keras模型预测只有一和零而不是概率

Question

这是我的模型和数据生成器，当我预测我得到 array([0., 0., 0., 1., 0., 0., 0., 0.] 而不是概率。我认为应该是概率。

model = keras.models.Sequential()
model.add ...
model.add(keras.layers.Dense(num_class))
model.add(keras.layers.Softmax())

sgd_opt = keras.optimizers.SGD(lr=0.001, momentum=0.9)
cce_loss = keras.losses.categorical_crossentropy

model.compile(optimizer=sgd_opt, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

train_generator = train_datagen.flow_from_directory(
train_data_dir,
target_size=IMAGE_SIZE,
batch_size=batch_size,
class_mode='sparse') 

validation_generator = train_datagen.flow_from_directory(
test_data_dir, # same directory as training data
target_size=IMAGE_SIZE,
batch_size=batch_size,
class_mode='sparse') 

Answer 1

看起来您的模型确实记住了样本的标签。
验证训练数据是不好的做法。尝试拆分数据集：

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(
  validation_split=0.2
)
train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=IMAGE_SIZE,
    batch_size=batch_size,
    subset='training',
    class_mode='sparse'
) 
validation_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=IMAGE_SIZE,
    batch_size=batch_size,
    subset='validation',
    class_mode='sparse')