运行 LiteRT / TFLite 模型 - Qualcomm Dragonwing Documentation

LiteRT(前称 TensorFlow Lite)是 Google 面向设备端 AI 的高性能运行时。借助 AI Engine Direct 中包含的 LiteRT 委托(delegate),您只需一行代码,即可在 Dragonwing 设备的 NPU 上运行现有的量化 LiteRT 模型(Python 或 C++)。

.tflite 文件从哪里来?

从 TensorFlow/Keras 转换而来 — 参见转换 TensorFlow 模型
从 Qualcomm AI Hub 下载预量化模型 — 参见 AI Hub
从 Edge Impulse 导出 — 参见 Edge Impulse

量化模型

NPU 仅支持 uint8/int8 量化模型。不受支持的模型或不受支持的层会自动回退到 CPU 上运行。您可以使用量化感知训练或训练后量化来量化您的 LiteRT 模型。请确保按照”全整数量化(Full integer quantization)“的步骤操作。

不想自己进行量化? 您可以从 Qualcomm AI Hub 下载一系列预量化模型,或使用 Edge Impulse 对新模型或现有模型进行量化。

在 NPU 上运行模型(Python)

要将模型卸载到 NPU,您只需加载 LiteRT 委托并将其传入解释器。例如:

from ai_edge_litert.interpreter import Interpreter, load_delegate

qnn_delegate = load_delegate("libQnnTFLiteDelegate.so", options={"backend_type": "htp"})
interpreter = Interpreter(
    model_path=...,
    experimental_delegates=[qnn_delegate]
)

在 NPU 上运行模型(C++)

要将模型卸载到 NPU,首先需要添加以下编译标志:

CFLAGS += -I${QNN_SDK_ROOT}/include
LDFLAGS += -L${QNN_SDK_ROOT}/lib/aarch64-ubuntu-gcc9.4 -lQnnTFLiteDelegate

然后,实例化 LiteRT 委托并将其传递给 LiteRT 解释器:

// == Includes ==
#include "QNN/TFLiteDelegate/QnnTFLiteDelegate.h"

// == Application code ==

// Get your interpreter...
tflite::Interpreter *interpreter = ...;

// Create QNN Delegate options structure.
TfLiteQnnDelegateOptions options = TfLiteQnnDelegateOptionsDefault();

// Set the mandatory backend_type option. All other options have default values.
options.backend_type = kHtpBackend;

// Instantiate delegate. Must not be freed until interpreter is freed.
TfLiteDelegate* delegate = TfLiteQnnDelegateCreate(&options);

TfLiteStatus status = interpreter->ModifyGraphWithDelegate(delegate);
// Check that status == kTfLiteOk

Python 示例

前提条件

应已刷写 Ubuntu OS
具备相应权限的终端访问
如果您之前尚未安装 PPA 软件包,请按照以下步骤安装:https://qualcomm-3.mintlify.io/devices/iq9075-evk/update-software/upgrade-ubuntu#4-upgrade-pre-built-packages

在开发板上打开终端,或通过 SSH 会话连接到开发板,创建一个新的 venv,并安装 LiteRT 运行时和 Pillow:

python3 -m venv .venv-litert-demo --system-site-packages
source .venv-litert-demo/bin/activate
pip3 install ai-edge-litert==1.3.0 Pillow
pip3 install opencv-python

要准备开发环境,请安装以下软件包。这些软件包提供了 GTK 绑定、Python 开发工具以及编译和运行应用程序所需的构建工具等基本组件。

sudo apt install -y python3-gi python3-gi-cairo gir1.2-gtk-3.0
sudo apt install -y build-essential python3-dev python3-pip python3-venv python3-full pkg-config meson
sudo apt install -y pkg-config cmake libcairo2-dev
sudo apt install -y libgirepository1.0-dev gir1.2-glib-2.0

