example for EP Context Cache

example/EP_Context_Cache/Makefile

@@ -0,0 +1,23 @@
+FLEXML_XTC ?= /proj/testcases/xtc/HEAD/tc/open/flexml
+MODEL_SOURCE_PATH := $(abspath $(dir $(lastword $(MAKEFILE_LIST))))
+
+ifeq ($(shell test -d $(MODEL_SOURCE_PATH)/../../../common && echo -n yes),yes)
+	PERFORCE_HOME = $(MODEL_SOURCE_PATH)/../../..
+else
+	PERFORCE_HOME = $(FLEXML_XTC)
+endif
+
+ifeq ($(shell test -d $(MODEL_SOURCE_PATH)/../../common && echo -n yes),yes)
+	PERFORCE_E2E_HOME = $(MODEL_SOURCE_PATH)/../../
+else
+	PERFORCE_E2E_HOME = $(FLEXML_XTC)/e2e-hw
+endif
+
+MODEL = ResNet18
+INPUT_MODEL = ResNet18.onnx
+TENSOR_SHAPE = 1x3x224x224
+VAIMLEP_RUN_SCRIPT = run.py
+DEVICE = stx
+FLEXML_USE_STD_SYSTEM_CALL = 1
+
+include $(PERFORCE_E2E_HOME)/common/Makefile-common.mk

example/EP_Context_Cache/README.md

@@ -0,0 +1,74 @@
+<table class="sphinxhide" width="100%">
+ <tr width="100%">
+    <td align="center"><img src="https://raw.githubusercontent.com/Xilinx/Image-Collateral/main/xilinx-logo.png" width="30%"/><h1> Ryzen™ AI EP Context Cache Example </h1>
+    </td>
+ </tr>
+</table>
+
+# Getting started with Ryzen AI EP Context Cache
+
+This is an example showing how to compile and run the ResNet18 model from https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet18-v2-7.onnx
+ on AMD's Ryzen AI NPU by new convenient EP Context Cache with ease of usage (start from Ryzen AI 1.5). 
+
+
+
+# Activate Ryzen AI conda environment
+
+
+```bash
+#Install Ryzen AI msi with relative NPU driver
+conda activate ryzen-ai-1.x
+```
+
+# Generate EP Context Cache directly, no need to quantize firstly
+
+```bash
+# https://github.com/onnx/models/blob/main/validated/vision/classification/resnet/model/resnet18-v2-7.onnx
+
+python compile.py resnet18-v2-7.onnx
+
+WARNING: Logging before InitGoogleLogging() is written to STDERR
+I20250729 08:31:30.094357 12396 vitisai_compile_model.cpp:1157] Vitis AI EP Load ONNX Model Success
+I20250729 08:31:30.094357 12396 vitisai_compile_model.cpp:1158] Graph Input Node Name/Shape (1)
+I20250729 08:31:30.094357 12396 vitisai_compile_model.cpp:1162]          data : [-1x3x224x224]
+I20250729 08:31:30.095352 12396 vitisai_compile_model.cpp:1168] Graph Output Node Name/Shape (1)
+I20250729 08:31:30.095352 12396 vitisai_compile_model.cpp:1172]          resnetv22_dense0_fwd : [-1x1000]
+Adding RYZEN_AI_INSTALLATION_PATH=C:\Program Files\RyzenAI\... to installation search path
+
+subpartition path = ....\resnet18-ep-context\resnet18-v2-7\vaiml_par_0\0
+[Vitis AI EP] No. of Operators : VAIML    60
+[Vitis AI EP] No. of Subgraphs : VAIML     1
+
+# EP context cache model is saved as resnet18-v2-7_ctx.onnx
+
+```
+
+# Run Inference with EP Context Cache
+```bash
+
+python run.py resnet18-v2-7_ctx.onnx
+WARNING: Logging before InitGoogleLogging() is written to STDERR
+I20250729 08:39:23.587743  2180 vitisai_compile_model.cpp:1157] Vitis AI EP Load ONNX Model Success
+I20250729 08:39:23.587743  2180 vitisai_compile_model.cpp:1158] Graph Input Node Name/Shape (1)
+I20250729 08:39:23.587743  2180 vitisai_compile_model.cpp:1162]          data : [-1x3x224x224]
+I20250729 08:39:23.587743  2180 vitisai_compile_model.cpp:1168] Graph Output Node Name/Shape (1)
+I20250729 08:39:23.587743  2180 vitisai_compile_model.cpp:1172]          resnetv22_dense0_fwd : [-1x1000]
+[Vitis AI EP] No. of Subgraphs supported by Vitis AI EP: VAIML     1
+Top 3 Probabilities
+[208 207 176]
+------------------------------------|------------
+Classification                      |Percentage
+------------------------------------|------------
+Labrador retriever                  |   67.43
+------------------------------------|------------
+golden retriever                    |    9.13
+------------------------------------|------------
+Saluki, gazelle hound               |    8.05
+------------------------------------|------------
+INFO: Test passed
+
+```
+
+
+
+

