tensorflow 2.0: An op outside of the function building code is being passed

Question

I'm getting an error:

TypeError: An op outside of the function building code is being passed
a "Graph" tensor. It is possible to have Graph tensors
leak out of the function building context by including a
tf.init_scope in your function building code.
For example, the following function will fail:
  @tf.function
  def has_init_scope():
    my_constant = tf.constant(1.)
    with tf.init_scope():
      added = my_constant * 2

Using an NVP layer like follows:

import tensorflow_probability as tfp
tfb = tfp.bijectors
tfd = tfp.distributions
class NVPLayer(tf.keras.models.Model):

    def __init__(self, *, output_dim, num_masked, **kwargs):
        super().__init__(**kwargs)
        self.output_dim = output_dim
        self.num_masked = num_masked
        self.shift_and_log_scale_fn = tfb.real_nvp_default_template(
            hidden_layers=[2], # HERE HERE ADJUST THIS
            activation=None, # linear
            )
        self.loss = None

    def get_nvp(self):
        nvp = tfd.TransformedDistribution(
            distribution=tfd.MultivariateNormalDiag(loc=[0.] * self.output_dim),
            bijector=tfb.RealNVP(
                num_masked=self.num_masked,
                shift_and_log_scale_fn=self.shift_and_log_scale_fn)
            )
        return nvp

    def call(self, *inputs):
        nvp = self.get_nvp()
        self.loss = tf.reduce_mean(nvp.log_prob(*inputs)) # how else to do this?
        # return nvp.bijector.forward(*inputs)
        return nvp.bijector.inverse(*inputs)

I'm not calling tf.init_scope anywhere. A simple version training a layer like seems to work.

I will try to get more more granular trace but I suspect this is something to do with non-eager mode stuff.

UPDATE: so this is definitely coming from the self.loss inclusion in some gradient tape layer. What is the correct way of doing this?

Answer 1

Just had the same problem couple of minutes ago, in my case I wanted to modify state inside my loss function class, here is how I would solve it in your case.

BTW @simon gave me an inspiration of how to properly asses this. So props to him!

It seems that you should create a tf.Variable for attributes you are going to change while training. Notice how you didn't have any problems with other attributes like self.output_dim, self.num_masked and other.

Try this:

import tensorflow_probability as tfp
tfb = tfp.bijectors
tfd = tfp.distributions
class NVPLayer(tf.keras.models.Model):

def __init__(self, *, output_dim, num_masked, **kwargs):
    super().__init__(**kwargs)
    self.output_dim = output_dim
    self.num_masked = num_masked
    self.shift_and_log_scale_fn = tfb.real_nvp_default_template(
        hidden_layers=[2], # HERE HERE ADJUST THIS
        activation=None, # linear
        )

    ###CHANGE HERE
    self.loss = tf.Variable(0.0)

def get_nvp(self):
    nvp = tfd.TransformedDistribution(
        distribution=tfd.MultivariateNormalDiag(loc=[0.] * self.output_dim),
        bijector=tfb.RealNVP(
            num_masked=self.num_masked,
            shift_and_log_scale_fn=self.shift_and_log_scale_fn)
        )
    return nvp

def call(self, *inputs):
    nvp = self.get_nvp()

    ### CHANGE HERE
    self.loss.assign(tf.reduce_mean(nvp.log_prob(*inputs)))
    # return nvp.bijector.forward(*inputs)
    return nvp.bijector.inverse(*inputs)

Check out this answer on github issue also, similar problem!

Answer 2

UPDATE: so this is definitely coming from the self.loss inclusion in some gradient tape layer. What is the correct way of doing this?

I think the correct way of doing this is by

self.add_loss(<your loss tensor>)

(https://www.tensorflow.org/api_docs/python/tf/keras/layers/Layer#add_loss for more on this)

(edit sorry, I wasn't paying attention to the date of your post, so I guess this wasn't very useful anymore lol)