AI_Cube目标检测导出的代码标签置信度全部一样

Viewed 77
model ai_cube micropython suggestion k230 support

问题描述

det_result.jpg
我使用AI_Cube示例数据进行训练，部署后生成的det_image.py中，每个目标的所有标签置信度都是一样的。但是我在AI_Cube模型评估中看到的测试集数据推理结果没有出现这样的问题。
50cd7b96af0954b1e28fc420384ff2fb_720.png
是部署出现了什么问题吗？我除了分辨率之外没有修改过生成的程序。

硬件板卡

亚博

软件版本

CanMV_K230_YAHBOOM_micropython_local_nncase_v2.9.0

其他信息

训练模型：AnchorFreeDet，其余参数为默认

Walnutzz3

而且不只是置信度，aicube.anchorfreedet_post_process输出的det_boxes只有[0]是不一样的，也就是只有tag不一样，其它参数都输出一样的。

Walnutzz•

分辨率改成多少了

Wy001•

部署时推理宽高是224*224

Walnutzz•
4 Answers

你好，排查后发现确实存在一些问题：

  1. 在进行模型部署的时候不要修改分辨率，用什么分辨率训练就用什么分辨率部署；
  2. 部署时参数选择NoClip和w:int16 d:uint8；
  3. 烧录最新的固件https://kendryte-download.canaan-creative.com/developer/releases/canmv_k230_micropython/daily_build/选择对应的开发板的固件下载烧录，使用cv_lite模块加载图片，使用openmv的image模块可能会有问题；
  4. 为了方便查看结果，调整代码的绘制结果部分；
  5. 推荐使用模型AnchorBaseDet，模型量化存在性能损失，需要通过调整阈值进行调优；
  6. 摄像头得到的数据是RGB的原始数据，和读入图片是不一样的，因此视频推理应该不会受影响，但是会受到数据集的影响，如果能使用230开发板采集的图片进行训练，可能能得到更好的效果；
    这里附上几张AICube测试和开发板推理的对比图：
  • 0094.jpg
    image.png
    image.png
  • 0180.jpg
    image.png
    image.png

这里把测试代码也附上：

import os
import ujson
import aicube
from media.sensor import *
from media.display import *
from media.media import *
from time import *
import nncase_runtime as nn
import ulab.numpy as np
import time
import image
import random
import gc
import utime
import cv_lite


DISPLAY_WIDTH = ALIGN_UP(1920, 16)
DISPLAY_HEIGHT = 1080


color_three = [(220, 20, 60), (119, 11, 32), (0, 0, 142), (0, 0, 230),
        (106, 0, 228), (0, 60, 100), (0, 80, 100), (0, 0, 70),
        (0, 0, 192), (250, 170, 30), (100, 170, 30), (220, 220, 0),
        (175, 116, 175), (250, 0, 30), (165, 42, 42), (255, 77, 255),
        (0, 226, 252), (182, 182, 255), (0, 82, 0), (120, 166, 157),
        (110, 76, 0), (174, 57, 255), (199, 100, 0), (72, 0, 118),
        (255, 179, 240), (0, 125, 92), (209, 0, 151), (188, 208, 182),
        (0, 220, 176), (255, 99, 164), (92, 0, 73), (133, 129, 255),
        (78, 180, 255), (0, 228, 0), (174, 255, 243), (45, 89, 255),
        (134, 134, 103), (145, 148, 174), (255, 208, 186),
        (197, 226, 255), (171, 134, 1), (109, 63, 54), (207, 138, 255),
        (151, 0, 95), (9, 80, 61), (84, 105, 51), (74, 65, 105),
        (166, 196, 102), (208, 195, 210), (255, 109, 65), (0, 143, 149),
        (179, 0, 194), (209, 99, 106), (5, 121, 0), (227, 255, 205),
        (147, 186, 208), (153, 69, 1), (3, 95, 161), (163, 255, 0),
        (119, 0, 170), (0, 182, 199), (0, 165, 120), (183, 130, 88),
        (95, 32, 0), (130, 114, 135), (110, 129, 133), (166, 74, 118),
        (219, 142, 185), (79, 210, 114), (178, 90, 62), (65, 70, 15),
        (127, 167, 115), (59, 105, 106), (142, 108, 45), (196, 172, 0),
        (95, 54, 80), (128, 76, 255), (201, 57, 1), (246, 0, 122),
        (191, 162, 208)]

