# 2022 / 09 / 20 進度

###### tags: `實驗`
[TOC]

## 上次的建議
1. Oversample重寫
-> 效果差不多(沒什麼必要)
2. Overlap 256x512 -> 280 x 512
-> performance沒有提升(甚至有點下降趨向)
3. 重新嘗試classification 按照病人切
-> 0.855是極限?

## 新的問題

## 實驗 ( Bone-scan ) Classification
| Model 50Epoch | AP 50 | AP 50-95 | AR 50-95 | AP50 rm hard case
| -------- | -------- | -------- | -------- | -------- |
|Faster Whole images no big box |0.2431|0.0790|0.2359|0.3740|
| FCOS Whole images no big box |0.2270|0.0790|0.2388|0.3790|
| Faster Whole images big box |0.3060|0.1486|0.3513|0.3650|
| FCOS Whole images big box |0.3236|0.1679|0.4015|0.3650|

base line (valid acc) : 0.855
嘗試各種加channel的方法/model大小/加黑圖或原圖/channel數/balance data/
極限(嘗試過加1channel和3channel/沒有前一張就用全黑圖代替):
ResNet18 ~= ResNet34 > ResNet50(需要降batch size)
不平衡data 0.8
平衡data 0.82
| 設定 | Best Acc |
| -------- | -------- |
| 不平衡data| 0.8|
| 每次隨機平衡data 400x400|0.8684 |
| 加上前一張圖片 400x400 | 0.9079 |
| Patch 256x512 | 0.8947 |

Patch
max val aupr 0.9073496240601504, saved at epoch 72
0.8947368421052632

Normal 300x300
Epoch: 48 Valid acc: 0.881578947368421, valid auc: 0.8232142857142857, valid pr 0.8896552822734803

Normal 400x400
0.8553

Add Last IMG 400x400
Epoch: 68 Valid acc: 0.9079, valid auc: 0.9214285714285714, valid pr 0.9543
Best Acc 0.9079

| 設定 | Best AUC |
| -------- | -------- |
| Unblance data| 0.8204 |
| Random balance data 400x400|0.8607 |
| Whole Img + 8Patch 256x512 | 0.8232 |
| Last Img feature mix 400x400 | 0.8839 |
| Add Last Img FC 400x400 | <font color=red>0.9214 |

## 其他

Next
1. confusion matrix
2. 改pos weight比例
3. k-folder
4. 想怎麼利用前後張圖片，key point、兩張圖對比
5. Patch classification

1. 改input channel(增加for two images)
2. kfold

老蘇的建議:
1. classification先做+channel看結果有沒有變好
2. object detection + Dynamic Head
3. 改label
4. 魔改head

補實驗
1. Faster-RCNN with crop
2. Faster-Rcnn with patch?
3. YOLO v7 with crop and 3class