以自定义张量作为变量的 TensorFlow 2.0 Keras 层

TensorFlow 2.0 Keras layers with custom tensors as variables

在 TF 1.x 中,可以使用自定义变量构建图层。这是一个例子:

import numpy as np
import tensorflow as tf

def make_custom_getter(custom_variables):
    def custom_getter(getter, name, **kwargs):
        if name in custom_variables:
            variable = custom_variables[name]
        else:
            variable = getter(name, **kwargs)
        return variable
    return custom_getter

# Make a custom getter for the dense layer variables.
# Note: custom variables can result from arbitrary computation;
#       for the sake of this example, we make them just constant tensors.
custom_variables = {
    "model/dense/kernel": tf.constant(
        np.random.rand(784, 64), name="custom_kernel", dtype=tf.float32),
    "model/dense/bias": tf.constant(
        np.random.rand(64), name="custom_bias", dtype=tf.float32),
}
custom_getter = make_custom_getter(custom_variables)

# Compute hiddens using a dense layer with custom variables.
x = tf.random.normal(shape=(1, 784), name="inputs")
with tf.variable_scope("model", custom_getter=custom_getter):
    Layer = tf.layers.Dense(64)
    hiddens = Layer(x)

print(Layer.variables)

构造的致密层的打印变量将是我们在custom_variables dict中指定的自定义张量:

[<tf.Tensor 'custom_kernel:0' shape=(784, 64) dtype=float32>, <tf.Tensor 'custom_bias:0' shape=(64,) dtype=float32>]

这允许我们创建 layers/models 直接使用 custom_variables 中提供的张量作为它们的权重,这样我们就可以进一步区分 layers/models 相对于任何张量的输出custom_variables 可能依赖于(对于在 modulating sub-nets, parameter generation, meta-learning 等中实现功能特别有用)。

变量作用域使使用自定义 getter 的作用域内嵌套所有图形构建变得容易,并在提供的张量之上构建模型作为它们的参数。由于在 TF 2.0 中不再建议使用会话和变量作用域(并且所有这些低级内容都移到了 tf.compat.v1),因此 最佳实践 是什么来实现以上使用 Keras 和 TF 2.0?

(相关 issue on GitHub。)

根据下面的评论回答

假设你有:

kernel = createTheKernelVarBasedOnWhatYouWant() #shape (784, 64)
bias = createTheBiasVarBasedOnWhatYouWant() #shape (64,)

Dense:

复制代码制作一个简单的函数
def custom_dense(x):
    inputs, kernel, bias = x

    outputs = K.dot(inputs, kernel)
    outputs = K.bias_add(outputs, bias, data_format='channels_last')
    return outputs

Lambda层使用函数:

layer = Lambda(custom_dense)
hiddens = layer([x, kernel, bias])

Warning: kernel and bias must be produced from a Keras layer, or come from an kernel = Input(tensor=the_kernel_var) and bias = Input(tensor=bias_var)


如果上面的警告对您不利,您可以随时使用 kernelbias “从外部”,例如:

def custom_dense(inputs):
    outputs = K.dot(inputs, kernel) #where kernel is not part of the arguments anymore
    outputs = K.bias_add(outputs, bias, data_format='channels_last')
    return outputs

layer = Lambda(custom_dense)
hiddens = layer(x)

最后一个选项使 save/load 模型变得有点复杂。

旧答案

您可能应该使用 Keras Dense 层并以标准方式设置其权重:

layer = tf.keras.layers.Dense(64, name='the_layer')
layer.set_weights([np.random.rand(784, 64), np.random.rand(64)])

如果你需要这些权重是不可训练的,在编译你设置的keras模型之前:

model.get_layer('the_layer').trainable=False

如果您想直接访问变量作为张量,它们是:

kernel = layer.kernel    
bias = layer.bias

还有很多其他选择,但这取决于您的确切意图,这在您的问题中并不清楚。

不完全确定我是否正确理解了你的问题,但在我看来,应该可以通过 custom layers and keras functional api.

的组合来做你想做的事

自定义层允许您以与 Keras 兼容的方式构建您想要的任何层,例如:

class MyDenseLayer(tf.keras.layers.Layer):
    def __init__(self, num_outputs):
        super(MyDenseLayer, self).__init__()
        self.num_outputs = num_outputs

    def build(self, input_shape):
        self.kernel = self.add_weight("kernel", 
                                      shape=[int(input_shape[-1]), 
                                             self.num_outputs],
                                      initializer='normal')

        self.bias = self.add_weight("bias", 
                                    shape=[self.num_outputs,],
                                    initializer='normal')

    def call(self, inputs):
        return tf.matmul(inputs, self.kernel) + self.bias

并且函数 api 允许您访问所述层的输出并重新使用它们:

inputs = keras.Input(shape=(784,), name='img')
x1 = MyDenseLayer(64, activation='relu')(inputs)
x2 = MyDenseLayer(64, activation='relu')(x1)
outputs = MyDenseLayer(10, activation='softmax')(x2)

model = keras.Model(inputs=inputs, outputs=outputs, name='mnist_model')

此处x1x2可以连接到其他子网。

下面是一个通用解决方案,适用于 TF2 中的任意 Keras 模型。

首先,我们需要定义一个辅助函数 canonical_variable_name 和一个具有以下签名的上下文管理器 custom_make_variable(参见 meta-blocks library 中的实现)。

def canonical_variable_name(variable_name: str, outer_scope: str):
    """Returns the canonical variable name: `outer_scope/.../name`."""
    # ...

@contextlib.contextmanager
def custom_make_variable(
    canonical_custom_variables: Dict[str, tf.Tensor], outer_scope: str
):
    """A context manager that overrides `make_variable` with a custom function.

    When building layers, Keras uses `make_variable` function to create weights
    (kernels and biases for each layer). This function wraps `make_variable` with
    a closure that infers the canonical name of the variable being created (of the
    form `outer_scope/.../var_name`) and looks it up in the `custom_variables` dict
    that maps canonical names to tensors. The function adheres the following logic:

    * If there is a match, it does a few checks (shape, dtype, etc.) and returns
      the found tensor instead of creating a new variable.
    * If there is a match but checks fail, it throws an exception.
    * If there are no matching `custom_variables`, it calls the original
      `make_variable` utility function and returns a newly created variable.
    """
    # ...

使用这些函数,我们可以使用自定义张量作为变量创建任意 Keras 模型:

import numpy as np
import tensorflow as tf

canonical_custom_variables = {
    "model/dense/kernel": tf.constant(
        np.random.rand(784, 64), name="custom_kernel", dtype=tf.float32),
    "model/dense/bias": tf.constant(
        np.random.rand(64), name="custom_bias", dtype=tf.float32),
}

# Compute hiddens using a dense layer with custom variables.
x = tf.random.normal(shape=(1, 784), name="inputs")
with custom_make_variable(canonical_custom_variables, outer_scope="model"):
    Layer = tf.layers.Dense(64)
    hiddens = Layer(x)

print(Layer.variables)