ImageDataGenerator 预测类 - 为什么预测不能正确地从概率转换为预测类？

Question

我有一个这样设置的目录：

图片

-- val
    --class1
    --class2
-- test
   --all_classes
-- train
    --class1
    --class2

每个目录中都有一组图像。我想预测测试中的每个图像是属于 1 类还是 2 类。

我写这个是为了读入训练和验证数据：

train_path = "/content/drive/train/"
valid_path = "/content/drive/val/"

train_datagen = ImageDataGenerator(
    rescale=1./255,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True)

test_datagen = ImageDataGenerator(rescale=1./255)

train_generator=train_datagen.flow_from_directory(
  directory=train_path,
  batch_size=32,
  class_mode='binary',
  target_size=(150,150)
)

validation_generator=test_datagen.flow_from_directory(
  directory=valid_path,
  batch_size=32,
  class_mode='binary',
  target_size=(150,150)
)

创建了一个网络：

def create_network(): 
  model = Sequential()
  model.add(Input(shape=(150,150,3)))

  model.add(Conv2D(32, kernel_size=3,strides=(1, 1),activation='relu', padding='valid', dilation_rate=1))
  model.add(MaxPooling2D(pool_size=(2, 2)))

  model.add(Conv2D(64, kernel_size=3, strides=(1, 1), activation='relu',padding='valid', dilation_rate=1))
  model.add(MaxPooling2D(pool_size=(2, 2)))

  model.add(Flatten())
  model.add(Dense(512, activation='relu'))

  model.add(Dense(1, activation='sigmoid'))
  plot_model(model, to_file='/content/drive/question1_model.png', show_shapes=True, show_layer_names=True)

  model.compile(optimizer = 'adam',
                   loss = 'binary_crossentropy', 
                   metrics = ['accuracy'])
  return model

拟合模型：

def fit_model(train_generator=train_generator, validation_generator=validation_generator,network=create_network()):
  checkpoint_path = "/content/drive/question1_checkpoint.h5"
  checkpoint_dir = os.path.dirname(checkpoint_path)

  callbacks_list = [
      callbacks.EarlyStopping(
          monitor = 'accuracy',
          patience = 5,
      ),

      callbacks.ModelCheckpoint(
          filepath=checkpoint_path,
          monitor = 'val_loss',
          #save_weights_only=True,
          save_best_only=True,
      ),

  ]

  model = network
  history = model.fit(train_generator,
                      epochs=200,
                      validation_data=validation_generator,
                      batch_size=32, 
                      callbacks = callbacks_list,
                      verbose=1
                      )
  return history,model,time_taken

history,model = fit_model(train_generator,validation_generator)

模型的accuracy和val准确率>80%，我重新加载进去预测：

model = load_model('/content/drive/question1_checkpoint.h5')

然后我想预测测试目录中的一组图像：

test_datagen = ImageDataGenerator(rescale=1./255)
test_path = "/content/drive/test/"

test_generator = test_datagen.flow_from_directory(
  directory=test_path,
  batch_size=16,
  class_mode='binary',
  target_size=(150,150),
  shuffle = False
)
test_generator.reset()
filenames = test_generator.filenames
nb_samples = len(filenames)
batch_size=16
predict = model.predict(test_generator,steps=test_generator.n/batch_size)

当我打印 predict 的开始时，我可以看到：

[[6.09035552e-01]
 [2.47541070e-02]
 [7.37663209e-02]
 [5.22839129e-02]
 [2.94408262e-01]
 [1.39171720e-01]
 [6.15863085e-01]

我认为这给了我第 1 类正确的概率。但是当我打印每个预测的类时：

predicted_class_indices=np.argmax(predict,axis=-1)
print(predicted_class_indices)

输出是：

[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0]

这意味着我的预测概率没有正确地转换为类，对吧？因为例如 2.47541070e-02 是 0.02，而 6.09035552e-01 是 0.60，所以不应该预测这些属于不同的类别吗？谁能告诉我哪里出错了？

Answer 1

这段代码给了我下面的输出：（在用相同的二进制类数据集训练模型后，我已经拍摄了 10 张图像进行测试 - 5 张狗的图像，5 张猫的图像）。

nb_samples = len(filenames)
batch_size=5
predict = model.predict(test_generator,steps=test_generator.n/batch_size)
predict

输出：

array([[0.06690815],
       [0.7787118 ],
       [0.109512  ],
       [0.39706784],
       [0.07243159],
       [0.61042166],
       [0.5808931 ],
       [0.86361384],
       [0.9961897 ],
       [0.61571515]], dtype=float32)

这是正确的，因为您在最后一层使用了sigmoid 激活函数，那么输出范围将从 0 到 1。

请不要将argmax 用于sigmoid 值。您可以使用argmax 方法，同时使用softmax 激活函数来查找其他类概率中类的最高概率值。

这里你可以使用下面的代码：

import tensorflow as tf

predictions = tf.where(predict <= 0.5, 0, 1)

print('Predictions:\n', predictions.numpy())

输出：

Predictions:

 [[0]
 [1]
 [0]
 [0]
 [0]
 [1]
 [1]
 [1]
 [1]
 [1]]