root_path="/sdcard/mp_deployment_source/"  # root_path要以/结尾
config_path=root_path+"deploy_config.json"
image_path=root_path+"0180.jpg"
deploy_conf={}
debug_mode=1

class ScopedTiming:
    def __init__(self, info="", enable_profile=True):
        self.info = info
        self.enable_profile = enable_profile

    def __enter__(self):
        if self.enable_profile:
            self.start_time = time.time_ns()
        return self

    def __exit__(self, exc_type, exc_value, traceback):
        if self.enable_profile:
            elapsed_time = time.time_ns() - self.start_time
            print(f"{self.info} took {elapsed_time / 1000000:.2f} ms")

def read_deploy_config(config_path):
    # 打开JSON文件以进行读取deploy_config
    with open(config_path, 'r') as json_file:
        try:
            # 从文件中加载JSON数据
            config = ujson.load(json_file)

            # 打印数据（可根据需要执行其他操作）
            #print(config)
        except ValueError as e:
            print("JSON 解析错误:", e)
    return config

def read_img(img_path):
    # openmv image加载图片
#    img_data = image.Image(img_path)
#    img_data_rgb888=img_data.to_rgb888()
#    img_hwc=img_data_rgb888.to_numpy_ref()
    # cv_lite加载图片
    img_hwc=cv_lite.load_image(img_path)[0]
    print(img_hwc.shape)
    shape=img_hwc.shape
    img_tmp = img_hwc.reshape((shape[0] * shape[1], shape[2]))
    img_tmp_trans = img_tmp.transpose()
    img_res=img_tmp_trans.copy()
    img_return=img_res.reshape((shape[2],shape[0],shape[1]))
    return img_return

def detection():
    print("--------------start-----------------")
    # 使用json读取内容初始化部署变量
    deploy_conf=read_deploy_config(config_path)
    kmodel_name=deploy_conf["kmodel_path"]
    labels=deploy_conf["categories"]
    # 注意调整阈值
    confidence_threshold = 0.45
    nms_threshold = 0.4
    img_size=deploy_conf["img_size"]
    num_classes=deploy_conf["num_classes"]
    nms_option = deploy_conf["nms_option"]
    model_type = deploy_conf["model_type"]
    if model_type == "AnchorBaseDet":
        anchors = deploy_conf["anchors"][0] + deploy_conf["anchors"][1] + deploy_conf["anchors"][2]
    kmodel_frame_size = img_size
    strides = [8,16,32]

    # ai2d输入输出初始化
    ai2d_input = read_img(image_path)
    frame_size = [ai2d_input.shape[2],ai2d_input.shape[1]]
    ai2d_input_tensor = nn.from_numpy(ai2d_input)
    ai2d_input_shape=ai2d_input.shape
    data = np.ones((1,3,img_size[0],img_size[1]),dtype=np.uint8)
    ai2d_out = nn.from_numpy(data)

    # 计算padding值
    ori_w = ai2d_input_shape[2]
    ori_h = ai2d_input_shape[1]
    width = kmodel_frame_size[0]
    height = kmodel_frame_size[1]
    ratiow = float(width) / ori_w
    ratioh = float(height) / ori_h
    if ratiow < ratioh:
        ratio = ratiow
    else:
        ratio = ratioh
    new_w = int(ratio * ori_w)
    new_h = int(ratio * ori_h)
    dw = float(width - new_w) / 2
    dh = float(height - new_h) / 2
    top = int(round(dh - 0.1))
    bottom = int(round(dh + 0.1))
    left = int(round(dw - 0.1))
    right = int(round(dw + 0.1))

