使用 Vgg16 output 并添加额外的自定义层时出错。 ValueError：检查输入时出错

Question

我正在尝试使用include_top = false获取通过 Vgg16 网络的图像输出（训练和验证），然后添加最后几层，如下面的代码所示。

我希望x存储完整的 model 以便我可以从中创建一个 tflite 文件（包括 vgg 和我添加的层）

from tensorflow.keras.models import Model
import os

x= vgg16.output
print(x.shape)
x = GlobalAveragePooling2D()(x)

x = Flatten()(x)
x = Dense(100)(x)
x = tf.keras.layers.LeakyReLU(alpha=0.2)(x)
x = (Dropout(0.5)) (x)
x = (Dense(50)) (x) 
x = tf.keras.layers.LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
x = Dense(num_classes, activation='softmax')(x)


# this is the model we will train
model = Model(inputs=vgg16.input, outputs=x)

# first: train only the top layers (which were randomly initialized)
# i.e. freeze all convolutional InceptionV3 layers
for layer in vgg16.layers:
    layer.trainable = False

model.compile(loss='categorical_crossentropy',
   optimizer=optimizers.RMSprop(lr=1e-4),
   metrics=['acc'])

# train the model on the new data for a few epochs
history = model.fit(train_data, train_labels, 
   epochs=15,
   batch_size=batch_size,
   validation_data=(validation_data, validation_labels))

model.save(top_model_weights_path)
(eval_loss, eval_accuracy) = model.evaluate( 
    validation_data, validation_labels, batch_size=batch_size, verbose=1)

x.shape 的x.shape为 (?, ?, ?, 512)

train_data.shape (1660, 2, 2, 512)

train_labels.shape (1660, 4)

validation_data.shape数据形状 (137, 4)

validation_labels.shape标签形状（137、2、2、512）

错误：

ValueError：检查输入时出错：预期 input_3 的形状为 (None, None, 3) 但数组的形状为 (2, 2, 512)

这个错误发生在线路上：

52 验证数据=（验证数据，验证标签））

---

如下所示的先前代码片段可以正常工作并提供准确的输出。 train_data存储vgg16.predict_generator()的 numpy 数组

model = Sequential() 
model.add(Flatten(input_shape=train_data.shape[1:])) 
model.add(Dense(100)) 
model.add(tf.keras.layers.LeakyReLU(alpha=0.2))
model.add(Dropout(0.5)) 
model.add(Dense(50)) 
model.add(tf.keras.layers.LeakyReLU(alpha=0.3))
model.add(Dropout(0.3)) 
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss='categorical_crossentropy',
   optimizer=optimizers.RMSprop(lr=1e-4),
   metrics=['acc'])
history = model.fit(train_data, train_labels, 
   epochs=15,
   batch_size=batch_size,
   validation_data=(validation_data, validation_labels),
   callbacks =[tensorboard])
model.save(top_model_weights_path)
(eval_loss, eval_accuracy) = model.evaluate( 
    validation_data, validation_labels, batch_size=batch_size,     verbose=1)
print("[INFO] accuracy: {:.2f}%".format(eval_accuracy * 100)) 
print("[INFO] Loss: {}".format(eval_loss)) 

通过 vgg16 传递所有图像（训练、验证、测试；这里仅显示训练）的这一步是针对上述代码片段完成的

train_data_dir = 'data/train'
validation_data_dir = 'data/validation'
test_data_dir = 'data/test'

# number of epochs to train top model 
epochs = 7 #this has been changed after multiple model run 
# batch size used by flow_from_directory and predict_generator 
batch_size = 32

#Loading vgc16 model
vgg16 = applications.VGG16(include_top=False, weights='imagenet')
datagen = ImageDataGenerator(rescale=1. / 255) 
generator = datagen.flow_from_directory( 
    validation_data_dir, 
    target_size=(img_width, img_height), 
    batch_size=batch_size, 
    class_mode=None, 
    shuffle=False) 

nb_train_samples = len(generator.filenames) 
num_classes = len(generator.class_indices) 

predict_size_train = int(math.ceil(nb_train_samples / batch_size)) 

train_data = vgg16.predict_generator(generator, predict_size_train) 

Answer 1

出色地...

1. 您定义了target_size=(img_width, img_height)并且如果(img_width, img_height)不是(224, 224)那么您还需要在 VGG model 中定义target_size ：

vgg16 = applications.VGG16(
  include_top=False, 
  weights='imagenet',
  target_size=(img_width, img_height, 3))

2. 为什么在datagen.flow_from_directory中使用class_mode=None ？ None是默认值。 如果你想让它分类写class_mode='categorical' ，但使用class_mode=None根本没有意义。

3. predict_generator返回预测。 现在predict_generator已被弃用，但您可以使用predict哪个适用于生成器。 但是predict应该在训练后使用。 使用生成器的正确方法是：

datagen = ImageDataGenerator(rescale=1. / 255) 
generator = datagen.flow_from_directory( 
  train_data_dir, 
  target_size=(img_width, img_height), 
  batch_size=batch_size, 
  shuffle=False)
// ...
history = model.fit(
  generator,
  epochs=15,
  steps_per_epoch=len(generator), 
  batch_size=batch_size,
  validation_data=validation_generator,
  validation_steps=len(validation_generator))

稍后，如果您想进行预测，请使用： model.predict(test_generator)

4. 在这种情况下，您不需要在GlobalAveragePooling2D之后使用Flatten 。 GlobalAveragePooling2D将 output 减少为一维数组。