数据集二分类 第一类1000张, 第二类600张
1. darknet + resnet50 Loss,训练出来测试的时候是NULL, 暂时不知道为什么, 将CUDA-10.0 换成 cuda-8.0 依然显示不出label Loss 也是3.8
2. darknet + alexnet Loss = 3.8 ,
3. keras + inceptionV3
1.) OpenCV4.0 调用失败, 提示FusedBatchNorm is_learning = True
2) https://github.com/opencv/opencv/issues/14236 修改代码,修复bug后,出现下面bug DataType = DT_BOOL 不被识别
3) 需要更新到最新版的opencv
4. keras + inception_resnetV2 同上
6. tensorflow + slim + inceptionV4 验证集准确率90%以上 , 但是OpenCV4.0 调用失败, slim 居然把DecodeJpeg层都加到 pb模型里去了, 暂时不知道怎么修改模型
1) 修改为
opencv 报错 :
2)修改为
opencv ---in 'Mul'
3) 修改为
opencv --- 错误同上
4) modify as
call in opencv , error shows as above
5)
sh 脚本提示不能使用Cast节点作为输入 , 必须用Placeholder , 于是下面更改input
6)
opencv Same as above
7)
Error
8)
fail in ‘Mul’ as well
9)
fail in ‘Mul’ as well
10)
fail in ‘Mul’ as well
7. 准确率比较
keras + mobilenetv2 + 224X224 val_acc = 91% epoch=20
keras + mobilenetv2 + 448X448 val_acc = 97% epoch=40
keras + mobilenetv2 + 640X360 val_acc = 79% epoch=45 and val_acc = 74% epoch=20