Simple RealNVP raises Keras error


Iam using TF 2.8.2 but I am not an expert fore sure. I need help to understand what is going on in this snippet and correct it. My concern is that I get an error related to legacy tf.layers.Layers

import numpy as np

import tensorflow as tf
import tensorflow_probability as tfp

from tensorflow_probability import bijectors as tfb
from tensorflow_probability import distributions as tfd
from tensorflow import keras as tfk
from tensorflow.keras import layers as tfkl

base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], tf.float32),name='base dist')
trans_dist = tfd.TransformedDistribution(

x_ = tfkl.Input(shape=(2,), dtype=tf.float32)
log_prob_ = trans_dist.log_prob(x_)
model = tfk.Model(x_, log_prob_)

it raises the following eror:

---> 19 model = tfk.Model(x_, log_prob_)

TypeError: The following are legacy tf.layers.Layers:
  <keras.legacy_tf_layers.core.Dense object at 0x7fa0ffd99ad0>
  <keras.legacy_tf_layers.core.Dense object at 0x7fa0ffd99c10>
  <keras.legacy_tf_layers.core.Dense object at 0x7fa0ffd99790>
  <keras.legacy_tf_layers.core.Dense object at 0x7fa19006dd10>
  <keras.legacy_tf_layers.core.Dense object at 0x7fa19006da90>
  <keras.legacy_tf_layers.core.Dense object at 0x7fa19006d090>
To use keras as a framework (for instance using the Network, Model, or Sequential classes), please use the tf.keras.layers implementation instead. (Or, if writing custom layers, subclass from tf.keras.layers rather than tf.layers)


Hi @jecampagne,

I think the error is because when you are declaring x_ which is Input Layer and its place holder for feeding into neural network. Then you are trying to apply log_prob on top of it which should not be the case. tf.keras.Model takes inputs and outputs which are connected together. I have created a gist here please go through it and let me know if this is what you are expecting and correct me if my understanding is not proper. I hope the gist will help you out in solving the issue.

Thanks & Regards,
Sravana Neeli.

Hi @Siva_Sravana_Kumar_N
Below you will find an update of my exercise where you will see a MAF, Invert MAF (aka IAF) and a NVP bijectors. The output of the test is

  1. MAF seems to accept the x_ structure
  2. Invert MAF raise an error at the log_prob_ line
  3. NVP raise an error at the “model line”.

So if I understand your comment I wander why MAF is ok ???
But more importantly would be to give a solution to train with X data which are (N,2), here is what I am using so far

                loss=lambda _, log_prob: -tf.reduce_mean(log_prob)) # signature of loss fn(y_true,output_of_the_model)

hist =,
            y=np.zeros((n_samples, 0), dtype=np.float32),


####### The test on MAF, IAF and NVP ##########

base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], DTYPE),name='base dist')
x_ = tfkl.Input(shape=(2,), dtype=tf.float32)

flow_bijector_IAF = tfb.Invert(tfb.MaskedAutoregressiveFlow(name ='MAF',
                params=2, hidden_units=[512, 512], activation='relu')))

flow_bijector_MAF=tfb.MaskedAutoregressiveFlow( name ='MAF', 
                params=2, hidden_units=[128, 128], activation='relu'))


def test(aBij):
  name =
    print(f">>>>>>>>> START {name} ")
    trans_dist = tfd.TransformedDistribution(

    log_prob_ = trans_dist.log_prob(x_)
    print(f"{name}:", log_prob_)

    model = tfk.Model(x_, log_prob_)

    print(f"<<<<<<<<< END {name} ")
  except Exception as e:
      print("Exception: " + str(e))

for aBij in [flow_bijector_MAF, flow_bijector_IAF, flow_bijector_NVP]:

here is the output

>>>>>>>>> START MAF 
MAF: KerasTensor(type_spec=TensorSpec(shape=(None,), dtype=tf.float32, name=None), name='tf.__operators__.add_165/AddV2:0', description="created by layer 'tf.__operators__.add_165'")
<<<<<<<<< END MAF 
>>>>>>>>> START invert_MAF 
Exception: You are passing KerasTensor(type_spec=TensorSpec(shape=(), dtype=tf.int32, name=None), inferred_value=[2], name='tf.math.reduce_prod_8/Prod:0', description="created by layer 'tf.math.reduce_prod_8'"), an intermediate Keras symbolic input/output, to a TF API that does not allow registering custom dispatchers, such as `tf.cond`, `tf.function`, gradient tapes, or `tf.map_fn`. Keras Functional model construction only supports TF API calls that *do* support dispatching, such as `tf.math.add` or `tf.reshape`. Other APIs cannot be called directly on symbolic Kerasinputs/outputs. You can work around this limitation by putting the operation in a custom Keras layer `call` and calling that layer on this symbolic input/output.
>>>>>>>>> START NVP 
NVP: KerasTensor(type_spec=TensorSpec(shape=(None,), dtype=tf.float32, name=None), name='tf.__operators__.add_171/AddV2:0', description="created by layer 'tf.__operators__.add_171'")
Exception: The following are legacy tf.layers.Layers:
  <keras.legacy_tf_layers.core.Dense object at 0x7f04081f00d0>
  <keras.legacy_tf_layers.core.Dense object at 0x7f040817d150>
  <keras.legacy_tf_layers.core.Dense object at 0x7f04081f0650>
  <keras.legacy_tf_layers.core.Dense object at 0x7f0408105a50>
  <keras.legacy_tf_layers.core.Dense object at 0x7f04081013d0>
  <keras.legacy_tf_layers.core.Dense object at 0x7f0408168990>
To use keras as a framework (for instance using the Network, Model, or Sequential classes), please use the tf.keras.layers implementation instead. (Or, if writing custom layers, subclass from tf.keras.layers rather than tf.layers)