File(spec).isFile()) return spec;                    // local path
+        String cacheDir = System.getProperty("user.home") + "/.cache/c5p_dnn/";
+        new File(cacheDir).mkdirs();
+        String name = spec.substring(spec.lastIndexOf('/') + 1);
+        String localOnnx = cacheDir + name;
+        if (spec.startsWith("http://") || spec.startsWith("https://")) {
+            httpFetch(spec, localOnnx);
+            try { httpFetch(spec + ".data", localOnnx + ".data"); } catch (Exception e) { /* no external data */ }
+        } else if (spec.matches("[^@/]+@[^:]+:.*")) {                 // user@host:/path (scp/rsync)
+            rsyncFetch(spec, localOnnx);
+            try { rsyncFetch(spec + ".data", localOnnx + ".data"); } catch (Exception e) { /* none */ }
+        } else {
+            throw new java.io.FileNotFoundException("CuasDnnInfer: model not found / unrecognized spec: " + spec);
+        }
+        return localOnnx;
+    }
+
+    static void rsyncFetch(String remote, String dest) throws Exception {
+        Process p = new ProcessBuilder("rsync", "-az", remote, dest).inheritIO().start();
+        if (p.waitFor() != 0) throw new RuntimeException("rsync failed for " + remote);
+    }
+
+    static void httpFetch(String url, String dest) throws Exception {
+        try (InputStream is = new URL(url).openStream()) {
+            Files.copy(is, new File(dest).toPath(), java.nio.file.StandardCopyOption.REPLACE_EXISTING);
+        }
+    }
+
+    // ---- standalone verification: Java ORT output vs the saved PyTorch raw output ----
+    static float[] readFloatsLE(String path) throws Exception {
+        ByteArrayOutputStream bos = new ByteArrayOutputStream();
+        try (DataInputStream in = new DataInputStream(new FileInputStream(path))) {
+            byte[] buf = new byte[8192]; int n;
+            while ((n = in.read(buf)) > 0) bos.write(buf, 0, n);
+        }
+        ByteBuffer bb = ByteBuffer.wrap(bos.toByteArray()).order(ByteOrder.LITTLE_ENDIAN);
+        float[] f = new float[bb.remaining() / 4];
+        for (int i = 0; i < f.length; i++) f[i] = bb.getFloat();
+        return f;
+    }
+
+    public static void main(String[] args) throws Exception {
+        // args: model(path/spec) testvec_in.bin [testvec_out.bin] [N H W]
+        String model = args[0], invec = args[1];
+        String outvec = args.length > 2 ? args[2] : null;
+        int N = args.length > 3 ? Integer.parseInt(args[3]) : 8;
+        int H = args.length > 4 ? Integer.parseInt(args[4]) : 24;
+        int W = args.length > 5 ? Integer.parseInt(args[5]) : 24;
+        float[] flat = readFloatsLE(invec);
+        if (flat.length != N * H * W) throw new RuntimeException("in size " + flat.length + " != " + (N * H * W));
+        float[][][] frames = new float[N][H][W];
+        int k = 0;
+        for (int n = 0; n < N; n++) for (int y = 0; y < H; y++) for (int x = 0; x < W; x++) frames[n][y][x] = flat[k++];
+        try (CuasDnnInfer infer = new CuasDnnInfer(model, N)) {
+            float[] out = infer.inferRaw(frames);
+            System.out.println("output len=" + out.length + "  det_logit=" + out[0] + "  s=" + sigmoid(out[0]));
+            if (outvec != null) {
+                float[] exp = readFloatsLE(outvec);
+                double maxabs = 0; int amax = 0;
+                for (int i = 0; i < Math.min(out.length, exp.length); i++) {
+                    double d = Math.abs(out[i] - exp[i]); if (d > maxabs) { maxabs = d; amax = i; }
+                }
+                System.out.printf("max abs diff vs PyTorch = %.6g (at ch %d: java %.5f vs torch %.5f)%n",
+                        maxabs, amax, out[amax], exp[amax]);
+                System.out.println(maxabs < 1e-3 ? "PASS - Java ORT matches PyTorch" : "MISMATCH - investigate");
+            }
+        }
+    }
+}

