丢失 function 与 image_dataset_from_directory 一起使用

Question

I am confused about the appropriate loss function to use as i am generating my dataset using image_dataset_from_directory.

Data Generator

Train

train_ds = tf.keras.utils.image_dataset_from_directory(
  '/content/dataset/train',
  validation_split=0.05,
  subset="training",
  seed=123,
  image_size=(IMAGE_SIZE, IMAGE_SIZE),
  batch_size=BATCH_SIZE)

Validation

val_ds =  tf.keras.preprocessing.image_dataset_from_directory(
    '/content/dataset/val', image_size=(IMAGE_SIZE, IMAGE_SIZE), batch_size=BATCH_SIZE
)

Model

rn50v2_model = Sequential()

pretrained_model = tf.keras.applications.ResNet50V2(
    include_top=False,
    weights="imagenet",
    input_shape=(IMAGE_SIZE, IMAGE_SIZE, 3),
    pooling='avg',
    classes = 2
)

print(pretrained_model.summary())

rn50v2_model.add(pretrained_model)

rn50v2_model.add(Flatten())

rn50v2_model.add(Dense(512, activation='relu'))

rn50v2_model.add(Dense(2, activation='softmax'))

#print(rn50v2_model.summary())

rn50v2_model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

when i tested my model, i got my result one hot encoded like below:

array([[0.24823777, 0.7517622 ]], dtype=float32)

I would prefer to use categorical_crossentropy but pls explain this behaviour i cant seem to find information on the official documentation

Answer 1

Your result is not one hot encoded. It is the result of your output layer squished by the softmax function. This function transforms all values between 0 and 1 and these values sum to 1. You can apply np.argmax(predictions, axis=-1) to these "probabilities" to get the corresponding class. To use categorical_crossentropy , try changing your label_mode to categorical , which will automatically generated one-hot-encoded labels:

train_ds = tf.keras.utils.image_dataset_from_directory(
  '/content/dataset/train',
  validation_split=0.05,
  subset="training",
  seed=123,
  label_mode = 'categorical',
  image_size=(IMAGE_SIZE, IMAGE_SIZE),
  batch_size=BATCH_SIZE)

val_ds =  tf.keras.preprocessing.image_dataset_from_directory(
    '/content/dataset/val', 
     subset="validation",
     seed=123, 
     label_mode = 'categorical',
     image_size=(IMAGE_SIZE, IMAGE_SIZE), 
     batch_size=BATCH_SIZE)

You could also consider using binary_crossentropy if you only have two classes. You would have to change your loss function and output layer:

rn50v2_model.add(Dense(1, activation='sigmoid'))