example/EP_Context_Cache/compile.py

@@ -0,0 +1,25 @@
+import os
+import sys
+import json
+import onnxruntime as ort
+
+onnx_model_path  = sys.argv[1]
+model_name = os.path.splitext(os.path.basename(onnx_model_path))[0]
+ctx_cache = model_name + "_ctx.onnx"
+
+# Delete prexisting EP context cache model
+if os.path.exists(ctx_cache):
+    print(f"INFO: EP context model {ctx_cache} already exists. Deleting it.")
+    os.remove(ctx_cache)
+
+session_options = ort.SessionOptions()
+session_options.add_session_config_entry('ep.context_enable', '1') 
+session_options.add_session_config_entry('ep.context_file_path', ctx_cache) 
+session_options.add_session_config_entry('ep.context_embed_mode', '1') 
+onnx_session = ort.InferenceSession(
+                   onnx_model_path,
+                   sess_options=session_options,
+                   providers=["VitisAIExecutionProvider"],
+                   provider_options=[{"cache_dir": os.getcwd(),
+                                      "cache_key": model_name,}]
+               )

example/EP_Context_Cache/image_utils.py

@@ -0,0 +1,79 @@
+from typing import Any, List, Tuple
+
+import numpy as np
+import numpy.typing as npt
+import torch
+from PIL import Image  # type: ignore [import-untyped]
+from torchvision import transforms
+
+
+def load_and_preprocess_image(image_file: str) -> torch.Tensor:
+    """
+    Load and preprocess image_file for inference test
+    It works for all imagenet images
+    """
+
+    img = Image.open(image_file).convert("RGB")
+    # preprocessing pipeline
+    preprocess = transforms.Compose(
+        [
+            transforms.Resize(256),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ]
+    )
+    img_preprocessed = preprocess(img)
+    return torch.unsqueeze(img_preprocessed, 0)
+
+
+def softmax(vector: npt.NDArray[np.float32]) -> npt.NDArray[np.float32]:
+    """
+    Calculate softmax of a vector
+    """
+    e = np.exp(vector)
+    res: npt.NDArray[np.float32] = e / e.sum()
+    return res
+
+
+def top_n_probabilities(
+    res: npt.NDArray[np.float32],
+    labels: List[str],
+    top_n: int = 3,
+    run_softmax: bool = False,
+) -> List[Tuple[Any, Any]]:
+    """
+    Compute probabilities of top 3 classifications from res
+    Inputs:
+        data_in: output as 1-D numpy array from full connected layer or softmax
+        run_softmax: whether or not to run softmax on data_in
+    """
+    indices = np.flip(np.argsort(res))
+    if run_softmax:
+        percentage = softmax(res) * 100
+    else:
+        percentage = res * 100
+
+    print(indices[:3])
+    top_n_result = [(labels[idx], percentage[idx].item()) for idx in indices[:3]]
+
+    return top_n_result
+
+
+def top3_probabilities(
+    data_in: npt.NDArray[np.float32], labels: List[str], run_softmax: bool = False
+) -> List[Tuple[Any, Any]]:
+    """
+    Helper function to get top 3 probabilities for backward compatibility
+    """
+    top3 = top_n_probabilities(data_in, labels, top_n=3, run_softmax=run_softmax)
+
+    return top3
+
+
+def load_labels(label_file: str) -> List[str]:
+    classes_fh = open(label_file)
+    labels = [line.strip() for line in classes_fh]
+
+    classes_fh.close()
+    return labels