Add a script to convert a jpg to the correct input for resnet50 with the vulkan gui (#362)

cpp/README.md

@@ -42,6 +42,16 @@ cmake --build build/
 wget https://storage.googleapis.com/shark_tank/latest/resnet50_tf/resnet50_tf.mlir
 iree-compile --iree-input-type=mhlo --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --iree-llvm-embedded-linker-path=`python3 -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])'`/iree/compiler/tools/../_mlir_libs/iree-lld --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --mlir-pass-pipeline-crash-reproducer=ist/core-reproducer.mlir --iree-llvm-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 resnet50_tf.mlir -o resnet50_tf.vmfb
 ```
+*Prepare the input*
+
+```bash
+python save_img.py
+```
+Note that this requires tensorflow, e.g.
+```bash
+python -m pip install tensorflow
+```
+
 *Run the vulkan_gui*
 ```bash
 ./build/vulkan_gui/iree-samples-vulkan-gui

cpp/save_img.py

@@ -0,0 +1,19 @@
+import numpy as np
+import tensorflow as tf
+from shark.shark_inference import SharkInference
+from shark.shark_downloader import download_tf_model
+
+
+def load_and_preprocess_image(fname: str):
+    image = tf.io.read_file(fname)
+    image = tf.image.decode_image(image, channels=3)
+    image = tf.image.resize(image, (224, 224))
+    image = image[tf.newaxis, :]
+    # preprocessing pipeline
+    input_tensor = tf.keras.applications.resnet50.preprocess_input(image)
+    return input_tensor
+
+
+data = load_and_preprocess_image("dog_imagenet.jpg").numpy()
+
+data.tofile("dog.bin")

cpp/vulkan_gui/vulkan_inference_gui.cc

@@ -17,6 +17,7 @@
 #include <cstring>
 #include <set>
 #include <vector>
+#include <fstream>
 
 #include "iree/hal/drivers/vulkan/api.h"
 
@@ -915,14 +916,20 @@ extern "C" int iree_main(int argc, char** argv) {
   char filename[] = "dog_imagenet.jpg";
   fprintf(stdout, "loading: %s\n", filename);
   int x,y,n;
-  unsigned char *image_raw = stbi_load(filename, &x, &y, &n, 3);
+  //unsigned char *image_raw = stbi_load(filename, &x, &y, &n, 3);
+  stbi_load(filename, &x, &y, &n, 3);
   fprintf(stdout, "res: %i x %i x %i\n", x, y, n);
 
+  /* Preprocessing needs to go here. For now use a buffer preprocessed in python.
+
   //convert image into floating point format
   for(int i=0;i<224*224*3;i++)
   {
     input_res50[i]= ((float)image_raw[i])/255.0f;
-  }
+  }*/
+
+  std::ifstream fin("dog.bin", std::ifstream::in | std::ifstream::binary);
+  fin.read((char*)input_res50, 224*224*3*sizeof(float));
 
   // load image again so imgui can display it
   int my_image_width = 0;
@@ -1060,7 +1067,7 @@ extern "C" int iree_main(int argc, char** argv) {
 
       // ImGui Inputs for two input tensors.
       // Run computation whenever any of the values changes.
-      static bool dirty = false;
+      static bool dirty = true;
       if (dirty) {
 
         // Synchronously invoke the function.