Vision Transformer
图像分类
目标检测

Vision Transformer

以下展示了如何使用 LiteRT 委托,在 CPU 和 NPU 上分别运行 Vision Transformer 模型(从 AI Hub 下载)。

创建推理脚本

创建 inference_vit.py 并添加以下代码:

import numpy as np
from ai_edge_litert.interpreter import Interpreter, load_delegate
from PIL import Image
import os, time, sys
import urllib.request

def curr_ms():
    return round(time.time() * 1000)

use_npu = True if len(sys.argv) >= 2 and sys.argv[1] == '--use-npu' else False

# Path to your quantized TFLite model and test image (will be download automatically)
MODEL_PATH = "vit-vit-w8a8.tflite"
IMAGE_PATH = "boa-constrictor.jpg"
LABELS_PATH = "vit-vit-labels.txt"

if not os.path.exists(MODEL_PATH):
    print("Downloading model...")
    model_url = 'https://cdn.edgeimpulse.com/qc-ai-docs/models/vit-vit-w8a8.tflite'
    urllib.request.urlretrieve(model_url, MODEL_PATH)

if not os.path.exists(LABELS_PATH):
    print("Downloading labels...")
    labels_url = 'https://cdn.edgeimpulse.com/qc-ai-docs/models/vit-vit-labels.txt'
    urllib.request.urlretrieve(labels_url, LABELS_PATH)

if not os.path.exists(IMAGE_PATH):
    print("Downloading image...")
    image_url = 'https://cdn.edgeimpulse.com/qc-ai-docs/examples/boa-constrictor.jpg'
    urllib.request.urlretrieve(image_url, IMAGE_PATH)

with open(LABELS_PATH, 'r') as f:
    labels = [line for line in f.read().splitlines() if line.strip()]

experimental_delegates = []
if use_npu:
    experimental_delegates = [load_delegate("libQnnTFLiteDelegate.so", options={"backend_type": "htp"})]

# Load TFLite model and allocate tensors
interpreter = Interpreter(
    model_path=MODEL_PATH,
    experimental_delegates=experimental_delegates
)
interpreter.allocate_tensors()

# Get input and output tensor details
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Load and preprocess image
def load_image(path, input_shape):
    # Expected input shape: [1, height, width, channels]
    _, height, width, channels = input_shape

    img = Image.open(path).convert("RGB").resize((width, height))
    img_np = np.array(img, dtype=np.uint8)  # quantized models expect uint8
    img_np = np.expand_dims(img_np, axis=0)
    return img_np

input_shape = input_details[0]['shape']
input_data = load_image(IMAGE_PATH, input_shape)

# Set tensor and run inference
interpreter.set_tensor(input_details[0]['index'], input_data)

# Run once to warmup
interpreter.invoke()

# Then run 10x
start = curr_ms()
for i in range(0, 10):
    interpreter.invoke()
end = curr_ms()

# Get prediction
q_output = interpreter.get_tensor(output_details[0]['index'])
scale, zero_point = output_details[0]['quantization']
f_output = (q_output.astype(np.float32) - zero_point) * scale

# Image classification models in AI Hub miss a Softmax() layer at the end of the model, so add it manually
def softmax(x, axis=-1):
    # subtract max for numerical stability
    x_max = np.max(x, axis=axis, keepdims=True)
    e_x = np.exp(x - x_max)
    return e_x / np.sum(e_x, axis=axis, keepdims=True)

# show top-5 predictions
scores = softmax(f_output[0])
top_k = scores.argsort()[-5:][::-1]
print("\nTop-5 predictions:")
for i in top_k:
    print(f"Class {labels[i]}: score={scores[i]}")

print('')
print(f'Inference took (on average): {(end - start) / 10}ms. per image')

在 CPU 上运行

python3 inference_vit.py

# INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
#
# Top-5 predictions:
# Class boa constrictor: score=0.6264431476593018
# Class rock python: score=0.047579940408468246
# Class night snake: score=0.006721484009176493
# Class mouse: score=0.0022421202156692743
# Class pick: score=0.001942973816767335
#
# Inference took (on average): 300.8ms. per image

