

微调模型¶

本工程通过PyTorch库进行卷积网络训练，其微调实现参考迁移学习

预训练模型¶

PyTorch提供了AlexNet的预训练模型

python文件¶

微调实现：py/finetune.py
自定义微调数据集类：py/utils/data/custom_finetune_dataset.py
自定义批量采样器类：py/utils/data/custom_batch_sampler.py
辅助函数：py/utils/util.py

训练参数¶

批量处理：每次训练128个图像，其中32个正样本，96个负样本
输入模型图像：缩放到(227, 227)，随机水平翻转，进行归一化操作
优化器
- 使用SGD：学习率为1e-3，动量大小为0.9
- 随步长衰减：每隔7轮衰减一次，衰减因子为0.1
迭代次数：25轮

训练结果¶

$ python finetune.py 
...
...
Epoch 0/24
----------
train Loss: 0.1906 Acc: 0.9163
val Loss: 0.3662 Acc: 0.8692

Epoch 1/24
----------
train Loss: 0.1204 Acc: 0.9499
val Loss: 0.4081 Acc: 0.8701

Epoch 2/24
----------
train Loss: 0.0958 Acc: 0.9608
val Loss: 0.4002 Acc: 0.8719

Epoch 3/24
----------
train Loss: 0.0825 Acc: 0.9663
val Loss: 0.4505 Acc: 0.8725

Epoch 4/24
----------
train Loss: 0.0726 Acc: 0.9707
val Loss: 0.5031 Acc: 0.8697

Epoch 5/24
----------
train Loss: 0.0662 Acc: 0.9733
val Loss: 0.5340 Acc: 0.8681

Epoch 6/24
----------
train Loss: 0.0611 Acc: 0.9754
Qval Loss: 0.5102 Acc: 0.8714

Epoch 7/24
----------
train Loss: 0.0505 Acc: 0.9799
val Loss: 0.5529 Acc: 0.8725

Epoch 8/24
----------
train Loss: 0.0489 Acc: 0.9806
val Loss: 0.5540 Acc: 0.8728

Epoch 9/24
----------
train Loss: 0.0479 Acc: 0.9810
val Loss: 0.5818 Acc: 0.8717

Epoch 10/24
----------
train Loss: 0.0465 Acc: 0.9815
val Loss: 0.5819 Acc: 0.8727

Epoch 11/24
----------
train Loss: 0.0452 Acc: 0.9821
val Loss: 0.5765 Acc: 0.8732

Epoch 12/24
----------
train Loss: 0.0456 Acc: 0.9819
val Loss: 0.5957 Acc: 0.8725

Epoch 13/24
----------
train Loss: 0.0449 Acc: 0.9823
val Loss: 0.5857 Acc: 0.8724

Epoch 14/24
----------
train Loss: 0.0438 Acc: 0.9827
val Loss: 0.5943 Acc: 0.8723

Epoch 15/24
----------
train Loss: 0.0442 Acc: 0.9823
val Loss: 0.5874 Acc: 0.8730

Epoch 16/24
----------
train Loss: 0.0443 Acc: 0.9824
val Loss: 0.5950 Acc: 0.8720

Epoch 17/24
----------
train Loss: 0.0437 Acc: 0.9828
val Loss: 0.5945 Acc: 0.8729

Epoch 18/24
----------
train Loss: 0.0434 Acc: 0.9828
val Loss: 0.5975 Acc: 0.8725

Epoch 19/24
----------
train Loss: 0.0435 Acc: 0.9827
val Loss: 0.5900 Acc: 0.8730

Epoch 20/24
----------
train Loss: 0.0432 Acc: 0.9828
val Loss: 0.5922 Acc: 0.8730

Epoch 21/24
----------
train Loss: 0.0426 Acc: 0.9832
val Loss: 0.5977 Acc: 0.8720

Epoch 22/24
----------
train Loss: 0.0426 Acc: 0.9832
val Loss: 0.5974 Acc: 0.8726

Epoch 23/24
----------
train Loss: 0.0435 Acc: 0.9828
val Loss: 0.6019 Acc: 0.8718

Epoch 24/24
----------
train Loss: 0.0432 Acc: 0.9828
val Loss: 0.5996 Acc: 0.8718

Training complete in 461m 14s
Best val Acc: 0.873228

模型在训练过程中呈现过拟合现象，可考虑调整学习率、添加权重衰减以及更换优化器的方式：

学习率从1e-3调整为1e-4
添加L2权重衰减，衰减因子为1e-4
使用Adam替换SGD