# 2022 / 11 / 16 進度 ###### tags: `實驗` [TOC] ### 嘗試K FOLD 發現原本的Validation資料因為有篩選過所以表現很好(0.6) 按照病人切5fold看看結果(雖然會有資料不均的情況) 上下浮動蠻大的,目前最好的方法AP50約0.34~0.42代表很吃資料 1. 切10FOLD會再好一點(Best : 0.492) 2. 可以明顯看到多加入一張圖片的Feature會變好 3. Validation overlap要想怎麼篩選 4. 實際把圖片都印出來看結果 5. 去掉所有大的bounding box反而效果不好但髖關節其實都抓不到,很喜歡框頭和右腳 6. 下次嘗試去掉box太密集地看看 7. Psuedo labels 8. 把Noise和Blur也加進去 validation有往上的趨勢了 9. 加入Deformable conv 10. 跑個一個兩層的Deformable conv 100 Epoch大概0.48就收斂了 ## 實驗 | - | Lowest mAP50 | Highest mAP50 | Lowest mAP30 | Highest mAP30 | | -------- | -------- | -------- | -------- | -------- | | 256overlap_5Fold | 0.27 | 0.35 | 0.32 | 0.41 | | Pair 256overlap_5Fold | 0.31 | 0.41 | 0.347 | 0.465 | | Pair 256overlap_10Fold | 0.42 | 0.484 | 0.464 | 0.554 | | Pair Deformable 256overlap_10Fold | - | 0.492 | - | 0.546| | Pair 2Deformable 256overlap_10Fold | - | 0.487 | - | 0.576 | | Pair Deform 256overlap_10Fold reverse | - | 0.507 | - | 0.571 | ## 實際Validation output
×
Sign in
Email
Password
Forgot password
or
By clicking below, you agree to our
terms of service
.
Sign in via Facebook
Sign in via Twitter
Sign in via GitHub
Sign in via Dropbox
Sign in with Wallet
Wallet (
)
Connect another wallet
New to HackMD?
Sign up