PaddlePaddle / PaddleDetection

Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.
Apache License 2.0
12.87k stars 2.9k forks source link

yolov3_mobilenetv3训练自己数据集损失一直降不下来 #7255

Open jiajunwang420 opened 2 years ago

jiajunwang420 commented 2 years ago

问题确认 Search before asking

请提出你的问题 Please ask your question

yolov3_mobilenetv3训练自己数据集损失一直降不下来,数据集大概6800张,两个类别,anchors重新计算过了(貌似没啥影响),但是损失一直在30左右,做一下预测,效果很差,基本上原数据集的图都框不出来, architecture: YOLOv3 use_gpu: true max_iters: 500000 log_iter: 20 save_dir: output snapshot_iter: 10000 metric: COCO pretrain_weights: weights/mobel_fire.pdparams weights: output/yolov3_mobilenet_v1_fire/best_model num_classes: 2 use_fine_grained_loss: false

YOLOv3: backbone: MobileNetV3 yolo_head: YOLOv3Head

MobileNetV3: norm_type: sync_bn norm_decay: 0. model_name: large scale: 1. extra_block_filters: [] feature_maps: [1, 2, 3, 4, 6]

YOLOv3Head: anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]] anchors: [[24, 33], [52, 69], [73, 139], [148, 84], [99, 265], [167, 163], [341, 137], [221, 305], [446, 385]] norm_decay: 0. yolo_loss: YOLOv3Loss nms: background_label: -1 keep_top_k: 100 nms_threshold: 0.45 nms_top_k: 1000 normalized: false score_threshold: 0.01

YOLOv3Loss: ignore_thresh: 0.7 label_smooth: false

LearningRate: base_lr: 0.00001 schedulers:

OptimizerBuilder: optimizer: momentum: 0.9 type: Momentum regularizer: factor: 0.0005 type: L2

READER: 'yolov3_reader.yml'

lyuwenyu commented 2 years ago

其他模型能跑出来嘛

jiajunwang420 commented 2 years ago

ppyolo可以跑出来,但是ppyolo无法部署在edgeboard上面,

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2022年11月3日(星期四) 下午2:52 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [PaddlePaddle/PaddleDetection] yolov3_mobilenetv3训练自己数据集损失一直降不下来 (Issue #7255)

其他模型能跑出来嘛

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

lyuwenyu commented 2 years ago
  1. ppyoloe或者picodet呢 这些能部署嘛
  2. 上边那个你有加载在coco的预训练参数没
jiajunwang420 commented 2 years ago

edgeboard好像是只支持yolov3系列的, 已经加载了预训练权重

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2022年11月3日(星期四) 下午3:16 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [PaddlePaddle/PaddleDetection] yolov3_mobilenetv3训练自己数据集损失一直降不下来 (Issue #7255)

picodet或者ppyoloe呢 这些能部署吧

上边那个你有加载在coco的预训练参数没

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

lyuwenyu commented 2 years ago
image

你这是修改了嘛,有试过默认的匹配能跑出来嘛

jiajunwang420 commented 2 years ago

你的意思是不修改anchor吗?

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2022年11月3日(星期四) 下午3:18 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [PaddlePaddle/PaddleDetection] yolov3_mobilenetv3训练自己数据集损失一直降不下来 (Issue #7255)

你这是修改了嘛,有试过默认的匹配能跑出来嘛

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

lyuwenyu commented 2 years ago

不不 1. 你这个featmap和anchor对不上吧 2. neck那个位置怎么没了

jiajunwang420 commented 2 years ago

不不 1. 你这个featmap和anchor对不上吧 2. neck那个位置怎么没了

我看官方给的yolov3_Mobilenetv3.yml里面的参数设置是这样的: image

jiajunwang420 commented 2 years ago
image

你这是修改了嘛,有试过默认的匹配能跑出来嘛

我昨天晚上用默认的yml配置跑了一遍,结果还是差不多,评估的准确率却特别地低 image