在 NPU 上运行

python3 inference_vit.py --use-npu

# INFO: TfLiteQnnDelegate delegate: 1382 nodes delegated out of 1633 nodes with 27 partitions.
#
# INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
#
# Top-5 predictions:
# Class boa constrictor: score=0.6113042235374451
# Class rock python: score=0.038359832018613815
# Class night snake: score=0.011630792170763016
# Class mouse: score=0.002294909441843629
# Class lens cap: score=0.0018960189772769809
#
# Inference took (on average): 13.9ms. per image

该模型在 NPU 上运行速度显著更快 — 但模型输出的精度略有下降。您还可以看到,对于该模型,并非所有层都能在 NPU 上运行(“1382 nodes delegated out of 1633 nodes with 27 partitions”)。

基于 GTK 的图像分类应用

以下展示了如何使用基于 GTK 的桌面应用程序,通过 AI Engine Direct 中的 LiteRT 委托,在 CPU 和 NPU 上运行图像分类模型 — 该模型从 AI Hub 下载。GoogLeNet_w8a8.tflite 模型来自 AI Hub,借助带有 QNN 委托加速的 TensorFlow Lite,实现高效的设备端推理。

属性	值
类型	桌面 GUI 应用程序
功能	使用 TFLite 进行图像分类
模式	CPU 和 QNN 委托
界面	基于 GTK 的 GUI
输出	带置信度条的 Top 预测结果

环境设置与导入

该脚本设置与显示相关的环境变量(适用于 Linux 系统),并导入 OpenCV、NumPy、GTK 和 TensorFlow Lite 等必要的库。

import cv2, numpy as np, os, time
from gi.repository import Gtk, GLib, GdkPixbuf
import ai_edge_litert.interpreter as tflite

GTK 用于 GUI,OpenCV 用于图像处理,TensorFlow Lite 用于推理。

配置常量

这些常量定义了模型、标签文件和委托库的路径。

TF_MODEL = "/etc/models/inception_v3_quantized.tflite"
LABELS = "/etc/labels/imagenet_labels.txt"
DELEGATE_PATH = "libQnnTFLiteDelegate.so"
DEVICE_OS = "Ubuntu"

下载 TFLite 模型

该脚本检查本地是否存在 TensorFlow Lite 模型文件,如果不存在,则从指定的 Hugging Face URL 下载。

import urllib.request
if not os.path.exists(TF_MODEL):
    print("Downloading model...")
    model_url = 'https://huggingface.co/qualcomm/GoogLeNet/resolve/main/GoogLeNet_w8a8.tflite'
    urllib.request.urlretrieve(model_url, TF_MODEL)

辅助函数

Softmax 计算 — 在将 logits 转换为概率时确保数值稳定性:

def stable_softmax(logits):
    logits = logits.astype(np.float32)
    shifted_logits = np.clip(logits - np.max(logits), -500, 500)
    exp_scores = np.exp(shifted_logits)
    return exp_scores / np.sum(exp_scores)

标签加载器 — 从文本文件加载类别标签:

def load_labels(label_path):
    with open(label_path, 'r') as f:
        return [line.strip() for line in f.readlines()]

图像预处理 — 为模型输入准备图像:

def preprocess_image(image_path, input_shape, input_dtype):
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (input_shape[2], input_shape[1]))
    img = img.astype(input_dtype)
    return np.expand_dims(img, axis=0)

执行推理

该函数加载模型(带或不带委托),准备输入,运行推理,应用 softmax,并返回带置信度分数的前 4 个预测结果。

def runInference(image, use_delegate):
    if use_delegate:
        try:
            delegate = tflite.load_delegate(DELEGATE_PATH, {'backend_type': 'htp'})
            model = tflite.Interpreter(model_path=TF_MODEL, experimental_delegates=[delegate])
        except:
            model = tflite.Interpreter(model_path=TF_MODEL)
    else:
        model = tflite.Interpreter(model_path=TF_MODEL)
    model.allocate_tensors()
    input_details = model.get_input_details()
    input_data = preprocess_image(image, input_details[0]['shape'], input_details[0]['dtype'])
    model.set_tensor(input_details[0]['index'], input_data)
    start_time = time.time()
    model.invoke()
    inference_time = time.time() - start_time
    output_data = model.get_tensor(model.get_output_details()[0]['index'])
    probabilities = stable_softmax(output_data[0])
    labels = load_labels(LABELS)
    top_indices = np.argsort(probabilities)[::-1][:4]
    results = [(labels[i], probabilities[i] * 100) for i in top_indices]
    return results, inference_time