    # init kpu and load kmodel
    kpu = nn.kpu()
    ai2d = nn.ai2d()
    kpu.load_kmodel(root_path+kmodel_name)
    ai2d.set_dtype(nn.ai2d_format.NCHW_FMT,
                                   nn.ai2d_format.NCHW_FMT,
                                   np.uint8, np.uint8)
    ai2d.set_pad_param(True, [0,0,0,0,top,bottom,left,right], 0, [114,114,114])
    ai2d.set_resize_param(True, nn.interp_method.tf_bilinear, nn.interp_mode.half_pixel )
    ai2d_builder = ai2d.build([1,3,ai2d_input_shape[1],ai2d_input_shape[2]], [1,3,height,width])
    with ScopedTiming("total",debug_mode > 0):
        ai2d_builder.run(ai2d_input_tensor, ai2d_out)
        kpu.set_input_tensor(0, ai2d_out)
        kpu.run()
        del ai2d_input_tensor
        del ai2d_out

        # get output
        results = []
        for i in range(kpu.outputs_size()):
            data = kpu.get_output_tensor(i)
            result = data.to_numpy()
            result = result.reshape((result.shape[0]*result.shape[1]*result.shape[2]*result.shape[3]))
            del data
            results.append(result)
        gc.collect()

        image_draw=image.Image(image_path).to_rgb565()
        # postprocess
        if model_type == "AnchorBaseDet":
            det_boxes = aicube.anchorbasedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, anchors, nms_option)
        elif model_type == "GFLDet":
            det_boxes = aicube.gfldet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
        else:
            det_boxes = aicube.anchorfreedet_post_process( results[0], results[1], results[2], kmodel_frame_size, frame_size, strides, num_classes, confidence_threshold, nms_threshold, nms_option)
        if det_boxes:
            for det_boxe in det_boxes:
                x1, y1, x2, y2 = det_boxe[2],det_boxe[3],det_boxe[4],det_boxe[5]
                w = float(x2 - x1)
                h = float(y2 - y1)
                image_draw.draw_rectangle(int(x1) , int(y1) , int(w) , int(h) , color=color_three[det_boxe[0]],thickness=3)
                label = labels[det_boxe[0]]
                score = str(round(det_boxe[1],2))
                image_draw.draw_string_advanced( int(x1) , int(y1)-20,15, label + " " + score , color=color_three[det_boxe[0]])
                print("\n" + label + " " + score)
            image_draw.compress_for_ide()
            image_draw.save(root_path+"det_result.jpg")
        else:
            print("No objects were identified.")
        del results

        del ai2d
        del ai2d_builder
        del kpu
        gc.collect()
    print("---------------end------------------")
    nn.shrink_memory_pool()

if __name__=="__main__":
    nn.shrink_memory_pool()
    detection()
Wy0013206

感谢大佬，其实视频流在AncherFree模型下标签参数还是重合，但是换了AncherBase之后就行了

Walnutzz•

测试图片正常说明训练没有问题，导出模型时分辨率是224*224，那训练的分辨率呢？量化的参数选择的是什么，你把生成的部署包发过来wangyan01@canaan-creative.com

Wy0013206

训练分辨率640640，量化参数int8int8，NoClip

Walnutzz•

已发送，麻烦了

Walnutzz•

你试一下导出模型的时候不改变分辨率正常吗

Wy001•

你好，可以调整一下阈值试试，主要调整下面的三个参数：
image.png

Wy0013206

改完之后它显示的标签是错误的，与AI_Cube的推理结果不同。

Walnutzz•

@Wy001大佬能不能再看看，还有什么需要我提供的信息吗

Walnutzz•

还在找问题

Wy001•

det_result.jpg
image.png

Walnutzz3

@Wy001 调整阈值后，标签显示与AI_Cube评估测试结果差别很大

Walnutzz•

使用和训练一样的分辨率后各个标签置信度依然一样，修改阈值和多标签选项后标签依然是错误的@Wy001

Walnutzz•