src/main/java/com/elphel/imagej/tileprocessor/IntersceneMatchParameters.java

@@ -1177,6 +1177,7 @@ min_str_neib_fpn	0.35
 	public boolean  curt_c5_white =        false; // linear 4D unsharp/whitening of the conv output before all nonlinearities: sharpens (reduces moment of inertia / Gram spread) at the cost of noise gain; filename gets -W<sp>_<vel> // By Claude on 06/13/2026
 	public double   curt_c5_white_sp =     0.5;   // whitening spatial strength a (3-tap unsharp (1+2a,-a,-a) in x and y); 0 - no spatial sharpening // By Claude on 06/13/2026
 	public double   curt_c5_white_vel =    0.0;   // whitening velocity strength a (3-tap unsharp in vx and vy); 0 - no velocity sharpening // By Claude on 06/13/2026
+	public String   curt_dnn_model =       "";    // C5P DNN front-end model (ONNX): empty = disabled; local path, scp user@host:path, or http(s) URL fetched to cache; overrides bundled resource (mirrors tile_processor_gpu) // By Claude on 06/13/2026
 	public boolean  curt_synth_src =       true;  // default set for the synthetic B-measurement experiment (set false for real-data runs); reads *-CUAS-SYNTHETIC-CUAS.tiff, output titles get -SYNTH // By Claude on 06/12/2026
 	public double   curt_synth_scale =     5.0;   // synthetic target peak, counts (synthetic file is peak-1 normalized; scaled at load) // By Claude on 06/12/2026
 	public boolean  curt_synth_bg =        true;  // add the real *-CUAS-MERGED-CUAS.tiff scene under the synthetic targets (label-matched frames); false - clean targets only // By Claude on 06/12/2026
@@ -3516,6 +3517,8 @@ min_str_neib_fpn	0.35
 				"Spatial unsharp strength a: s' = (1+2a)s - a(s_left+s_right) in x and y across the ROI. Larger = sharper spatial blob + more spatial noise. 0 - off."); // By Claude on 06/13/2026
 		gd.addNumericField("Whitening velocity strength",            this.curt_c5_white_vel, 6,8,"", // By Claude on 06/13/2026
 				"Velocity unsharp strength a: same 3-tap in vx and vy within each pixel's velocity grid. Sharpens the velocity-cell spread. 0 - off (start spatial-only)."); // By Claude on 06/13/2026
+		gd.addStringField ("C5P DNN model (ONNX)",                   this.curt_dnn_model, 60, // By Claude on 06/13/2026
+				"Trained DNN front-end model location. Empty = disabled (use matched-filter/posterior path). Local path, scp user@host:/path (fetched to cache), or http(s) URL. Overrides any bundled resource - same default-vs-override scheme as the GPU kernel sources."); // By Claude on 06/13/2026
 		gd.addCheckbox ("Use synthetic input",                       this.curt_synth_src, // By Claude on 06/11/2026
 				"Read *-CUAS-SYNTHETIC-CUAS.tiff (generated test targets) instead of *-CUAS-MERGED-CUAS.tiff from the same model directory; all output titles get a -SYNTH mark."); // By Claude on 06/11/2026
 		gd.addNumericField("Synthetic target peak",                   this.curt_synth_scale, 6,8,"counts", // By Claude on 06/12/2026
@@ -5060,6 +5063,7 @@ min_str_neib_fpn	0.35
 		this.curt_c5_white =            gd.getNextBoolean();     // By Claude on 06/13/2026
 		this.curt_c5_white_sp =         gd.getNextNumber();      // By Claude on 06/13/2026
 		this.curt_c5_white_vel =        gd.getNextNumber();      // By Claude on 06/13/2026
+		this.curt_dnn_model =           gd.getNextString().trim(); // By Claude on 06/13/2026
 		this.curt_synth_src =           gd.getNextBoolean();     // By Claude on 06/11/2026
 		this.curt_synth_scale =         gd.getNextNumber();      // By Claude on 06/12/2026
 		this.curt_synth_bg =            gd.getNextBoolean();     // By Claude on 06/12/2026
@@ -6424,6 +6428,7 @@ min_str_neib_fpn	0.35
 		properties.setProperty(prefix+"curt_c5_white",        this.curt_c5_white+"");       // boolean // By Claude on 06/13/2026
 		properties.setProperty(prefix+"curt_c5_white_sp",     this.curt_c5_white_sp+"");    // double // By Claude on 06/13/2026
 		properties.setProperty(prefix+"curt_c5_white_vel",    this.curt_c5_white_vel+"");   // double // By Claude on 06/13/2026
+		properties.setProperty(prefix+"curt_dnn_model",       this.curt_dnn_model);         // String // By Claude on 06/13/2026
 		properties.setProperty(prefix+"curt_synth_src",       this.curt_synth_src+"");      // boolean // By Claude on 06/11/2026
 		properties.setProperty(prefix+"curt_synth_scale",     this.curt_synth_scale+"");    // double  // By Claude on 06/12/2026
 		properties.setProperty(prefix+"curt_synth_bg",        this.curt_synth_bg+"");       // boolean // By Claude on 06/12/2026
@@ -6823,6 +6828,7 @@ min_str_neib_fpn	0.35
 		if (properties.getProperty(prefix+"curt_c5_white")!=null)        this.curt_c5_white=Boolean.parseBoolean(properties.getProperty(prefix+"curt_c5_white")); // By Claude on 06/13/2026
 		if (properties.getProperty(prefix+"curt_c5_white_sp")!=null)     this.curt_c5_white_sp=Double.parseDouble(properties.getProperty(prefix+"curt_c5_white_sp")); // By Claude on 06/13/2026
 		if (properties.getProperty(prefix+"curt_c5_white_vel")!=null)    this.curt_c5_white_vel=Double.parseDouble(properties.getProperty(prefix+"curt_c5_white_vel")); // By Claude on 06/13/2026
+		if (properties.getProperty(prefix+"curt_dnn_model")!=null)       this.curt_dnn_model=(String) properties.getProperty(prefix+"curt_dnn_model"); // By Claude on 06/13/2026
 
 		if (properties.getProperty(prefix+"curt_synth_src")!=null)       this.curt_synth_src=Boolean.parseBoolean(properties.getProperty(prefix+"curt_synth_src"));         // By Claude on 06/11/2026
 		if (properties.getProperty(prefix+"curt_synth_scale")!=null)     this.curt_synth_scale=Double.parseDouble(properties.getProperty(prefix+"curt_synth_scale"));       // By Claude on 06/12/2026
@@ -9104,6 +9110,7 @@ min_str_neib_fpn	0.35
 		imp.curt_c5_white =         this.curt_c5_white;       // By Claude on 06/13/2026
 		imp.curt_c5_white_sp =      this.curt_c5_white_sp;    // By Claude on 06/13/2026
 		imp.curt_c5_white_vel =     this.curt_c5_white_vel;   // By Claude on 06/13/2026
+		imp.curt_dnn_model =        this.curt_dnn_model;     // By Claude on 06/13/2026
 		imp.curt_synth_src =        this.curt_synth_src;     // By Claude on 06/11/2026
 		imp.curt_synth_scale =      this.curt_synth_scale;   // By Claude on 06/12/2026
 		imp.curt_synth_bg =         this.curt_synth_bg;      // By Claude on 06/12/2026