GTK GUI 与程序入口

该 GUI 包含图像显示区域、用于选择 CPU 或委托的单选按钮、用于选择和重新处理图像的按钮,以及带标签和进度条的结果显示。

class FileBrowser(Gtk.FileChooserDialog):
    def __init__(self):
        super().__init__(title="Choose an image", action=Gtk.FileChooserAction.OPEN)
        self.add_buttons(Gtk.STOCK_CANCEL, Gtk.ResponseType.CANCEL, Gtk.STOCK_OPEN, Gtk.ResponseType.OK)
    def run_and_get_file(self):
        if self.run() == Gtk.ResponseType.OK:
            return self.get_filename()
        self.destroy()

def main():
    app = MainWindow()
    app.connect("destroy", Gtk.main_quit)
    app.show_all()
    Gtk.main()
if __name__ == "__main__":
    success, _ = Gtk.init_check()
    if not success:
        print("GTK could not be initialized.")
        exit(1)
    main()

运行应用程序

python3 classification.py

从互联网下载任意图片。在此示例中,我们使用了一张消防车的图片。通过 scp 命令将图片复制到设备上:

scp xxx.jpg ubuntu@IP_address:/home/ubuntu/

在 GUI 上选择 CPU 作为运行时选项:

在 GUI 上选择委托(delegate)作为运行时选项:

完整参考代码

# -----------------------------------------------------------------------------
#
# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
# SPDX-License-Identifier: BSD-3-Clause
#
# -----------------------------------------------------------------------------
import cv2
import gi
import numpy as np
import os
os.environ['xDG_RUNTIME_DIR'] = '/run/user/1000/'
os.environ['WAYLAND_DISPLAY'] = 'wayland-1'
os.environ['DISPLAY'] = ':0'
import time
import urllib.request
gi.require_version("Gtk", "3.0")
from gi.repository import Gtk, GLib, GdkPixbuf

# ========= Constants =========
TF_MODEL = "/home/ubuntu/GoogLeNet_w8a8.tflite"
LABELS = "/etc/labels/imagenet_labels.txt"
DELEGATE_PATH = "libQnnTFLiteDelegate.so"
DEVICE_OS="Ubuntu"
UNAME = os.uname().nodename

import ai_edge_litert.interpreter as tflite

if not os.path.exists(TF_MODEL):
    print("Downloading model...")
    model_url = 'https://huggingface.co/qualcomm/GoogLeNet/resolve/main/GoogLeNet_w8a8.tflite'
    urllib.request.urlretrieve(model_url, TF_MODEL)

# ========= Helper Functions =========
def stable_softmax(logits):
    logits = logits.astype(np.float32)
    shifted_logits = logits - np.max(logits)
    shifted_logits = np.clip(shifted_logits, -500, 500)
    exp_scores = np.exp(shifted_logits)
    probabilities = exp_scores / np.sum(exp_scores)
    return probabilities

def load_labels(label_path):
    with open(label_path, 'r') as f:
        return [line.strip() for line in f.readlines()]

def resizeImage(pixbuf):
    original_width = pixbuf.get_width()
    original_height = pixbuf.get_height()
    max_width = 800
    max_height = 600
    scale = min(max_width / original_width, max_height / original_height)
    new_width = int(original_width * scale)
    new_height = int(original_height * scale)
    return new_width, new_height

def preprocess_image(image_path, input_shape, input_dtype):
    img = cv2.imread(image_path)
    if img is None:
        raise ValueError(f"Failed to load image at {image_path}")
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (input_shape[2], input_shape[1]))
    img = img.astype(input_dtype)
    img = np.expand_dims(img, axis=0)
    return img

