毕设周记

最新推荐文章于 2021-09-18 22:08:17 发布

marsggbo

最新推荐文章于 2021-09-18 22:08:17 发布

阅读量4.7k

点赞数 2

CC 4.0 BY-SA版权

分类专栏： python 机器学习

本文链接：https://blog.youkuaiyun.com/marsggbo/article/details/79774499

本文记录了2018年3月31日作者在毕设中使用Pytorch在SIW-13数据集上测试vgg16_bn、vgg16_bn.features+SPP+vgg16_bn.classification和ResNet50模型的过程。训练、验证和测试集的准确度详细列出，显示出过拟合问题。过拟合可能源于网络结构复杂度和数据不平衡，建议降低网络复杂度或采用数据增强策略。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

周记实验记录

时间：2018年3月31日

地点：寝室
运行环境：Pytorch，实验室服务器，SIW-13数据集

这一周我使用了如下几种方法对数据集进行测试，结果分别如下：

1.vgg16_bn

直接将每张图片缩放到固定大小224*224，之后传入vgg16模型中进行训练。

其中在训练过程中，当准确率分别达到85%，90%，95%以上时都会对模型进行保存，以便于对后面的验证集和测试集进行准确度检验。

参数设置：LR=1e-05, dropout=0.4, EPOCH=8, batch_size=20

最终损失值及准确度:

1）训练集

Loss:0.0088, Train Accuracy:97.1970%

Epoch: 0
----------
Step 99, Train Loss:0.1219, Train Accuracy:18.4500%, file=f
Step 199, Train Loss:0.1155, Train Accuracy:25.1750%, file=f
Step 299, Train Loss:0.1075, Train Accuracy:32.2833%, file=f
Step 399, Train Loss:0.1003, Train Accuracy:37.9750%, file=f
Loss:0.0987, Train Accuracy:39.2887%, file=f
Training time is:0m 43s
__________


Epoch: 1
----------
Step 99, Train Loss:0.0643, Train Accuracy:64.6000%, file=f
Step 199, Train Loss:0.0614, Train Accuracy:66.0500%, file=f
Step 299, Train Loss:0.0580, Train Accuracy:67.7333%, file=f
Step 399, Train Loss:0.0556, Train Accuracy:68.8625%, file=f
Loss:0.0552, Train Accuracy:69.0260%, file=f
Training time is:0m 43s
__________


Epoch: 2
----------
Step 99, Train Loss:0.0419, Train Accuracy:78.0000%, file=f
Step 199, Train Loss:0.0408, Train Accuracy:78.1250%, file=f
Step 299, Train Loss:0.0396, Train Accuracy:78.7667%, file=f
Step 399, Train Loss:0.0383, Train Accuracy:79.5500%, file=f
Loss:0.0380, Train Accuracy:79.7197%, file=f
Training time is:0m 43s
__________


Epoch: 3
----------
Step 99, Train Loss:0.0307, Train Accuracy:84.0000%, file=f
Step 199, Train Loss:0.0301, Train Accuracy:84.3000%, file=f
Step 299, Train Loss:0.0292, Train Accuracy:84.6333%, file=f
Step 399, Train Loss:0.0283, Train Accuracy:85.0250%, file=f
save vgg16bn_0.85_params.pkl
Loss:0.0282, Train Accuracy:85.1136%, file=f
Training time is:0m 47s
__________


Epoch: 4
----------
Step 99, Train Loss:0.0230, Train Accuracy:89.1500%, file=f
Step 199, Train Loss:0.0225, Train Accuracy:89.2250%, file=f
Step 299, Train Loss:0.0219, Train Accuracy:89.5000%, file=f
Step 399, Train Loss:0.0213, Train Accuracy:89.7875%, file=f
save vgg16bn_0.85_params.pkl
Loss:0.0212, Train Accuracy:89.8599%, file=f
Training time is:0m 46s
__________


Epoch: 5
----------
Step 99, Train Loss:0.0172, Train Accuracy:92.2000%, file=f
Step 199, Train Loss:0.0171, Train Accuracy:91.8500%, file=f
Step 299, Train Loss:0.0165, Train Accuracy:92.4833%, file=f
Step 399, Train Loss:0.0160, Train Accuracy:92.812