# 11/07/24 Meeting Notes #12
# This Week's Progress 這週進度
## Classification Model #1
- No progress

## Classification Model #2
- Added validation dataset (20% of training dataset)
`已新增驗證數據集（佔訓練數據集的 20%）`
- Improved the model`改進模型`:
-- Channel Bottlenecking in Attention Modules
`在注意力模塊中加入通道瓶頸`
-- Added layer normalization
`添加層正規化`
-- Data augmentation
`數據增強`

- Running the improved model with 50 epochs

- Running on a pretrained model (https://github.com/facebookresearch/ConvNeXt?tab=readme-ov-file)
`目前正在使用預訓練模型進行運行`

## Generative Model
- Finished setting docker container, but still need to download cuda for WSL
`完成 Docker 容器的設置，但仍需下載 CUDA給WSL`

- Started looking for model for adding landmark to the clothes
`已開始尋找模型來為衣物添加標記點`
- Need to contact Style3D (module for implementing the mesh into the body)
`需要聯繫 Style3D（用於將網格實現於身體的模組）`
- https://studio.style3d.com/en/home

## Datasets
- Made a python script to divide the datasets to different tags (provided by authors). The authors provided two types of labels (noisy and cleaned).
`製作了一個 Python 腳本來將數據集劃分至不同的標籤（由作者提供）。作者提供了兩種標籤（噪音標籤和清理後的標籤）`
> -- Noisy labels refer to labels that may contain inaccuracies or ambiguities, as they are the initial tags and haven’t been thoroughly verified. 噪音標籤指的是可能包含不準確或模糊的標籤，這些是初始標籤，尚未經過全面驗證。
> -- Cleaned labels refer to labels that have been reviewed and corrected to be more accurate, with ambiguities clarified.清理後的標籤指的是已審核並更正的標籤，標籤更加準確且已澄清模糊之處
- Total number of tags: 66, but we only chose 29
`標籤總數：66，但我們僅選擇了 29 個標籤`
> Top, T-Shirt, Blouse, Shirt, Cardigan, Cape, Blazer, Sweatshirt, Vest, Sweater, Jacket, Dress, Coat, Skirt, Pants, Jeans, Shorts, Jumper, Romper, Leggings, Shoes, Sandals, Boots, Pumps, Flats, Loafers, Sneakers, Tie, Swimwear

-- Purpose目的: Dividing the dataset into these specific tags simplifies dataset management and allows us to organize the images more effectively.
`將數據集劃分為這些特定標籤，有助於簡化數據集管理，並讓圖像組織更加有效`
-- The script also took out duplicates.
`該腳本也去除了重複數據`

-- Casual 95
-- Formal 36
-- Semi Formal 80
-- Sporty 0
-- Vintage 6

New dataset found:
- trends vogue Computer Vision Project (2.1k)
<https://universe.roboflow.com/fashion-jywg7/trends-vogue>
- Runway vogue Computer Vision Project (655)
<https://universe.roboflow.com/fashion-jywg7/runway-vogue>

The dataset are smaller in number but newer and more up-to-date to today's style.
Can add it into the current one we have (fashion550k)

# To Do 需做
- Before the 13th, need ppt ready for professor
- Prof. might ask questions about the papers, so read and understand them.
- 摘要，研究動機與研究問題，文獻回顧與探討，研究方法與步驟，預期結果，需要指導教授指導內容，產考文獻（citation）
- 摘要：用什麼技術，解決什麼問題，爲什麼要做什麼。（basically to do summary lah, keywords and all).
- 研究動機與研究問題：Motive and how we solve them.
- Abstract, research motivation and research questions, literature review and discussion, research methods and steps, expected results, content required for guidance from the professor, production examination documents (citation)
If we want to do this, need to tell TAs as prof ned to prep documents.
大專生國科會計劃 (National Science Council Program for College Students)
—　not sure that our identity will affect.

# Next Next Weeks's Meeting 下週會議
- Train classification model #1 with actual datasets
- Generative (FULLY finish setting up envirionment)
---
Previous: [10/31/24 Meeting Notes #11](https://hackmd.io/@emps-113up/meeting11)
Next: [11/21/24 Meeting Notes #13](https://hackmd.io/@emps-113up/meeting13)
Full Content List [here](https://hackmd.io/@emps-113up/full-list)