# ====== Inference Function ======
def runInference(image, use_delegate):
    results = []
    print(f"Running on {DEVICE_OS} using Delegate:{use_delegate}")
    if use_delegate:
        try:
            delegate_options = { 'backend_type': 'htp' }
            delegate = tflite.load_delegate(DELEGATE_PATH, delegate_options)
            model = tflite.Interpreter(model_path=TF_MODEL, experimental_delegates=[delegate])
            print("INFO: Loaded QNN delegate with HTP backend")
        except Exception as e:
            print(f"WARNING: Failed to load QNN delegate: {e}")
            print("INFO: Continuing without QNN delegate")
            model = tflite.Interpreter(model_path=TF_MODEL)
    else:
        model = tflite.Interpreter(model_path=TF_MODEL)

    model.allocate_tensors()
    input_details = model.get_input_details()
    input_shape = input_details[0]['shape']
    input_dtype = input_details[0]['dtype']
    input_data = preprocess_image(image, input_shape, input_dtype)
    model.set_tensor(input_details[0]['index'], input_data)
    model.get_signature_list()

    try:
        start_time = time.time()
        model.invoke()
        end_time = time.time()
        print("Interpreter invoked successfully.")
    except Exception as e:
        print(f"Error during model invocation: {e}")
        return []

    inference_time = end_time - start_time
    output_details = model.get_output_details()
    output_data = model.get_tensor(output_details[0]['index'])
    labels = load_labels(LABELS)
    predicted_index = np.argmax(output_data)
    predicted_label = labels[predicted_index]
    print("Predicted index:", predicted_index)
    print("Predicted label:", predicted_label)

    logits = output_data[0]
    probabilities = stable_softmax(logits)
    top_k = 4
    top_indices = np.argsort(probabilities)[::-1][:top_k]
    for i in top_indices:
        result = (labels[i], probabilities[i] * 100)
        results.append(result)

    return results, inference_time

# ====== GTK GUI Classes ======
class FileBrowser(Gtk.FileChooserDialog):
    def __init__(self):
        super().__init__(title="Choose an image", action=Gtk.FileChooserAction.OPEN)
        self.add_buttons(Gtk.STOCK_CANCEL, Gtk.ResponseType.CANCEL, Gtk.STOCK_OPEN, Gtk.ResponseType.OK)

    def run_and_get_file(self):
        response = super().run()
        if response == Gtk.ResponseType.OK:
            print("Selected file:", self.get_filename())
            self.selected_file = self.get_filename()
        self.destroy()
        return self.selected_file

class MainWindow(Gtk.Window):
    def __init__(self):
        super().__init__(title="Image Classification")
        self.set_default_size(800, 600)
        self.imageFilepath = ""
        self.mainBox = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=10)
        self.mainBox.set_margin_top(10)
        self.mainBox.set_margin_bottom(10)
        self.mainBox.set_margin_start(10)
        self.mainBox.set_margin_end(10)
        self.add(self.mainBox)

        self.image = Gtk.Image()
        try:
            MAIN_IMAGE = "MainWindowPic.jpg"
            self.image.set_from_file(MAIN_IMAGE)
        except Exception as e:
            print("Error loading main image:", e)
            self.image.set_from_icon_name("image-missing", Gtk.IconSize.DIALOG)

        self.mainBox.pack_start(self.image, True, True, 0)
        self.infoBox = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=10)

        delegate_label = Gtk.Label(label="Select Inference Mode:")
        self.infoBox.pack_start(delegate_label, False, False, 10)
        self.cpu_radio = Gtk.RadioButton.new_with_label_from_widget(None, "CPU")
        self.delegate_radio = Gtk.RadioButton.new_with_label_from_widget(self.cpu_radio, "Delegate")
        self.infoBox.pack_start(self.cpu_radio, False, False, 0)
        self.infoBox.pack_start(self.delegate_radio, False, False, 0)
        self.cpu_radio.connect("toggled", self.on_radio_toggled)
        self.delegate_radio.connect("toggled", self.on_radio_toggled)

