[Draft][ONNX] Enable Android Build Support for ONNX Model with Protobuf Integration

.gitignore

@@ -4,6 +4,9 @@
 # ignore build directory
 /build
 /builddir
+/third_party
+/protobuf-25.2
+/subprojects
 .cache/
 .idea/
 
@@ -71,3 +74,8 @@ test/jni/googletest/
 ctre-unicode.hpp
 encoder.hpp
 json.hpp
+
+*.bin
+onnx__*
+*.onnx
+*.weight

Applications/ONNX/jni/Android.mk

@@ -44,12 +44,12 @@ LOCAL_ARM_NEON := true
 LOCAL_CFLAGS += -std=c++17 -Ofast -mcpu=cortex-a53 -Ilz4-nougat/lib
 LOCAL_LDFLAGS += -Llz4-nougat/lib/obj/local/$(TARGET_ARCH_ABI)/
 LOCAL_CXXFLAGS += -std=c++17 -frtti
-LOCAL_CFLAGS += -pthread -fexceptions -fopenmp
+LOCAL_CFLAGS += -pthread -fexceptions -fopenmp -static-openmp
 LOCAL_LDFLAGS += -fexceptions
 LOCAL_MODULE_TAGS := optional
 LOCAL_ARM_MODE := arm
 LOCAL_MODULE := nntrainer_onnx_example
-LOCAL_LDLIBS := -llog -landroid -fopenmp
+LOCAL_LDLIBS := -llog -landroid -fopenmp -static-openmp
 
 LOCAL_SRC_FILES := main.cpp
 

Applications/ONNX/jni/main.cpp

@@ -1,48 +1,125 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Copyright (C) 2025 Seungbaek Hong <[email protected]>
- *
+ * Copyright (C) 2025 Sachin Singh <[email protected]>
  * @file   main.cpp
- * @date   26 Feb 2025
+ * @date   14 October 2025
  * @brief  onnx example using nntrainer-onnx-api
  * @see    https://github.com/nnstreamer/nntrainer
- * @author Seungbaek Hong <sb92.honge@samsung.com>
+ * @author Sachin Singh <sachin.3@samsung.com>
  * @bug    No known bugs except for NYI items
+ *
+ * Updated on 16 Oct 2025 to add debug output for Android execution
  */
 
+#include <fstream>
 #include <iostream>
 #include <layer.h>
 #include <model.h>
 #include <nntrainer-api-common.h>
 #include <optimizer.h>
 #include <util_func.h>
 
+void saveToRaw(const float *data, size_t size, const std::string &filename) {
+  std::ofstream out(filename, std::ios::binary);
+  if (!out) {
+    std::cerr << "Error: Cannot open file " << filename << " for writing.\n";
+    return;
+  }
+
+  out.write(reinterpret_cast<const char *>(data), size * sizeof(float));
+  out.close();
+
+  std::cout << std::endl << ".bin generated successfully !";
+}
+
 int main() {
   auto model = ml::train::createModel();
 
+  std::cout << "--------------------------------------Create Model "
+               "Done--------------------------------------"
+            << std::endl;
   try {
-    std::string path = "../../../../Applications/ONNX/jni/add_example.onnx";
+    // std::string path = "/storage_data/snap/sumon/sumon-98/nntrainer/Applications/ONNX/jni/qwen3_model_one_layer_no_cast.onnx";
+    std::string path = "./qwen3_model_one_layer_no_cast.onnx";
     model->load(path, ml::train::ModelFormat::MODEL_FORMAT_ONNX);
   } catch (const std::exception &e) {
     std::cerr << "Error during load: " << e.what() << "\n";
     return 1;
   }
 
+  std::cout << "--------------------------------------Load Model "
+               "Done--------------------------------------"
+            << std::endl;
   try {
-    model->compile();
+    model->compile(ml::train::ExecutionMode::INFERENCE);
   } catch (const std::exception &e) {
     std::cerr << "Error during compile: " << e.what() << "\n";
     return 1;
   }
 
+  std::cout << "--------------------------------------Compile Model "
+               "Done--------------------------------------"
+            << std::endl;
   try {
     model->initialize();
   } catch (const std::exception &e) {
     std::cerr << "Error during initialize: " << e.what() << "\n";
     return 1;
   }
 
