# 03/20/25 Meeting Notes #26
# This Week's Progress 這週進度
## Classification model #2
- Dataset count after adding augmentation to `Bohemian and Hippie` dataset. (Due to insufficient dataset count) `增強Bohemian and Hippie 後的資料集數量(之前的資料集數量不足)`:

Full dataset:
    1570 files in bohemian_and_hippie
    1560 files in streetwear
    1572 files in sporty
    1527 files in formal
    1586 files in casual
    1703 files in semi-formal
total: 9518

Split dataset: 
    train dataset: 
        1099 files in bohemian_and_hippie
        1092 files in streetwear
        1100 files in sporty
        1068 files in formal
        1110 files in casual
        1192 files in semi-formal
    total: 6661

    val dataset: 
        235 files in bohemian_and_hippie
        234 files in streetwear
        235 files in sporty
        229 files in formal
        237 files in casual
        255 files in semi-formal
    total: 1425

    test dataset: 
        236 files in bohemian_and_hippie
        234 files in streetwear
        237 files in sporty
        230 files in formal
        239 files in casual
        256 files in semi-formal
    total: 1432

- With learning rate `學習率` 0.0003
- With learning rate `學習率` 0.0001
- Difference`差異` : learning rate `學習率` 。Higher accuracy has lower learning rate. `更高的準確度對應較低的學習率`
- 0.0003 lowered to 0.0001, because with 0.0003, the graph fluctuate too much (first graph). `從 0.0003 調整到 0.0001，因為 0.0003 會導致圖表波動過大（見第一張圖）`
- But even with 0.001, it still fluctuate, so more tuning is needed. `即使使用 0.001，波動仍然存在，因此需要進一步調整`
- Not only fluctuate, also over-fitting. `僅波動，還有過擬合的問題`
- With learning rate `學習率` 0.0003
- With learning rate `學習率` 0.0001
- Can see that the accuracy only increased by a little. `可以看到準確度只略微提高`
---
- We tried training the model without `Bohemian and Hippie` dataset. 我們嘗試在沒有「Bohemian and Hippie」數據集的情況下訓練模型。
- Dataset count `數據集計數`:

full dataset:
    1560 files in streetwear
    1572 files in sporty
    1527 files in formal
    1586 files in casual
    1703 files in semi-formal
total: 7948

Split dataset:
    train dataset: 
        1092 files in streetwear
        1100 files in sporty
        1068 files in formal
        1110 files in casual
        1192 files in semi-formal
    total: 5562

    val dataset: 
        234 files in streetwear
        235 files in sporty
        229 files in formal
        237 files in casual
        255 files in semi-formal
    total: 1190

    test dataset: 
        234 files in streetwear
        237 files in sporty
        230 files in formal
        239 files in casual
        256 files in semi-formal
    total: 1196

- Without `bohemian and hippie` dataset, with learning rate 0.0001. `在沒有「bohemian and hippie」數據集的情況下，使用學習率 0.0001。`
- The training accuracy doesn't increases much, so we suppose its the model that is the problem. Will try to tune the model and try again. `訓練精準度沒有增加太多，所以我們認為問題出在模型。將嘗試調整模型並重試`

## Generative model
- landmarks
- 1000 iterations (regular)
- 3000 iterations (trial)
- Problem:
I thought increasing the number of iterations would improve the output, but it actually made it worse. The landmarks in the output don’t match their positions in the reference image. This is the problem, and I haven’t found a solution yet. `我本來以為增加迭代次數會改善輸出，但結果反而變得更糟。輸出中的標記點與參考圖片中的位置不匹配。這就是問題所在，但我還沒找到解決辦法`

## Frontend
- The wardrobe screens are mostly done, but they’re not connected to the database yet and have no API. Because of this, they don’t work as intended. `衣櫥界面大致完成了，但還沒有連接到資料庫，也沒有 API。因此，它的運作方式與預期不同。`

## Backend
- Currently having problems with the environment. Will try to solve it as soon as possible. `目前遇到環境問題，會盡快嘗試解決`

# To Do 需做
- High-Accuracy Clothing and Style Classification via Multi-Feature Fusion is the paper for model 2, the accuracy is 69.66%. Needa explain to prof that the accuracy is higher than the original paper. May also needa explain why we don't use the other paper.
https://link.springer.com/article/10.1007/s10489-024-05683-9/tables/3
explain why the accuracy is higher blah blah

Meeting:
- A demo for the prof (everything). (No need to be complete by next week, but needa have a lil smtg)
- Every input and output of each model.

# Next Weeks's Meeting 下週會議

---
Previous:
Next:
Full Content List [here](https://hackmd.io/@emps-113up/full-list)