        open_button = Gtk.Button(label="Select Image")
        open_button.connect("clicked", self.on_open_file_clicked)
        self.infoBox.pack_start(open_button, False, True, 10)

        reprocess_button = Gtk.Button(label="Reprocess Image")
        reprocess_button.connect("clicked", self.on_reprocess_image_clicked)
        self.infoBox.pack_start(reprocess_button, False, True, 10)

        self.results = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=10)
        self.infoBox.pack_start(self.results, True, True, 0)
        self.mainBox.pack_start(self.infoBox, True, True, 0)

    def use_delegate(self):
        return self.delegate_radio.get_active()

    def on_radio_toggled(self, button):
        if button.get_active():
            print(f"Selected option: {button.get_label()}")

    def process_file(self, filepath):
        try:
            pixbuf = GdkPixbuf.Pixbuf.new_from_file(filepath)
            new_width, new_height = resizeImage(pixbuf)
            scaled_pixbuf = pixbuf.scale_simple(new_width, new_height, GdkPixbuf.InterpType.BILINEAR)
            self.image.set_from_pixbuf(scaled_pixbuf)

            use_delegate = self.use_delegate()
            print("delegate: " , use_delegate)
            options, inference_time = runInference(filepath, use_delegate)

            for child in self.results.get_children():
                self.results.remove(child)

            for label, percent in options:
                textBox = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=10)
                barBox = Gtk.Box(orientation=Gtk.Orientation.HORIZONTAL, spacing=10)
                text = Gtk.Label(label=label, xalign=0)
                text.set_size_request(100, -1)
                bar = Gtk.ProgressBar()
                bar.set_fraction(percent / 100.0)
                bar.set_text(f"{percent:.2f}%")
                bar.set_show_text(True)
                textBox.pack_start(text, False, False, 0)
                barBox.pack_start(bar, True, True, 0)
                self.results.pack_start(textBox, False, False, 0)
                self.results.pack_start(barBox, False, False, 0)
                self.results.show_all()

            time_label = Gtk.Label(label=f"Inference Time : {inference_time * 1000:.2f} ms")
            self.results.pack_start(time_label, False, False, 50)
            self.results.show_all()
        except Exception as e:
            print("Error reading file:", e)

    def on_open_file_clicked(self, widget):
        dialog = FileBrowser()
        selected_file = dialog.run_and_get_file()
        self.imageFilepath = selected_file
        if selected_file:
            self.process_file(selected_file)

    def on_reprocess_image_clicked(self, widget):
        self.process_file(self.imageFilepath)

    def on_destroy(self, widget):
        Gtk.main_quit()

# === Main Entry Point ===
def main():
    app = MainWindow()
    app.connect("destroy", Gtk.main_quit)
    app.show_all()
    Gtk.main()

if __name__ == "__main__":
    success, _ = Gtk.init_check()
    if not success:
        print("GTK could not be initialized. Check environmental variables")
        exit(1)
    main()

使用 OpenCV 和 Wayland 显示进行目标检测

该 Python 脚本使用量化的 YOLOv8 TensorFlow Lite 模型对视频文件执行实时目标检测,并通过 GStreamer 在 Wayland 显示器上展示带标注的帧。它通过 QNN TFLite 委托利用硬件加速,针对边缘 AI 场景进行了优化。

YOLOv8 模型默认不可用。请按照 Qualcomm Intelligent Multimedia SDK 中的步骤 6 导出 YOLOv8 量化模型。

将模型推送到目标设备:

scp xxxx.tflite ubuntu@IP_address:/home/ubuntu/

配置

定义了模型、标签、输入视频和委托的路径。同时设置了帧尺寸、FPS、置信度阈值和缩放因子等常量,用于预处理和后处理。

参数	值
`MODEL_PATH`	`yolov8_det_quantized.tflite`
`LABEL_PATH`	`coco_labels.txt`
`VIDEO_IN`	`video.mp4`
`DELEGATE_PATH`	`libQnnTFLiteDelegate.so`

