# 2022 / 12 / 28 進度

###### tags: `實驗`
[TOC]

## 實驗結果
valid都是第6個fold
<font size=5, color='blue'>一張圖片</font>
| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
|單純一張圖片|0.2153|||
|一張圖片+放大bbox|0.3565|||
|一張圖片+放大bbox+Deformable|0.3585|||
|一張圖片+放大bbox+DIoU Deformable|0.3614|||

<font size=5, color='blue'>加上另一張圖片(Special)</font>
| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
|Baseline兩張原圖直接0.5+0.5|0.2627|0.3564|0.3867|
|兩張原圖直接0.8+0.2|0.4087|0.4769|0.4530|
|整張圖registration|<b>0.4514|<b>0.5151|0.4885|
|一半一半registration|0.4363|0.4955|<b>0.5045|

<font size=5, color='blue'>參考時間資訊(Temporal)</font>
| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
| Baseline LSTM 1/log(e+天數) |0.4038|0.4993| 0.4529|
| 1/t沒有間隔限制(左右都registration) |<b>0.4582|0.5101|0.4606|
| e沒有間隔限制(左右都registration) |0.4439|<b>0.5286|<b>0.5286|

<font size=5, color='blue'>只參考特定時間內的資訊(Time constrant)</font>
| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
|Baseline LSTM 1/log(e+天數)|0.4038|<b>0.4993|0.4529|
|天數>=200不考慮|<b>0.4262|0.4849|<b>0.4852|
|天數>=300不考慮|0.4257|0.4639|0.4564|
|天數>=365不考慮|0.3764|0.4433|0.4630|
|天數>=400不考慮|0.3904|0.4337|0.4798|

===============================================
<font size=5, color='blue'>AI自己學兩張圖片的比重(Attention)</font>

| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|half|0.4991|0.4320|0.3053|0.5858|0.4048|0.2095|
|whole|0.5270|0.4900|0.3956|0.5823|0.4920|0.2269|

| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
|Attention whole|<b>0.4887|<b>0.5305|<b>0.5108|
|Attention half|0.4285|0.4932|0.4821|
|Attention whole with time||||

Attention whole

Patch No attention : 0.4729
Patch attention 8:2 : 0.4576

發現使用Pseudo label使performance變好 但是參考上一張圖片的情況消失了(1/0) 之前0.488的那個是(0.95:0.05)

3.Pseudo label (att weight : 1/0)

| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|best|0.5349|0.5009|0.4154|0.5238|0.5635|0.1976|

[0.4399432715312221, 0.5320373957854486, 0.6326539170600971]

| Name | mAP50 | mAP30 | mAR50:95 |
| -------- | -------- | -------- | -------- |
|best Pseudo label|0.5009|0.5349|0.5075|
|Lastest Pseudo label|0.4926|0.5373|0.5044|

* 嘗試用<b>DeformFPN</b> +　Attention + Residual Whole
(att weight : 1/1e-20)

| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|best |0.5910|0.5312|0.4659|0.5976|0.5301|0.0.2773|

mAP 0.30:0.95 = 0.362

[0.5164648309276482, 0.5976064066156258, 0.6586510863301804]

* 嘗試用<b>DeformAtt</b> + Attention + Residual <b>Whole</b>
(att weight : 1/1e-13)

| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|best |0.5547|0.5134|0.4074|0.6208|0.5121|0.2683|

[0.4411639431210541, 0.6284983532771768, 0.5943222475821961]

* 嘗試用<b>DeformFPN</b> + Attention + Residual <b>Patch</b>
(att weight : 1/1e-25)

| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|best |0.6030|0.5528|0.5621|0.5413|0.5550*|0.1961|

Dual Flow Module
| Name | mAP30 | mAP50 | Benign | Equivocal | Malignant| mAP75 |
| -------- | -------- | -------- | -------- | -------- | -------- | -------- |
|best |0.5427|0.4685|0.3159|0.5413|0.5620|0.2213|

[0.40640543675439, 0.5706984574186, 0.6509455636143]

1.然後要嘗試一下兩張圖片的augmentation (1+2/1+1/2+2/2+1)

4.可能切patch的時候就已經有誤差了 導致registration對的不好，可能要先做registration，不過這樣等於要存每張圖片以及上一張做registration的圖

2.Unsupervised data label

whole registration 改回來看看

|half regis|0.4991|0.4320|0.3053|0.5858|0.4048|0.2095|
|whole regis|0.5270|0.4900|0.3956|0.5823|0.4920|0.2269|