Search code examples
tensorflowkerasregularized

Regularization function using weights from multiple layers?


I don't know if it is feasible but I'm asking just in case. Here is the (simplified) architecture of my model.

Layer (type)           Output Shape        Param  #Connected to
==========================================
input_1 (InputLayer)   [(None, 7, 7, 1024)  0
conv (Conv2D)          (None, 7, 7, 10)     10240 input_1[0][0]

where each of the 10 filters in "conv" is a 1x1x1024 convolutional filter (with no bias but it's irrelevant for this particular issue). I am currently using a custom regularization function on "conv" to make sure that the (1x1)x1024x10 matrix of filter weights has a nice property (basically that all vectors are pairwise orthogonal) and so far, everything is working as expected. Now I also want the ability to disable training on some of these 10 filters. The only way I know how to do that would be to implement 10 filters independently as follows

Layer (type)                    Output Shape         Param #     Connected to                     
=========================================================
input_1 (InputLayer)            [(None, 7, 7, 1024) 0
conv_1 (Conv2D)          (None, 7, 7, 1)     1024       input_1[0][0]
conv_2 (Conv2D)          (None, 7, 7, 1)     1024       input_1[0][0]
conv_3 (Conv2D)          (None, 7, 7, 1)     1024       input_1[0][0]
...
conv_10 (Conv2D)          (None, 7, 7, 1)     1024       input_1[0][0]

followed by a Concatenate layer, then to set the "trainable" parameter to True/False on each conv_i layer as I see fit. However, now I don't know how to implement my regularization function which must be computed on the weights of all layers conv_i simultaneously rather than independently. Is there a trick that I can use to implement such function? Or conversely, is there a way to freeze only part of the weights of a convolutional layer? Thanks!

Solution

For those interested, here is the working code for my problem following the advice provided by @LaplaceRicky.

class SpecialRegularization(tf.keras.Model):
   """ In order to avoid a warning message when saving the model, 
   I use the solution indicated here 
   https://github.com/tensorflow/tensorflow/issues/44541
   and now inherit from tf.keras.Model instead of Layer
   """
    def __init__(self,nfilters,**kwargs):
        super().__init__(**kwargs)
        self.inner_layers=[Conv2D(1,(1,1)) for _ in range(nfilters)]

    def call(self, inputs):
        outputs=[l(inputs) for l in self.inner_layers]
        self.add_loss(self.define_your_regularization_here())
        return tf.concat(outputs,-1)

    def set_trainable_parts(self, trainables):
        """ Set the trainable attribute independently on each filter """
        for l,t in zip(self.inner_layers,trainables):
            l.trainable = t

    def define_your_regularization_here(self):
        #reconstruct the original kernel
        large_kernel=tf.concat([l.kernel for l in self.inner_layers],-1)
        return tf.reduce_sum(large_kernel*large_kernel[:,:,:,::-1])
        

Solution

  • One way to achieve this is to have a custom keras layer that wraps all of the small conv layers and is responsible for computing the regularization loss.

    Example Codes:

    import tensorflow as tf
    
    def _get_losses(model,x):
        model(x)
        return model.losses
    
    def _get_grads(model,x):
      with tf.GradientTape() as t:
        model(x)
        reg_loss=tf.math.add_n(model.losses)
      return t.gradient(reg_loss,model.trainable_weights)
    
    class SpecialRegularization(tf.keras.layers.Layer):
        def __init__(self, **kwargs):
            self.inner_layers=[tf.keras.layers.Conv2D(1,(1,1)) for i in range(10)]
            super().__init__(**kwargs)
    
        def call(self, inputs,training=None):
            outputs=[l(inputs,training=training) for l in self.inner_layers]
            self.add_loss(self.define_your_regularization_here())
            return tf.concat(outputs,-1)
    
        def define_your_regularization_here(self):
          #reconstruct the original kernel
          large_kernel=tf.concat([l.kernel for l in self.inner_layers],-1)
          #just giving an example here
          #you should define your own regularization using the entire kernel
          return tf.reduce_sum(large_kernel*large_kernel[:,:,:,::-1])
    
    tf.random.set_seed(123)
    inputs = tf.keras.Input(shape=(7,7,1024))
    outputs = SpecialRegularization()(inputs)
    model = tf.keras.Model(inputs=inputs, outputs=outputs)
    
    #get_losses, get_grads are for demonstration purpose
    get_losses=tf.function(_get_losses)
    get_grads=tf.function(_get_grads)
    data=tf.random.normal((64,7,7,1024))
    print(get_losses(model,data))
    print(get_grads(model,data)[0])
    print(model.layers[1].inner_layers[-1].kernel*2)
    model.summary()
    '''
    [<tf.Tensor: shape=(), dtype=float32, numpy=-0.20446025>]
    tf.Tensor(
    [[[[ 0.02072023]
       [ 0.12973154]
       [ 0.11631528]
       ...
       [ 0.00804012]
       [-0.07299817]
       [ 0.06031524]]]], shape=(1, 1, 1024, 1), dtype=float32)
    tf.Tensor(
    [[[[ 0.02072023]
       [ 0.12973154]
       [ 0.11631528]
       ...
       [ 0.00804012]
       [-0.07299817]
       [ 0.06031524]]]], shape=(1, 1, 1024, 1), dtype=float32)
    Model: "model"
    _________________________________________________________________
    Layer (type)                 Output Shape              Param #   
    =================================================================
    input_1 (InputLayer)         [(None, 7, 7, 1024)]      0         
    _________________________________________________________________
    special_regularization (Spec (None, 7, 7, 10)          10250     
    =================================================================
    Total params: 10,250
    Trainable params: 10,250
    Non-trainable params: 0
    _________________________________________________________________
    '''