知一組再生晶圓瑕疵特徵圖(Reclaim wafer defective feature map)之
資料(如附件)，請以 Keras API 建構有效之 MLP 及 CN 分類器 (Classifier)，將此資料供分類成所需之十種分類，必須以一系列實驗(DOE)決定訓練(及預測)之相關參數，以期得到最佳分類及預測效果，並以 words 完成相關報告；使用 Python 請依以下敘述，逐步完成所需之功能 (請依功能分題完成，以利評分；另，take-home exam 為榮譽考試，請勿討論， 共 110 分，最高以 100 計)
python程式代寫需求：
- 建立一 databalancing 函式(or method): 將 10 類資料讀入系統中(各類數量不同)，並將各類資料的數量複製到最多數量那一個 class 的數量(只要接近即可)，同時要執行shuffle and split 後 回 傳 train_feature_list 、 train_target_list 、 test_feature_list 及
test_target_list，分別列印於螢幕上；最後請確認資料讀入正確，並分別儲存成 feature 及
target 四個 txt 檔案，以利後續讀取 (以自建之函式，其他方式不給分，請同時要 scale
成 [0, 1]的區間資料) (10%)
- 比較 MLP之 data balacing: 建構 MLP 之各種架構之 model，分別比較有無 data balancing 之訓練與測試之結果(Note:自行判定並決定停止條件，必須記錄所有訓練、測試、model 之數據及過程，包括: train acc, val acc, train loss, val loss, model structure, model, confusion matrix，並以 excel 繪製圖形做比較，以數據說明 data balacing 是否有提升訓練效果， 並提供最佳 model 及其 acc、loss 等資訊) (20%)
- 比較 CNN之 data balacing: 建構 CNN 之各種架構之 model，分別比較有無 data balancing
之訓練與測試之結果(Note:同上) (20%)
- 比較 MLP之 data augment: 呈上題，建構 MLP 之各種架構 model，比較有無 data augment 對分類之影像，data augment 必須包括:上下 flipping、左右 flipping、旋轉 45 度
(或更多角度)，需求與前二題相同(同樣必須提供訓練、測試過程、最佳 model、辨識率等資訊，並告知 data augment 是否能提升辨識率)(20%)
- 比較 CNN之 data augment: 同上題，但使用 CNN (20%)
- 確認性實驗: 將上述所得到之 MLP、CNN最佳網路，做重複性實驗 (各 run 三次以上)， 比較 MLP 和 CNN 對於該數據之平均辨識能力之比較，做個人之最後結論 (結論包括: 整體實驗結果，並評論 data balancing、data augment 對 MLP 及 CNN 訓練之影響；MLP 及 CNN 之比較等)(20%)
圖表建議格式: Note:
| 網路 | 網路
架構 |
訓練參數 | Training
辨識率 /loss |
Testing
辨識率 /loss |
整體辨識率 | 訓練圖形 |
| MLP 1 | “max_iter”: 3000, “steps”: 10, “structure”:(5), “activation”:’tanh’,”solver”: ‘lbfgs’,”verbose”:False, … | …… | ….. | …. |
 |
|
| MLP 2 | ||||||
| MLP 3 | ||||||
| …. | ||||||
| …. |
Note: 請使用一方式可以清楚表示網路之架構，如:
model_C8(K3_ReLU_MP3)_F32(ReLu)_F2(SM)，並說明表示之方式
The Python program writes a set of Reclaim wafer defective feature maps.
Information (such as attachments), please use the Keras API to construct a valid MLP and CN Classifier
A set of known flaw characteristics in FIG reconstituted wafer (Reclaim wafer defective feature map) of
Information ( such as accessories ) , please Keras API construct valid MLP and CN classifier (. Classifier) , this information for classification into the required ten kinds of classification, must be based on a series of experiments (DOE) decided to train ( and predict ) the Relevant parameters, in order to get the best classification and prediction results, and complete the relevant report with words ; use Python , according to the following description, gradually complete the required functions ( please complete the function according to the function, to benefit the score; another, take-home exam to honor the exam, do not discuss, a total of 110 points, up to 100 dollars )
Python program write requirements:
- Create a databalancing function (or method): Read 10 types of data into the system (various types), and copy the number of types of data to the maximum number of classes (as long as they are close), and at the same time After performing shuffle and split, return train_feature_list , train_target_list , test_feature_list , and
Test_target_list, which is printed on the screen; please confirm that the data is correctly read and stored as feature and
Target Four txt files for subsequent reading (in self-built function, other methods do not give points, please scale at the same time
Interval data of [0, 1] (10%)
- Comparing MLP data balacing: Constructing models of various architectures of MLP, comparing the results of training and testing with data balancing respectively (Note: self-determination and decision to stop conditions, all training, testing, model data and processes must be recorded, including: Train acc, val acc, train loss, val loss, model structure, model, confusion matrix, and draw graphics with excel for comparison, to show whether data balacing has improved training effect, and provide the best model and its acc, loss, etc. Information) (20%)
- Compare CNN’s data balacing: Construct models of CNN’s various architectures, compare and compare data balancing
Training and testing results (Note: ibid.) (20%)
- Compare MLP data augment: Submit the above questions, construct various architecture models of MLP, compare data augment to classified images, data augment must include: up and down flapping, left and right flipping, rotation 45 degrees
(or more), the requirements are the same as the first two questions (the same training, test process, best model, recognition rate, etc. must be provided, and data augment can be raised to improve the recognition rate) (20%)
- Compare CNN’s data augment: Same as above, but use CNN (20%)
- Confirmatory experiment: The above-mentioned MLP and CNN optimal network were repetitive experiments (three times each run), and the comparison of the average discriminating ability of MLP and CNN for the data was compared, and the final conclusion of the individual was made (conclusion includes : Overall experimental results, and comment on the impact of data balancing, data augment on MLP and CNN training; comparison of MLP and CNN, etc. (20%)
Chart suggested format: Note:
| network | networkArchitecture | Training parameter | TrainingIdentification rate
/loss |
TestingIdentification rate
/loss |
Overall recognition rate | Training graphics |
| MLP 1 | “max_iter”: 3000, “steps”: 10, “structure”: (5), “activation”: ‘tanh’, “solver”: ‘lbfgs’, “verbose”: False, … | …… | ….. | …. |
 |
|
| MLP 2 | ||||||
| MLP 3 | ||||||
| …. | ||||||
| …. |
Note: Please use a way to clearly indicate the architecture of the network, such as:
model_C8(K3_ReLU_MP3)_F32(ReLu)_F2(SM), and explain how it is represented