请使用适合目标检测模型的视频文件。为获得最佳效果,请选择主体清晰、光照良好且运动模糊较少的视频(例如街景、仓库地面、静态摄像头画面)。

模型加载与委托设置

加载硬件委托以实现加速推理,并使用量化的 YOLOv8 模型初始化 TensorFlow Lite 解释器。

delegate_options = { 'backend_type': 'htp' }
delegate = tflite.load_delegate(DELEGATE_PATH, delegate_options)
interpreter = tflite.Interpreter(model_path=MODEL_PATH, experimental_delegates=[delegate])
interpreter.allocate_tensors()

GStreamer 流水线设置

使用 appsrc 创建 GStreamer 流水线,将帧流式传输到 Wayland sink,以实时显示处理后的帧。

pipeline = Gst.parse_launch(
    'appsrc name=src is-live=true block=true format=time '
    'caps=video/x-raw,format=BGR,width=1600,height=900,framerate=30/1 '
    '! videoconvert ! waylandsink')

推理与后处理

对每一帧运行推理,对输出进行反量化,应用置信度阈值,并使用非极大值抑制(NMS)去除重叠的检测框。

interpreter.set_tensor(in_det[0]['index'], input_tensor)
interpreter.invoke()
boxes_q = interpreter.get_tensor(out_det[0]['index'])[0]
scores_q = interpreter.get_tensor(out_det[1]['index'])[0]
classes_q = interpreter.get_tensor(out_det[2]['index'])[0]

# Filter by confidence threshold
mask = scores >= CONF_THRES
boxes_f = boxes[mask]
scores_f = scores[mask]
classes_f = classes[mask]

标注与显示

使用 OpenCV 在帧上绘制边界框和标签,然后流式传输到 Wayland 显示器。

cv2.rectangle(frame_rs, (x1i, y1i), (x2i, y2i), (0,255,0), 2)
cv2.putText(frame_rs, f"{lab} {sc:.2f}", (x1i, max(10,y1i-5)),
            cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 2)

运行应用程序

python3 ObjectDetection.py

完整参考代码

# -----------------------------------------------------------------------------
#
# Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
# SPDX-License-Identifier: BSD-3-Clause
#
# -----------------------------------------------------------------------------

#!/usr/bin/env python3

import cv2
import numpy as np
import gi
gi.require_version('Gst', '1.0')
from gi.repository import Gst
import ai_edge_litert.interpreter as tflite

Gst.init(None)

# -------------------- Parameters --------------------
MODEL_PATH = "/etc/models/yolox_quantized.tflite"
LABEL_PATH = "/etc/labels/coco_labels.txt"
VIDEO_IN = "/etc/media/video.mp4"
DELEGATE_PATH = "libQnnTFLiteDelegate.so"

FRAME_W, FRAME_H = 1600, 900
FPS_OUT = 30
CONF_THRES = 0.25
NMS_IOU_THRES = 0.50
BOX_SCALE = 3.2108588218688965
BOX_ZP = 31.0
SCORE_SCALE = 0.0038042240776121616

# -------------------- Load Model --------------------
delegate_options = { 'backend_type': 'htp' }
delegate = tflite.load_delegate(DELEGATE_PATH, delegate_options)
interpreter = tflite.Interpreter(model_path=MODEL_PATH, experimental_delegates=[delegate])
interpreter.allocate_tensors()

in_det = interpreter.get_input_details()
out_det = interpreter.get_output_details()
in_h, in_w = in_det[0]["shape"][1:3]

# -------------------- Load Labels --------------------
labels = [l.strip() for l in open(LABEL_PATH)]

# -------------------- GStreamer Pipeline --------------------
pipeline = Gst.parse_launch(
    'appsrc name=src is-live=true block=true format=time '
    'caps=video/x-raw,format=BGR,width=1600,height=900,framerate=30/1 '
    '! videoconvert ! waylandsink'
)
appsrc = pipeline.get_by_name('src')
pipeline.set_state(Gst.State.PLAYING)

