NPU usage for AI models

[Mohan51]

Hi, I am trying to utilize NPU for my currency model..for that i converted my tensorflow model to quantized tflite model (using EIQ). but still it is using Cores.I tried with the nxp example and it run successfully, i am attaching the screenshot for your reference.

While running my inference code, the logs it showing external delegate is created for NXP example. but it is not showing the message 'external delegate is created ' for my custom model.

In the last line it is shown that tensorflow lite delegate xnnpack is created for cpu. but why it is creating for cpu?

this is my inference code:

from fastapi import FastAPI, File, UploadFile, HTTPException

from fastapi.responses import HTMLResponse, JSONResponse

from fastapi.staticfiles import StaticFiles

import tflite_runtime.interpreter as tflite

import cv2

import numpy as np

import os

USE_HARDWARE_INFERENCE = True

if os.environ.get('USE_HARDWARE_INFERENCE') == '0':

USE_HARWARE_INFERENCE = False

if (USE_HARDWARE_INFERENCE):

delegates = [tflite.load_delegate('/usr/lib/libethosu_delegate.so')]

else:

delegates = []

app = FastAPI()

# Load class names and TensorFlow Lite Interpreter for currency classification

currency_model_path = "/home/root/currency/Xception_saved_converted.tflite" # Update with the actual path to your classification model

currency_class_names_path = "labels.txt" # Update with the actual path to your class names file

with open(currency_class_names_path, "r") as file:

currency_class_names = [line.strip() for line in file.readlines()]

currency_interpreter = tflite.Interpreter(model_path=currency_model_path,experimental_delegates = delegates)

currency_interpreter.allocate_tensors()

# Serve static files

app.mount("/static", StaticFiles(directory="static"), name="static")

@app.get("/", response_class=HTMLResponse)

async def get_index():

with open("static/index.html", "r") as file:

html_content = file.read()

return HTMLResponse(content=html_content)

def classify_currency(image

input_details = currency_interpreter.get_input_details()

output_details = currency_interpreter.get_output_details()

input_shape = input_details[0]['shape']

img = cv2.resize(image, (input_shape[1], input_shape[2]))

img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

input_data = np.expand_dims(img_rgb, axis=0).astype(np.float32)

currency_interpreter.set_tensor(input_details[0]['index'], input_data)

currency_interpreter.invoke()

output_data = currency_interpreter.get_tensor(output_details[0]['index'])

class_index = np.argmax(output_data)

class_name = currency_class_names[class_index]

return class_name

@app.post("/capture-and-detect/")

async def capture_and_detect(image_file: UploadFile = File(...)):

contents = await image_file.read()

nparr = np.frombuffer(contents, np.uint8)

img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)

if img is None:

raise HTTPException(status_code=400, detail="Invalid image file")

# Classify the image

classification_result = classify_currency(img)

result = {

"classification": classification_result

}

return JSONResponse(content=result)

if __name__ == "__main__":

import uvicorn

import os

# Ensure the static directory exists

if not os.path.exists("static"

os.makedirs("static")

uvicorn.run(app, host="0.0.0.0", port=81)

[Bio_TICFSL]

Hello,

Which board are you using? Which BSP? this run normal with NPU but you are using FASTAPI this does not include support for NPU that's the reason of it.

Regards