Add L2 regularization when using high level tf.layers

user_6255101 0
Viewed 28688

Is it possible to add an L2 regularization when using the layers defined in tf.layers?

It seems to me that since tf.layers is an high level wrapper, there is no easy way to get access to the filter weights.

With tf.nn.conv2d

regularizer = tf.contrib.layers.l2_regularizer(scale=0.1)

weights = tf.get_variable(
    name="weights",
    regularizer=regularizer
)

#Previous layers

...

#Second layer 
layer 2 = tf.nn.conv2d(
input,
weights,
[1,1,1,1],
[1,1,1,1])

#More layers
...

#Loss
loss = #some loss

reg_variables = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
reg_term = tf.contrib.layers.apply_regularization(regularizer, reg_variables)
loss += reg_term

Now what would that look like with tf.layers.conv2d?

Thanks!

tensorflow
4 Answers

I see two incomplete answers, so here is the complete one:

regularizer = tf.contrib.layers.l2_regularizer(scale=0.1)
    layer2 = tf.layers.conv2d(
        inputs,
        filters,
        kernel_size,
        kernel_regularizer=regularizer)

alternatively:

layer2 = tf.layers.conv2d(inputs, 
     filters, 
     kernel_size,                        
     kernel_regularizer= tf.contrib.layers.l2_regularizer(scale=0.1))

don't forget to add it to the final loss:

l2_loss = tf.losses.get_regularization_loss()
....
loss += l2_loss

Basically, add regularization when defining a layer and then make sure you add regularization loss to your loss.

user_48337730

In eager execution, there are two ways.

  1. Calculate by hand with tf.add_n([tf.square(i) for i in layer.variables]) * l2_coef .
  2. Using layer.losses when the layer is created with kernel_regularizer.

As shown in the official examples: densenet_test.py

rand_input = tf.random_uniform((10, 3, 32, 32))
weight_decay = 1e-4

conv = tf.keras.layers.Conv2D(
    3, (3, 3),
    padding='same',
    use_bias=False,
    kernel_regularizer=tf.keras.regularizers.l2(weight_decay))

optimizer = tf.train.GradientDescentOptimizer(0.1)
conv(rand_input)  # Initialize the variables in the layer

def compute_true_l2(vs, wd):
  return tf.reduce_sum(tf.square(vs)) * wd

true_l2 = compute_true_l2(conv.variables, weight_decay)
keras_l2 = tf.add_n(conv.losses)
self.assertAllClose(true_l2, keras_l2)

with tf.GradientTape() as tape_true, tf.GradientTape() as tape_keras:
  loss = tf.reduce_sum(conv(rand_input))
  loss_with_true_l2 = loss + compute_true_l2(conv.variables, weight_decay)
  loss_with_keras_l2 = loss + tf.add_n(conv.losses)

true_grads = tape_true.gradient(loss_with_true_l2, conv.variables)
keras_grads = tape_keras.gradient(loss_with_keras_l2, conv.variables)
self.assertAllClose(true_grads, keras_grads)
user_79841840
Related