+  std::cout << "--------------------------------------Initialize Model Done--------------------------------------" << std::endl;
+  std::cout << "Skipping model summary..." << std::endl;
+
+  std::cout << "Starting model summary..." << std::endl;
   model->summarize(std::cout, ML_TRAIN_SUMMARY_MODEL);
+  std::cout << "Finished model summary." << std::endl;
+
+  std::cout << "--------------------------------------Summarize Model "
+               "Done--------------------------------------"
+            << std::endl;
+
+  std::cout << "Loading weights..." << std::endl;
+  std::string weight_path = "./qwen_weights_one_layer_no_cast/";
+  std::cout << "Loading weights from: " << weight_path << std::endl;
+  std::cout << "Loading weights from: " << weight_path << std::endl;
+  try {
+        model->load(weight_path, ml::train::ModelFormat::MODEL_FORMAT_BIN);
+        std::cout << "Weights loaded successfully" << std::endl;
+      } catch (std::exception &e) {
+        std::cerr << "Error during loading weights: " << e.what() << "\n";
+        return 1;
+      }
+  std::cout<<"starting inferencing!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!";
+  float *input = new float[1];
+  float *sin = new float[128];
+  float *cos = new float[128];
+  float *epsilon = new float[1];
+
+  input[0] = 52;
+
+  for (int i = 0; i < 128; i++) {
+    sin[i] = 0;
+    cos[i] = 1;
+  }
+  epsilon[0] = 1e-6;
+
+  std::vector<float *> in;
+
+  in.push_back(epsilon);
+  in.push_back(sin);
+  in.push_back(cos);
+  in.push_back(input);
+
+  auto ans = model->inference(1, in);
+
+  std::cout << "-------------------------------------------Inference "
+               "Done--------------------------------------------"
+            << std::endl;
+
+  for (auto it : ans) {
+    saveToRaw(it, 151936,
+              "../../../../Applications/ONNX/jni/nntrainer_logits.bin");
+  }
 
   return 0;
 }

Applications/ONNX/python/qwen3/compare.py

@@ -0,0 +1,9 @@
+import numpy as np
+
+arr1 = np.fromfile("./modelling_logits.bin",dtype="float32").reshape(1,151936)
+arr2 = np.fromfile("../../jni/nntrainer_logits.bin",dtype="float32").reshape(1,151936)
+
+if(np.allclose(arr1,arr2,atol=1e-4,rtol=1e-4)):
+    print("equal")
+else:
+    print("not equal")    

Applications/ONNX/python/qwen3/create_bin.py

@@ -0,0 +1,47 @@
+import onnx
+import numpy as np
+import json
+import os
+import shutil
+from onnx import numpy_helper, TensorProto
+
+def cleanName(name):
+    if name.startswith('/'):
+        name = name[1:]
+
+    name = name.replace('/', '_')
+    name = name.replace('.', '_')
+    name = name.replace(':', '_')
+    name = name.lower()
+
+    return name
+
+
+model = onnx.load("./qwen3_model.onnx", load_external_data=True)
+
+metadata = {}
+
+script_dir = os.path.dirname(os.path.abspath(__file__))
+folder = os.path.join(script_dir, "bins")
+if os.path.exists(folder):
+    shutil.rmtree(folder) 
+os.makedirs(folder)
+
+for tensor in model.graph.initializer:
+    arr = numpy_helper.to_array(tensor).astype(np.float32)
+
+    filename = f"./bins/{cleanName(tensor.name)}.bin"
+    arr.tofile(filename)
+
+    # Save metadata (name, dtype, shape, file)
+    metadata[tensor.name] = {
+        "file": filename,
+        "tensor name": tensor.name,
+        "dtype": TensorProto.DataType.Name(tensor.data_type),
+        "shape": list(arr.shape)
+    }
+
+    print(f"Saved {tensor.name} -> {filename}, dtype={arr.dtype}, shape={arr.shape}")
+
+with open("./weights_metadata.json", "w") as f:
+    json.dump(metadata, f, indent=4)