finishing tidying up example code

Author: stefanwebb

bootcamp/tutorials/quickstart/onnx_example/java/app/src/main/java/org/example/App.java

@@ -7,62 +7,68 @@
 import java.util.Map;
 
 import ai.onnxruntime.*;
-// import ai.onnxruntime.extensions.*;
 import ai.onnxruntime.extensions.*;
 
 @SuppressWarnings("unused")
 public class App {
-    public String getGreeting() {
-        return "Hello World 2!";
-    }
-
     public static void main(String[] args) {
         var env = OrtEnvironment.getEnvironment();
 
         try {
             var sess_opt = new OrtSession.SessionOptions();
+
+            // NOTE: ONNXRuntimeExtensions for Java on Apple Silicon isn't currently
+            // available
+            // Hence I'll comment out these lines for my machine
             // sess_opt.registerCustomOpLibrary(OrtxPackage.getLibraryPath());
+
+            // Depending on how you call App within Visual Studio, may need to add app/ to
+            // filenames below
             System.out.println(System.getProperty("user.dir"));
 
-            // var tokenizer = env.createSession("app/tokenizer.onnx", new
-            // OrtSession.SessionOptions());
+            // Try out tokenizer
+            // var tokenizer = env.createSession("tokenizer.onnx", sess_opt);
 
+            // Get input and output node names for tokenizer
             // var inputName = tokenizer.getInputNames().iterator().next();
             // var outputName = tokenizer.getOutputNames().iterator().next();
-
             // System.out.println(inputName);
             // System.out.println(outputName);
 
-            var session = env.createSession("app/model.onnx", sess_opt);
+            // Try out embedding model
+            var session = env.createSession("model.onnx", sess_opt);
 
             // Get input and output names
-
             var inputName = session.getInputNames().iterator().next();
             var outputName = session.getOutputNames().iterator().next();
+            System.out.println(inputName);
+            System.out.println(outputName);
 
-            // System.out.println(inputName);
-            // System.out.println(outputName);
-
+            // Since I wasn't able to run tokenizer via ONNX on Apple Silicon, hardcode
+            // token ids from Python
+            // This is for "The quick brown fox..."
             long[][] tokens = { { 101, 1996, 4248, 2829, 4419, 14523, 2058, 1996, 13971, 3899, 1012, 102 } };
             long[][] masks = { { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 } };
             long[][] token_type_ids = { { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 } };
 
+            // Wrap native Java types in OnnxTensor's and put in input map
             var test_tokens = OnnxTensor.createTensor(env, tokens);
             var test_mask = OnnxTensor.createTensor(env, masks);
             var test_token_type_ids = OnnxTensor.createTensor(env, token_type_ids);
-
             var inputs = Map.of("input_ids", test_tokens, "attention_mask", test_mask, "token_type_ids",
                     test_token_type_ids);
-            var results = session.run(inputs).get("embeddings");
 
-            // System.out.println("type");
-            // System.out.println(results.get().getType());
-
-            // float[][][] embeddings = (float[][][]) results.get().getValue();
+            // Run embedding model on tokens and convert back to native Java type
+            var results = session.run(inputs).get("embeddings");
             float[][] embeddings = (float[][]) results.get().getValue();
 
+            // Print the first 16 dimensions of the resulting embedding
             var result = Arrays.toString(Arrays.copyOfRange(embeddings[0], 0, 16));
             System.out.println(result);
+
+            // Comparing this to our Python notebook, we see it is identical! I.e. calling
+            // the ONNXRuntime on the same model from Python and Java produces same results
+            // (up to numerical precision etc.)
         } catch (Exception e) {
             System.out.println(e);
         }

bootcamp/tutorials/quickstart/onnx_example/python/export_hf_to_onnx.ipynb

@@ -688,7 +688,9 @@
    "id": "987319b9",
    "metadata": {},
    "source": [
-    "The embeddings between the ONNX model and the saved HuggingFace weights will differ as they're from different trainings."
+    "The embeddings between the ONNX model and the saved HuggingFace weights will differ as they're from different trainings.\n",
+    "\n",
+    "In the corresponding Java app, you can compare the embedding below to the one produced via Java."
    ]
   },
   {