机械臂项目2 不同数据集和模型效果比较

最新推荐文章于 2023-07-08 10:07:53 发布

_CoCoNutNut_

最新推荐文章于 2023-07-08 10:07:53 发布

阅读量1k

点赞数

本文链接：https://blog.csdn.net/weixin_42319408/article/details/97666586

版权

本文对比了不同数据集和模型在机械臂项目中的表现。使用了ssdlite_mobilenet_v2_coco、faster_rcnn_inception_v2_coco以及ssd_mobilenet_v2_coco模型，通过调整模型参数和数据集大小，观察loss的变化。实验结果显示，数据集的扩充和模型的选择对性能有显著影响。

摘要由CSDN通过智能技术生成

数据集

原图：100
扩充后：600
训练集：480
测试集：120

模型

ssdlite_mobilenet_v2_coco_2018_05_09

I0729 11:57:47.076869 4638664128 learning.py:507] global step 209: loss = 3.6843 (2.739 sec/step)
I0729 11:57:49.732738 123145492389888 supervisor.py:1117] Saving checkpoint to path training/model.ckpt

I0729 12:07:49.414452 4638664128 learning.py:507] global step 427: loss = 2.6105 (2.729 sec/step)
I0729 12:07:49.732372 123145492389888 supervisor.py:1117] Saving checkpoint to path training/model.ckpt

I0729 12:27:39.913542 4638664128 learning.py:507] global step 857: loss = 2.2021 (2.720 sec/step)
I0729 12:27:42.616056 4638664128 learning.py:507] global step 858: loss = 2.6953 (2.702 sec/step)
I0729 12:27:45.337074 4638664128 learning.py:507] global step 859: loss = 2.1534 (2.720 sec/step)
I0729 12:27:48.036956 4638664128 learning.py:507] global step 860: loss = 2.3363 (2.699 sec/step)
I0729 12:27:49.732676 123145492389888 supervisor.py:1117] Saving checkpoint to path training/model.ckpt

1. model.ckpt-427

loss=2.6105
threshold=0.1

2. model.ckpt-860

loss=2.3363
threshold=0.1

数据集

原图：100
扩充后：1200
训练集：960
测试集：240

模型

上一步中model.ckpt-860

I0729 14:13:53.092765 4595254720 learning.py:507] global step 193: loss = 2.4848 (2.997 sec/step)
I0729 14:13:53.362606 123145565872128 supervisor.py:1117] Saving checkpoint to path training/model.ckpt

I0729 14:23:51.449750 4595254720 learning.py:507] global step 390: loss = 2.6236 (2.975 sec/step)
I0729 14:23:53.363627 123145565872128 supervisor.py:1117] Saving checkpoint to path training/model.ckpt