如何在 TF2.1 中保存/加载用于微调/迁移学习的部分模型？

Question

我想建立自己的基础模型并使用大数据集对其进行训练。训练后，我保存了基础模型。我有另一个自定义模型，我想从基础模型中加载前两层的权重。我应该如何在 Tensorflow 2.1.0 中实现它，谢谢。

示例代码：

import os
os.environ["CUDA_VISIBLE_DEVICES"]="" 
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

class BaseModel():
    def __init__(self):
        inputs = keras.Input(shape=(32, 32, 3))
        x = inputs
        x = layers.Conv2D(32, 3, padding='same', activation=tf.nn.relu)(x)
        x = layers.MaxPool2D()(x)
        x = layers.Conv2D(64, 3, padding='same', activation=tf.nn.relu)(x)

        x = layers.Flatten()(x)

        x = layers.Dense(500, activation=tf.nn.relu)(x)

        outputs = layers.Dense(1000, activation=tf.nn.softmax)(x)

        self.model = keras.Model(inputs=inputs, outputs=outputs)

    def __call__(self, inputs):
        return self.model(inputs)

bm = BaseModel()  # the model for pretraining
bm.model.save_weights('base_model') # save the pretrained model


class MyModel():
    def __init__(self):
        inputs = keras.Input(shape=(32, 32, 3))
        x = inputs
        x = layers.Conv2D(32, 3, padding='same', activation=tf.nn.relu)(x)
        x = layers.MaxPool2D()(x)
        x = layers.Conv2D(64, 3, padding='same', activation=tf.nn.relu)(x)

        x = layers.Conv2D(128, 3, padding='same', activation=tf.nn.relu)(x)

        x = layers.Flatten()(x)

        x = layers.Dense(1000, activation=tf.nn.relu)(x)

        outputs = layers.Dense(10, activation=tf.nn.softmax)(x)

        self.model = keras.Model(inputs=inputs, outputs=outputs)

    def __call__(self, inputs):
        return self.model(inputs)


mm = MyModel()  # the model for my customized applications
mm.model.load_weights('base_model')  # load the pretrained model with the first two conv layers

# further fine-tuning or transfer learning 

Answer 1

我使用了 TF 网站上提供的自定义模型来演示这个想法。使用子类模型与 Keras Sequential 和函数模型几乎没有什么不同。我已经使用子类模型来演示这个想法，如下所示。

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

基本模型

class ThreeLayerMLP(keras.Model):

  def __init__(self, name=None):
    super(ThreeLayerMLP, self).__init__(name=name)
    self.dense_1 = layers.Dense(64, activation='relu', name='dense_1')
    self.dense_2 = layers.Dense(64, activation='relu', name='dense_2')
    self.pred_layer = layers.Dense(10, name='predictions')

  def call(self, inputs):
    x = self.dense_1(inputs)
    x = self.dense_2(x)
    return self.pred_layer(x)

def get_model():
  return ThreeLayerMLP(name='3_layer_mlp')

base_model = get_model()

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape(60000, 784).astype('float32') / 255
x_test = x_test.reshape(10000, 784).astype('float32') / 255

base_model.compile(loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              optimizer=keras.optimizers.RMSprop())
history = base_model.fit(x_train, y_train,
                    batch_size=64,
                    epochs=1)
#saving weights
base_model.save_weights('./base_model_weights', save_format='tf')

自定义模型

class MyCustomModel(keras.Model):

  def __init__(self, name=None):
    super(MyCustomModel, self).__init__(name=name)
    self.dense_1 = layers.Dense(64, activation='relu', name='dense_1')
    self.dense_2 = layers.Dense(64, activation='relu', name='dense_2')
    self.pred_layer = layers.Dense(10, name='predictions')

  def call(self, inputs):
    x = self.dense_1(inputs)
    x = self.dense_2(x)
    return self.pred_layer(x)

def get_custom_model():
  return MyCustomModel(name='my_custom_model')

my_custom_model = get_custom_model()

my_custom_model.compile(loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  optimizer=keras.optimizers.RMSprop())

当保存的模型在模型中有任何custom_objects或自定义层时，以下步骤是导入

# This initializes the variables used by the optimizers,
# as well as any stateful metric variables
my_custom_model.train_on_batch(x_train[:1], y_train[:1])

# Load the state of the old model (to load weights for all layers)
# my_custom_model.load_weights('path_to_my_weights')

layer_dict = dict([(layer.name, layer) for layer in base_model.layers])
print(layer_dict)

# my_custom_model.trainable = True
# loading the weights from base_model
for layer in my_custom_model.layers:
  layer_name = layer.name
  #print(layer.name)
  layer.set_weights(layer_dict[layer_name].get_weights())