Keras 的稀疏分类交叉熵形状问题

Question

I have a multiclass problem where an image can be one of three classes (Masked, UnMasked, Hybrid).

I am using image_dataset_from_directory from keras preprocessing module which makes things easier.

Load dataset

def load_from_directory(shuffle=False):
  train_ds = tfk.preprocessing.image_dataset_from_directory(
    directory=TRAINING_PATH,
    image_size=IMAGE_SIZE,
    validation_split=VALIDATION_SPLIT,
    batch_size=BATCH_SIZE,
    seed=SEED,
    subset='training',
    label_mode='int',
    shuffle=shuffle
  )

  val_ds = tfk.preprocessing.image_dataset_from_directory(
    directory=TRAINING_PATH,
    image_size=IMAGE_SIZE,
    validation_split=VALIDATION_SPLIT,
    batch_size=BATCH_SIZE,
    seed=SEED,
    subset='validation',
    label_mode='int',
    shuffle=False
  )

  test_ds = tfk.preprocessing.image_dataset_from_directory(
    directory=TESTING_PATH,
    labels=None,
    image_size=IMAGE_SIZE,
    batch_size=BATCH_SIZE,
    seed=SEED,
    label_mode='int',
    shuffle=False
  )

  return train_ds, val_ds, test_ds

On the keras documentation it states for the label_model the following: 'int': means that the labels are encoded as integers (eg for sparse_categorical_crossentropy loss) ... ( here is the link). Since my folder has the following structure:

• MaskDataset
  • training
    • 0/
      • img1, img2, …, imgN
    • 1/
      • img1, img2, …, imgN
    • 2/
      • img1, img2, …, imgN
  • testing
    • img1, img2, …, imgN

And for the model by final layers are the following:

  x = tfkl.SeparableConv2D(1024, 3, padding='same')(x)
  x = tfkl.BatchNormalization()(x)
  x = tfkl.Activation('relu')(x)  

  # GlobalAveragePooling + Dropout
  x = tfkl.GlobalAveragePooling2D()(x)
  x = tfkl.Dropout(0.5)(x) 
  
  # Softmax
  outputs = tfkl.Dense(units=len(CLASS_NAMES), activation='softmax')(x)

  model = tfk.Model(inputs, outputs)

which I get the following complain: "ValueError: Shapes (None, 3) and (None, 1) are incompatible"

If I change the label_mode to 'categorical', when loading the dataset and change the loss function from SparseCategoricalCrossEntropy to CategoricalCrossEntropy and the accuracy from SparseCategoricalAccuracy to CategoricalAccuracy it works, but I really wanted to understand why I am not able to use the SparseCrossEntropy loss function and how to fix it.

Edit: Add model compile and fit

def compile_model(model, plot=False):
  model.compile(
    optimizer=tf.optimizers.Adam(1e-3),
    loss=tf.losses.SparseCategoricalCrossentropy(name='loss'),
    metrics=[
      tfk.metrics.SparseCategoricalAccuracy(name='accuracy'), 
      tfk.metrics.Precision(name='precision'), 
      tfk.metrics.Recall(name='recall'), 
    ]
  )

  model.summary()
  if plot: tfk.utils.plot_model(model, show_shapes=True)

def train_model(model, debug_mode=False):
  callbacks = [tfk.callbacks.EarlyStopping(patience=5, monitor='val_loss', restore_best_weights=True)]
  if debug_mode:
    callbacks.append(tfk.callbacks.ModelCheckpoint(filepath='model.{epoch:02d}-{val_loss:.2f}.h5', save_best_only=True, monitor='val_loss'))
    callbacks.append(tfk.callbacks.TensorBoard(log_dir='./tensorboard'))

  history = model.fit(
    x=train_ds, 
    validation_data=val_ds,
    epochs=100, 
    callbacks=callbacks,
    # steps_per_epoch=len(train_ds),
    # validation_steps=len(val_ds),
  )

  return history

Answer 1

tf.keras.metrics.SparseCategoricalAccuracy() --> is for SparseCategorical (int) class. tf.keras.metrics.Recall() --> is for categorical (one-hot) class.

You have to use a one-hot class if you want to use any metric naming without the 'Sparse'.

update:

num_class=4
def get_img_and_onehot_class(img_path, class):
    img = tf.io.read_file(img_path)
    img = tf.io.decode_jpeg(img, channels=3)
    """ Other preprocessing of image."""
    return img, tf.one_hot(class, num_class)

when you got the one-hot class:

loss=tf.losses.CategoricalCrossentropy
METRICS=[tf.keras.metrics.CategoricalAccuracy(name='accuracy'),
            tf.keras.metrics.Precision(name='precision'),
            tf.keras.metrics.Recall(name='recall'),]

model.compile(
        optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001),
        loss=loss,
        metrics= METRICS)

Ref: #72217176