# -------------------- Video Input --------------------
cap = cv2.VideoCapture(VIDEO_IN)
sx, sy = FRAME_W / in_w, FRAME_H / in_h
frame_rs = np.empty((FRAME_H, FRAME_W, 3), np.uint8)
input_tensor = np.empty((1, in_h, in_w, 3), np.uint8)
frame_cnt = 0

# -------------------- Main Loop --------------------
while True:
    ok, frame = cap.read()
    if not ok:
        break
    frame_cnt += 1

    cv2.resize(frame, (FRAME_W, FRAME_H), dst=frame_rs)
    cv2.resize(frame_rs, (in_w, in_h), dst=input_tensor[0])

    interpreter.set_tensor(in_det[0]['index'], input_tensor)
    interpreter.invoke()

    boxes_q = interpreter.get_tensor(out_det[0]['index'])[0]
    scores_q = interpreter.get_tensor(out_det[1]['index'])[0]
    classes_q = interpreter.get_tensor(out_det[2]['index'])[0]

    boxes = BOX_SCALE * (boxes_q.astype(np.float32) - BOX_ZP)
    scores = SCORE_SCALE * scores_q.astype(np.float32)
    classes = classes_q.astype(np.int32)

    mask = scores >= CONF_THRES
    if np.any(mask):
        boxes_f = boxes[mask]
        scores_f = scores[mask]
        classes_f = classes[mask]

        x1, y1, x2, y2 = boxes_f.T
        boxes_cv2 = np.column_stack((x1, y1, x2 - x1, y2 - y1))

        idx_cv2 = cv2.dnn.NMSBoxes(
            bboxes=boxes_cv2.tolist(),
            scores=scores_f.tolist(),
            score_threshold=CONF_THRES,
            nms_threshold=NMS_IOU_THRES
        )

        if len(idx_cv2):
            idx = idx_cv2.flatten()
            sel_boxes = boxes_f[idx]
            sel_scores = scores_f[idx]
            sel_classes = classes_f[idx]

            if frame_cnt % 100 == 0:
                print(f"[{frame_cnt:4d}] max score = {sel_scores.max():.3f}")

            sel_boxes[:, [0,2]] *= sx
            sel_boxes[:, [1,3]] *= sy
            sel_boxes = sel_boxes.astype(np.int32)
            sel_boxes[:, [0,2]] = np.clip(sel_boxes[:, [0,2]], 0, FRAME_W-1)
            sel_boxes[:, [1,3]] = np.clip(sel_boxes[:, [1,3]], 0, FRAME_H-1)

            for (x1i, y1i, x2i, y2i), sc, cl in zip(sel_boxes, sel_scores, sel_classes):
                cv2.rectangle(frame_rs, (x1i, y1i), (x2i, y2i), (0,255,0), 2)
                lab = labels[cl] if cl < len(labels) else str(cl)
                cv2.putText(frame_rs, f"{lab} {sc:.2f}", (x1i, max(10,y1i-5)),
                            cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0,255,0), 2)

    data = frame_rs.tobytes()
    buf = Gst.Buffer.new_allocate(None, len(data), None)
    buf.fill(0, data)
    buf.duration = Gst.util_uint64_scale_int(1, Gst.SECOND, FPS_OUT)
    timestamp = cap.get(cv2.CAP_PROP_POS_MSEC) * Gst.MSECOND
    buf.pts = buf.dts = int(timestamp)
    appsrc.emit('push-buffer', buf)

# -------------------- Finish --------------------
appsrc.emit('end-of-stream')
pipeline.set_state(Gst.State.NULL)
cap.release()
print("Done – video streamed to Wayland sink")

​量化模型

​在 NPU 上运行模型(Python)

​在 NPU 上运行模型(C++)

​Python 示例

​Vision Transformer

​基于 GTK 的图像分类应用

​使用 OpenCV 和 Wayland 显示进行目标检测

量化模型

在 NPU 上运行模型(Python)

在 NPU 上运行模型(C++)

Python 示例

Vision Transformer

基于 GTK 的图像分类应用

使用 OpenCV 和 Wayland 显示进行目标检测