TensorFlow 2.0 Keras 层以自定义张量作为变量

Question

In TF 1.x, it was possible to build layers with custom variables. Here's an example:

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)

The printed variables of the constructed dense layer will be custom tensors we specified in the custom_variables dict:

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

This allows us to create layers/models that use provided tensors in custom_variables directly as their weights, so that we could further differentiate the output of the layers/models with respect to any tensors that custom_variables may depend on (particularly useful for implementing functionality in modulating sub-nets , parameter generation , meta-learning , etc.).

Variable scopes used to make it easy to nest all off graph-building inside scopes with custom getters and build models on top of the provided tensors as their parameters. Since sessions and variable scopes are no longer advisable in TF 2.0 (and all of that low-level stuff is moved to tf.compat.v1 ), what would be the best practice to implement the above using Keras and TF 2.0?

(Related issue on GitHub .)

Answer 1

Answer based on the comment below

Given you have:

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

Make a simple function copying the code from Dense :

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

Use the function in a Lambda layer:

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)

---

If the warning above is bad for you, you can always use kernel and bias "from outside", like:

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)

This last option makes it a bit more complicated to save/load models.

Old answer

You should probably use a Keras Dense layer and set its weights in a standard way:

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

If you need that these weights are not trainable, before compiling the keras model you set:

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

If you want direct access to the variables as tensors, they are:

kernel = layer.kernel    
bias = layer.bias

---

There are plenty of other options, but that depends on your exact intention, which is not clear in your question.

Answer 2

Below is a general-purpose solution that works with arbitrary Keras models in TF2.

First, we need to define an auxiliary function canonical_variable_name and a context manager custom_make_variable with the following signatures (see implementation in meta-blocks library ).

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.
    """
    # ...

Using these functions, we can create arbitrary Keras models with custom tensors used as variables:

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)

Answer 3

Not entirely sure I understand your question correctly, but it seems to me that it should be possible to do what you want with a combination ofcustom layers and keras functional api .

Custom layers allow you to build any layer you want in a way that is compatible with Keras, eg:

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

and the functional api allows you to access the outputs of said layers and re-use them:

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')

Here x1 and x2 can be connected to other subnets.