找到属于 2 个类的 0 个经过验证的图像文件名

Question

1 个数据集。 我正在使用目录中的 Keras Flow 来读取图像。

这是我的代码。

train_generator_flow = train_generator.flow_from_dataframe(df_train,
                                                           x_col='training_path', 
                                                           y_col='labels', 
                                                           class_mode='sparse', 
                                                           subset='training', 
                                                           target_size=(512, 512),
                                                           shuffle=True, 
                                                           seed=3407)

valid_generator_flow = valid_generator.flow_from_dataframe(df_valid, 
                                                           x_col='validation_path',
                                                           y_col='labels', 
                                                           class_mode='sparse', 
                                                           subset='validation', 
                                                           target_size=(512, 512),
                                                           seed=3407)

火车发电机和有效发电机是

train_generator = tf.keras.preprocessing.image.ImageDataGenerator(rotation_range=0.1, zoom_range=0.2) # very basic data augmentation

valid_generator = tf.keras.preprocessing.image.ImageDataGenerator() # no data augmentation for validation/testing

df_train和df_valid是

# let's wrap all the above stuff in a function, so that we can reuse all of that for testing 

def custom_data_loader(distribution='training', PATH = 'MURA-v1.1/', CSV_NAME = 'train_image_paths.csv'):
        df = pd.read_csv(os.path.join(PATH, CSV_NAME), dtype='str', header=None)
        df.columns = [f'{distribution}_path'] # rename first column
        df['category'] = df[f'{distribution}_path'].map(lambda x: x.split('/')[2])
        df['labels'] = df[f'{distribution}_path'].map(lambda x: 'positive' if 'positive' in x else 'negative')
        df['patient_id'] = df[f'{distribution}_path'].map(lambda x: x.split('/')[3])

        return df

df_train = custom_data_loader(distribution='training', PATH = 'MURA-v1.1/', CSV_NAME = 'train_image_paths.csv')

df_valid = custom_data_loader(distribution='validation', PATH = 'MURA-v1.1/', CSV_NAME = 'valid_image_paths.csv')

两个数据帧的头都是

df_train.head()

    training_path   category    labels  patient_id
0   MURA-v1.1/train/XR_SHOULDER/patient00001/study...   XR_SHOULDER positive    patient00001
1   MURA-v1.1/train/XR_SHOULDER/patient00001/study...   XR_SHOULDER positive    patient00001
2   MURA-v1.1/train/XR_SHOULDER/patient00001/study...   XR_SHOULDER positive    patient00001
3   MURA-v1.1/train/XR_SHOULDER/patient00002/study...   XR_SHOULDER positive    patient00002
4   MURA-v1.1/train/XR_SHOULDER/patient00002/study...   XR_SHOULDER positive    patient00002

df_valid.head()
validation_path category    labels  patient_id
0   MURA-v1.1/valid/XR_WRIST/patient11185/study1_p...   XR_WRIST    positive    patient11185
1   MURA-v1.1/valid/XR_WRIST/patient11185/study1_p...   XR_WRIST    positive    patient11185
2   MURA-v1.1/valid/XR_WRIST/patient11185/study1_p...   XR_WRIST    positive    patient11185
3   MURA-v1.1/valid/XR_WRIST/patient11185/study1_p...   XR_WRIST    positive    patient11185
4   MURA-v1.1/valid/XR_WRIST/patient11186/study1_p...   XR_WRIST    positive    patient11186

当我创建valid_generator_flow时，我得到以下输出。

Found 0 validated image filenames belonging to 2 classes.

对于训练来说

Found 36808 validated image filenames belonging to 2 classes.

除了作为子集正在训练的经过验证的单词之外，这似乎很好。 这是一个二元分类问题，我尝试过 class_mode='binary', class_mode='categorical', class_mode='sparse'（这个例子），但每次我得到 0 个经过验证的图像到 2 个类。

但是路径是正确的，因为我尝试使用 df_valid。

import cv2
img = cv2.imread(df_valid['validation_path'][0])

plt.imshow(img)

它工作正常并显示图像。

如果我打印类索引等，我会得到正确的输出。

a, b = next(train_generator_flow)
i,l = next(valid_generator_flow)
print("training input images patch shape  : ", a.shape)
print("training input labels patch shape  : ",b.shape)
print("training labels:",train_generator_flow.class_indices)
print ("________________________________________")
print("validation input images patch shape: ", a.shape)
print("validation input labels patch shape: ",b.shape)
print("validation labels:",valid_generator_flow.class_indices)
 

training input images patch shape  :  (32, 512, 512, 3)
training input labels patch shape  :  (32, 2)
training labels: {'negative': 0, 'positive': 1}
________________________________________
validation input images patch shape:  (32, 512, 512, 3)
validation input labels patch shape:  (32, 2)
validation labels: {'negative': 0, 'positive': 1}