overkill_leakage.py

# 針對每個元件計算其Overkill rate和Leakage rate
# 目前只適用於batch size = 1的情形
# 用來產生overkill和leakage數值的dataframe
import pandas as pd
import torch

test_df_mapping2_label = test_df.copy()                                          #複製一份要mapping到2個label的testing資料
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 0, 'class'] = 0    #將大於1的label轉成1
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 1, 'class'] = 1
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 2, 'class'] = 1
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 3, 'class'] = 1
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 4, 'class'] = 1
test_df_mapping2_label.loc[test_df_mapping2_label['class'] == 5, 'class'] = 1

name_of_each_component = test_df_mapping2_label['component_name'].value_counts().index.tolist()         # 取得每個元件的名稱
num_of_image_in_each_component = test_df_mapping2_label['component_name'].value_counts().values         # 取得每個元件的圖片總數量
test_component_name_df = pd.DataFrame(list(zip(name_of_each_component, num_of_image_in_each_component)), columns =['component_name', 'total'])   #產生包含所有元件名稱和總數的資料

for name in set(test_df_mapping2_label['component_name'].values):                                 # 對所有的元件名稱
    temp_data = test_df_mapping2_label.loc[(test_df_mapping2_label["component_name"] == name)]    # 選取出符合指定元件名稱的資料(如R0402)
    for k, v in zip(temp_data['class'].value_counts().keys(), temp_data['class'].value_counts()): # 取得該元件下的k(class名稱)及v(class中樣本的數量)
        if k == 0:  #如果k=0，即Good label，新增新欄位good，給予在該元件名稱下其label為Good的樣本數量
            test_component_name_df.loc[test_component_name_df['component_name'] == name, 'good'] = temp_data['class'].value_counts().sort_index().values[0]
        elif k ==1: #如果k=1，即Bad label，新增新欄位bad，給予在該元件名稱下其label為bad的樣本數量
            try:    # 通常來說第0個為Good的資料、第1個為Bad的資料；但有可能有些元件只有Bad的資料，資料只會在第0個出現，這時就會執行except的部份
                test_component_name_df.loc[test_component_name_df['component_name'] == name, 'bad'] = temp_data['class'].value_counts().sort_index().values[1]
            except:
                print(f"{name} only contains bad label.")
                test_component_name_df.loc[test_component_name_df['component_name'] == name, 'bad'] = temp_data['class'].value_counts().sort_index().values[0]
test_component_name_df['good'] = test_component_name_df['good'].fillna(0).astype(int)          # 將NAN值填0並轉成INT格式
test_component_name_df['bad'] = test_component_name_df['bad'].fillna(0).astype(int)            # 將NAN值填0並轉成INT格式
test_component_name_df = test_component_name_df[['component_name', 'total', 'good', 'bad']]    # 欄位排序
col = {'overkill': 0, 'leakage': 0}                                                            # 新增overkill和leakage的預設值資料
test_component_name_df = test_component_name_df.assign(**col)                                  # 使用assign將所有的overkill和leakage欄位都填上0


def testing(test_component_name_df, 略, .........):
    略.....
    preds = torch.tensor([(lambda i: 1 if i > 1 else i)(i) for i in preds]).cuda()    # 將預測結果大於1的轉成1
    target = torch.tensor([(lambda i: 1 if i > 1 else i)(i) for i in target]).cuda()  # 將label大於1的轉成1
    略.....
    if args.batch_size==1:                                        # batch size等於1的情況下才執行以下動作，省去處理元件名稱的問題，1筆資料對1筆資料
        if target.data == 0 and target.data == preds.data:        # True Positive的情況
            tp += 1
        elif target.data == 0 and target.data != preds.data:      # False Negative(Overkill)的情況
            fn += 1
            test_component_name_df.loc[(test_component_name_df["component_name"] == ''.join(component_name)), 'overkill'] +=1    #找該batch下的元件名稱資料，並在overkill中加1
        elif target.data == 1 and target.data == preds.data:      # True Negative的情況
            tn += 1
        elif target.data == 1 and target.data != preds.data:      # False Postive(Leakage)的情況
            fp += 1
            test_component_name_df.loc[(test_component_name_df["component_name"] == ''.join(component_name)), 'leakage'] +=1     #找該batch下的元件名稱資料，並在leakage中加1
    return test_component_name_df  #回傳蒐集到的資料

# 計算overkill rate和leakage rate、取到下數後2位、加上百分比符號
test_component_name_df['overkill_rate'] = (test_component_name_df['overkill'] / test_component_name_df['total'] * 100).round(decimals = 2).astype(str) + '%'
test_component_name_df['leakage_rate'] = (test_component_name_df['leakage'] / test_component_name_df['total'] * 100).round(decimals = 2).astype(str) + '%'

# 計算整體測試樣本最終之overkill rate和leakage rate
test_component_name_df = test_component_name_df.append({'total':sum(test_component_name_df['total']),'good':sum(test_component_name_df['good']),'bad':sum(test_component_name_df['bad']),'overkill':sum(test_component_name_df['overkill']), 'leakage':sum(test_component_name_df['leakage']), 'overkill_rate':sum(test_component_name_df['overkill'])/sum(test_component_name_df['total']), 'leakage_rate': sum(test_component_name_df['leakage'])/sum(test_component_name_df['total'])}, ignore_index=True)

# 將資料輸出成Excel格式
test_component_name_df.to_excel('./overkill_and_leakage.xlsx', index=False)