@tf.function( input_signature ) 在 class scope 之外定义的对象方法

Question

Say I have a Custom Layer :

class Custom_Layer(keras.layers.Layer):
    def __init__(self, **kwargs):
        self.w_0 = tf.Variable(tf.random_uniform_initializer(),trainable=True)
        self.b_0 = tf.Variable(tf.zeros_initializer(),trainable=True)    
        ....
    def call(self, inputs):
        output = A_Method(self, inputs)
        return output

def A_Method(self, TensorA):
    ....
    return something

if I want to decorate @tf.function(with input_signature) to A_Method to control tracing

@tf.function(input_signature=[???,  tf.TensorSpec(shape=None)])
def A_Method(self, TensorA):
    ....
    return something

what spec should i put for self ? I tried putting tf.TensorSpec but it raised an error

___Updated the question___:

Im quite new to tensorflow sorry if the code is weird or doesnt make sense. the reason I do this is I found RNN took a long time for first epoch to get started, I dont know if this custom layer can do something alike but is taking less time. but ultimately I believe the slow initialize time is because of tensorflow retracing repeatedly even on same input_spec - input_shape . I use this layer repeatedly as,

input_layer = Input(shape=( X_.shape[1],X_.shape[2]),  name='input')
for loop :
Hard_Code_RNN_Layer(input_layer[:,:, slicing])

then I ran .experimental_get_tracing_count() count is 300 which really shouldn't be above 10, thats why I wanted to take this method out def Mimic_RNN(self, step_input, step_state) remove it from the class and try giving it an input_signature. Please see below:

def Initialize_Variable(input_dim, units):
    w_init = tf.random_normal_initializer()
    b_init = tf.zeros_initializer()
    w_0 = tf.Variable(initial_value=w_init(shape=(input_dim, units))) 
    b_0 = tf.Variable(initial_value=b_init(shape=(units)))
    return w_0, b_0

def Initialize_One_Variable(input_dim, units):
    w_init = tf.random_uniform_initializer()
    R_kernal = tf.Variable(initial_value=w_init(shape=(input_dim, units))) 
    return R_kernal

class Hard_Code_RNN_Layer(keras.layers.Layer):
    def __init__(self, input_tuple, Sequencee=True, **kwargs):
        super(Hard_Code_RNN_Layer, self).__init__()
        input_shape, units = input_tuple       
        self.Hidden_Size = (int)(input_shape * 0.85)
        self.inputshape = input_shape
        self.units = units
        self.thiseq = Sequencee
        self.Uz = Initialize_One_Variable(self.Hidden_Size, self.Hidden_Size)
        self.Ur = Initialize_One_Variable(self.Hidden_Size, self.Hidden_Size)
        self.w_hz, self.b_hz    = Initialize_Variable(self.units, self.Hidden_Size)
        self.w_out, self.b_out  = Initialize_Variable(self.Hidden_Size,self.units)
        self.w_0, self.b_0  = Initialize_Variable(self.inputshape,self.units)

    def get_config(self):
        cfg = super().get_config()
        return cfg 

    def Layer_Method(inputs, w_h, b_h):
        return tf.matmul(inputs, w_h) + b_h
    
    def Mimic_RNN(self, step_input, step_state):  <-----------input_signature_this
        x__j = self.Layer_Method(step_input, self.w_0, self.b_0)
        r = tf.sigmoid(tf.matmul(step_state, self.Ur))
        z = tf.sigmoid(tf.matmul(step_state, self.Uz))
        h__ = tf.nn.relu(tf.matmul(x__j, self.w_hz) +  tf.multiply(r, step_state) + self.b_hz) 
        h = (1-z) * h__ + z * step_state
        output__ = tf.nn.relu(tf.matmul(h, self.w_out) + self.b_out)
        
        return output__, h
    
    def call(self, inputs):
        unstack = tf.unstack(inputs, axis=1)
        out1, hiddd = self.Mimic_RNN(step_input=unstack[0], step_state=tf.zeros_like(unstack[0][:,0:self.Hidden_Size]))
        out2, hiddd = self.Mimic_RNN(step_input=unstack[1], step_state=hiddd)
        out3, hiddd = self.Mimic_RNN(step_input=unstack[2], step_state=hiddd)
        
        if(self.thiseq):
            return tf.stack([out1, out2, out3], axis =1 )  
        else:
            return out3

Answer 1

You actually can use input_signature with class methods. The initial self parameter simply needs to be ignored in the input signature specification, so you just provide the tf.TensorSpec for the other parameters.

For example:

import tensorflow as tf

class MyClass:
    @tf.function(input_signature=(tf.TensorSpec([None], tf.float32),
                                  tf.TensorSpec([None], tf.float32)))
    def my_method(self, a, b):
        return a + b

tf.print(MyClass().my_method([1, 2, 3], [4]))
# 5, 6, 7

Answer 2

If you specify an input signature, all inputs to the python function must be convertible to Tensor . self in that case, holds the reference to the instance on which the method is called on, and cannot be convertible as a Tensor. You just can't specify an input_signature on your A_method function.

However, it is still possible to decorate a method from a class, because TensorFlow will detect if the function to decorate is a method, and if that's the case, will automatically remove the self argument. You can check the source code :

if self._is_method:
  # Remove `self`: default arguments shouldn't be matched to it.
  # TODO(b/127938157): Should this error out if there is no arg to
  # be removed?
  args = fullargspec.args[1:]

It's worth noting that if a method is defined outside of the class, then this check will fail. (The check relies on the ismethod function from the standard library inspect module). As self is not convertible to a Tensor, the decorated method will throw an error when called.

It's not best practice to define a method outside of the class definition: it makes the code harder to read, and harder to use. You can give a look to that question for more details: Define a method outside of class definition? . The python way of reusing logic between classes is either to use inheritance, or to define a function that does not depend on the attributes of the object (or where those attributes are passed as argument to the function).