I have a CNN model and using it to predict the class of an image:

model = load_model(modelName)
model.compile(loss='binary_crossentropy',
              optimizer='rmsprop',
              metrics=['accuracy'])

img = image.load_img(filename1, target_size=(img_width, img_height), color_mode="grayscale")
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)

images = np.vstack([x])


predict = model.predict(images, batch_size=8) 
classes = np.argmax(predict, axis=1)

print(classes)

Output: [25]. I have 36 folders, each with a different name (https://imgur.com/a/rJdEmfJ), full of images that I used to train the CNN and I didn't label them in preprocessing. (class_names were left as comment, shown below):

ds_train = tf.keras.preprocessing.image_dataset_from_directory(
    directory = 'D:/dataset2/',
    labels = 'inferred',
    label_mode = 'int', 
    # class_names=['0', '1', '2', '3', ...]
    color_mode = 'grayscale',
    batch_size = batchSize,
    image_size = (imgHeight, imgWidth),
    shuffle = True,
    seed = 123,
    validation_split = 0.2,
    subset = "training",
)

How do I know which class is which, does it sort the folders in some way or what?

CodePudding user response：

If you set labels = 'inferred' and do not specify class_names, then the ordering of the classes is alphanumeric. If you specify class_names then the order of the classes will be the order of the list of class_names.

For example, I have a directory that contains 30 subdirectories and each of these subdirectories contains image files of musical instruments. In the code below train_data is a data set where the class_names are not specified so the order will be alphanumeric.

The dataset reverse_train_data is created by first listing the content of the main directory, then using the python function sorted with reverse=True to get a reversed alphanumeric list, and then setting class_names equal to that reversed list.

The print-out in the code shows the resultant order of the classes for each case. Note you can get the class name order using class_names=train_data.class_names

train_dir=r'C:\Temp\instruments\train'
classlist=os.listdir(train_dir)# Note per python documentation list_dir returns an arbitrary ordered list
sorted_classlist=sorted(classlist, reverse=True) # this is a list in reverse alphanumeric order
train_data=tf.keras.utils.image_dataset_from_directory(train_dir, labels='inferred', label_mode='categorical',  class_names=None,
                                                        color_mode='rgb', batch_size=32,  image_size=(224,224), shuffle=False,
                                                        seed=None,  validation_split=None, subset=None,    interpolation='bilinear',
                                                        follow_links=False,   crop_to_aspect_ratio=False)
reverse_train_data=tf.keras.utils.image_dataset_from_directory(train_dir, labels='inferred', label_mode='categorical',  class_names=sorted_classlist,
                                                        color_mode='rgb', batch_size=32,  image_size=(224,224), shuffle=False,
                                                        seed=None,  validation_split=None, subset=None,    interpolation='bilinear',
                                                        follow_links=False,   crop_to_aspect_ratio=False)
class_names=train_data.class_names
reverse_class_names=reverse_train_data.class_names
print('{0:^25s}{1:^25s}{2:^25s}'.format('CLASS NAMES', 'REVERSE CLASS NAMES', 'SORTED CLASS LIST'))
for i in range (len(class_names)):
    print('{0:^25s}{1:^25s}{2:^25s}'.format(class_names[i], reverse_class_names[i], sorted_classlist[i]))

The result of the printout is shown below

Found 4793 files belonging to 30 classes.
Found 4793 files belonging to 30 classes.
       CLASS NAMES          REVERSE CLASS NAMES       SORTED CLASS LIST    
       Didgeridoo                 violin                   violin          
       Tambourine                  tuba                     tuba           
        Xylophone                 trumpet                  trumpet         
        acordian                 trombone                 trombone         
         alphorn                steel drum               steel drum        
        bagpipes                   sitar                    sitar          
          banjo                  saxaphone                saxaphone        
       bongo drum                  piano                    piano          
         casaba                   ocarina                  ocarina         
        castanets                 marakas                  marakas         
        clarinet                   harp                     harp           
       clavichord                harmonica                harmonica        
       concertina                 guitar                   guitar          
          drums                    guiro                    guiro          
        dulcimer                   flute                    flute          
          flute                  dulcimer                 dulcimer         
          guiro                    drums                    drums          
         guitar                 concertina               concertina        
        harmonica               clavichord               clavichord        
          harp                   clarinet                 clarinet         
         marakas                 castanets                castanets        
         ocarina                  casaba                   casaba          
          piano                 bongo drum               bongo drum        
        saxaphone                  banjo                    banjo          
          sitar                  bagpipes                 bagpipes         
       steel drum                 alphorn                  alphorn         
        trombone                 acordian                 acordian         
         trumpet                 Xylophone                Xylophone        
          tuba                  Tambourine               Tambourine        
         violin                 Didgeridoo               Didgeridoo