PyTorch如何列出分類錯誤之原始圖片路徑？

參考資料

How to get the filename of a sample from a DataLoader?

前提

本文假設batch size使用1的情況下，將分類錯誤的原始圖片路徑、實際標籤、預測標籤寫入到missclassified_file_path.txt檔案當中。

因為每張圖片都必須要有路徑，PyTorch的DataLoader才能順利載入圖片，也因此，如果希望獲得圖片的路徑，修改DataLoader讓它能夠return圖片原始路徑即可。如程式碼的第10至14行

關鍵在程式碼的第51~58行

if args.batch_size==1 and pred_index.cpu() != target.cpu():
    with open("missclassified_file_path.txt", "a") as f:
        true_label = str(target.cpu().tolist()[0])
        pred_label = str(pred_index.cpu().tolist()[0])
        write_content = ' '.join(path)
        write_content = write_content + ' ' + true_label + ' ' + pred_label              
        f.write(write_content+"\n")

程式碼

根據txt文件中路徑，顯示圖片的預測 label及實際 label

import pandas as pd
import matplotlib.pyplot as plt
from pylab import *
df = pd.read_csv('../missclassified_file_path.txt', sep=" ", header=None)
df.columns = ["path", "true", "prediction", "component_name"]
df.loc[df['true'] == 0 , 'true'] = 'Good'
df.loc[df['true'] == 1 , 'true'] = 'Bad'
df.loc[df['prediction'] == 0 , 'prediction'] = 'Good'
df.loc[df['prediction'] == 1 , 'prediction'] = 'Bad'
data = df.loc[df['component_name'] == 'XXXXXX'][0::]
data_list = [path for path in data['path']]
    
predict_class = data['prediction'].tolist()
true_class = data['true'].tolist()
c_file_list = data['path'].tolist()
iclass = iter(true_class)
prediction = iter(predict_class)
def showImagesHorizontally(list_of_files):
    fig = plt.figure(figsize=(20, 10), dpi=80)
    number_of_files = len(list_of_files)
    for i in range(number_of_files):
        a=fig.add_subplot(5,9,i+1)
        plt.title(f"True: {next(iclass)}\n Prediction: {next(prediction)}")
        plt.tight_layout()
        image = imread(list_of_files[i])
        imshow(image,cmap='Greys_r')
        axis('off')
        
showImagesHorizontally(data_list)