如何复制 tf.keras.models.Model 子类？

我需要复制 keras 模型，但我无法知道可以做什么，除非该模型is not a tf.keras.models.Model()子类。

Note:使用copy.deepcopy()将在没有任何错误的情况下工作，但是每当使用副本时都会导致另一个错误。

Example:

import tensorflow as tf

class MyModel(tf.keras.Model):

  def __init__(self):
    super(MyModel, self).__init__()
    self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu)
    self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax)
    self.dropout = tf.keras.layers.Dropout(0.5)

  def call(self, inputs, training=False):
    x = self.dense1(inputs)
    if training:
      x = self.dropout(x, training=training)
    return self.dense2(x)


if __name__ == '__main__':
    model1 = MyModel()
    model2 = tf.keras.models.clone_model(model1)

结果是：

Traceback (most recent call last):
  File "/Users/emadboctor/Library/Application Support/JetBrains/PyCharm2020.3/scratches/scratch.py", line 600, in <module>
    model2 = tf.keras.models.clone_model(model1)
  File "/usr/local/lib/python3.8/site-packages/tensorflow/python/keras/models.py", line 430, in clone_model
    return _clone_functional_model(
  File "/usr/local/lib/python3.8/site-packages/tensorflow/python/keras/models.py", line 171, in _clone_functional_model
    raise ValueError('Expected `model` argument '
ValueError: Expected `model` argument to be a functional `Model` instance, but got a subclass model instead.

目前，我们还不能使用tf.keras.models.clone_model对于子类化模型 API，而我们可以对于顺序 API 和函数式 API。从doc https://www.tensorflow.org/api_docs/python/tf/keras/models/clone_model,

model   Instance of Model (could be a functional model or a Sequential model).

这是满足您需求的解决方法。如果我们需要复制经过训练的模型，我们可以在其中获得一些优化的参数，这是有意义的。因此，主要任务是我们需要通过复制现有模型来创建新模型。目前这种情况最方便的方法是get训练有素的体重和set到新创建的模型实例。首先构建一个模型，对其进行训练，然后获取权重矩阵并将其设置到新模型。

import tensorflow as tf
import numpy as np

class ModelSubClassing(tf.keras.Model):
    def __init__(self, num_classes):
        super(ModelSubClassing, self).__init__()
        self.conv1 = tf.keras.layers.Conv2D(32, 3, strides=2, activation="relu")
        self.gap   = tf.keras.layers.GlobalAveragePooling2D()
        self.dense = tf.keras.layers.Dense(num_classes)

    def call(self, input_tensor, training=False):
        # forward pass: block 1 
        x = self.conv1(input_tensor)
        x = self.gap(x)
        return self.dense(x)

    def build_graph(self, raw_shape):
        x = tf.keras.layers.Input(shape=raw_shape)
        return tf.keras.Model(inputs=[x], outputs=self.call(x))


# compile 
sub_classing_model = ModelSubClassing(10)
sub_classing_model.compile(
        loss      = tf.keras.losses.CategoricalCrossentropy(),
        metrics   = tf.keras.metrics.CategoricalAccuracy(),
        optimizer = tf.keras.optimizers.Adam())

# plot for debug 
tf.keras.utils.plot_model(
    sub_classing_model.build_graph(x_train.shape[1:]),
    show_shapes=False,
    show_dtype=False,
    show_layer_names=True,
    expand_nested=False,
    dpi=96,
)

DataSet

(x_train, y_train), (_, _) = tf.keras.datasets.mnist.load_data()

# train set / data 
x_train = np.expand_dims(x_train, axis=-1)
x_train = x_train.astype('float32') / 255
# train set / target 
y_train = tf.keras.utils.to_categorical(y_train, num_classes=10)

# fit 
sub_classing_model.fit(x_train, y_train, batch_size=128, epochs=1)
# 469/469 [==============================] - 2s 2ms/step - loss: 8.2821

新模型/副本

对于子类模型，我们必须启动类对象。

sub_classing_model_copy = ModelSubClassing(10)
sub_classing_model_copy.build((x_train.shape))
sub_classing_model_copy.set_weights(sub_classing_model.get_weights()) # <- get and set wg

# plot for debug ; same as original plot
# but know, layer name is no longer same
# i.e. if, old: conv2d_40 , new/copy: conv2d_41
tf.keras.utils.plot_model(
    sub_classing_model_copy.build_graph(x_train.shape[1:]),
    show_shapes=False,
    show_dtype=False,
    show_layer_names=True,
    expand_nested=False,
    dpi=96,
)