多分类脚本结果明细.md

多分类——173类 （8650：50：50）

建议每个高度：训练集1K = 10*121（张）（训练集数据增强）
12000--8700 可能有一些类中微球没有高度区别，未离开光阱

回归——将代码改为回归问题 改dataloader 模型减小 目标检测算法

训练数据分类结果： 1. 没有训练集数据增广Train = 6*173;无transform数据增广; batch = 8; 层：256,128,240

Best valid accuracy: epoch 78 
train accuracy：53.5645% 
valid accuracy：15.3179% 
2. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,240

Best valid Accuracy：epoch 104
Train Accuracy: 52.1195%, 
Valid Accuracy: 39.5954%,
3. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,128,64,240

Best valid Accuracy：epoch 528
Train Accuracy: 25.7225%
Valid Accuracy: 48.5549%
4. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,240

Best valid Accuracy：epoch 546
Train Accuracy: 29.86%
Valid Accuracy: 45.08%
5. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,32,16,240

Best valid Accuracy：epoch 325
Train Accuracy: 2.1%
Valid Accuracy: 3.4%
6. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=15),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,32,173

Best valid Accuracy：epoch  2335
Train Accuracy: 20.13%
Valid Accuracy: 41.9%
7. 数据集增广 train = 8*121*173
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,173

Best valid Accuracy：epoch 15
Train Accuracy: 57.85%
Valid Accuracy: 61.27%
8. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,32,173

Best valid Accuracy：epoch 967
Train Accuracy: 28.6%
Valid Accuracy: 45.08%
9. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,173
LEARNING RATE =1e-4

Best valid Accuracy：epoch 967
Train Accuracy: 36.12%
Valid Accuracy: 41.04%
10. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,173
LEARNING RATE =1e-3

Best valid Accuracy：epoch 
Train Accuracy: 0.578%？
Valid Accuracy: 0.578%
11. 没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,173
LEARNING RATE =3e-4

Best valid Accuracy：epoch 123
Train Accuracy: 10.79%
Valid Accuracy: 25.72%
12. 数据集增广 train = 8*121*173
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 32; 层：256,128,64,173
LEARNING RATE =1e-4
Resnet101

Best valid Accuracy：epoch 
Train Accuracy: %
Valid Accuracy: %
13. resnet101 小样本数据
没有训练集数据增广Train = 6*173;
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 16; 层：256,173
learning r = 1e-4

Best valid Accuracy：epoch 135
Train Accuracy: 38%
Valid Accuracy: 39.31%
14. 默认学习率
15. 数据集增广 train = 8*121*173
transforms.RandomResizedCrop(size=256, scale=(0.8, 1.0)),
transforms.RandomRotation(degrees=45),
transforms.RandomHorizontalFlip()；
batch = 16;层：256,128,64,173；drop out = 0.4；learning rate = 1e-4

Best valid Accuracy：epoch 29
Train Accuracy: 61.06%
Valid Accuracy: 60.78%