Models using a KLDivergenceRegularizer cannot be saved

[jsilter]

I've been working through the TFP VAE tutorial, and tried to save the model after training. However, this was not possible in SavedModel format.

Command:

model.save(saved_model_path, save_format="tf", save_traces=True)

Produces the following error:

` File "/Users/xxx/venvs/ml_venv/lib/python3.9/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1251, in call * return self._kl_divergence_fn(distribution_a)

File "/Users/xxx/venvs/ml_venv/lib/python3.9/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1370, in _fn  **  
    kl = kl_divergence_fn(distribution_a, distribution_b_)   

File "/Users/xxx/venvs/ml_venv/lib/python3.9/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1354, in kl_divergence_fn. 
    distribution_a.log_prob(z) - distribution_b.log_prob(z),  
AttributeError: 'Tensor' object has no attribute 'log_prob'`. 

Command:

model.save(saved_model_path, save_format="tf", save_traces=False)

Produces the following error:

<tensorflow_probability.python.layers.distribution_layer.KLDivergenceRegularizer object at 0x16b79c5b0> does not implement get_config()

Indeed it does not, KLDivergenceRegularizer does not have a get_config implementation.

Sidenote: I cannot load from h5 because MultivariateNormalTriL requires a positional argument: TypeError: __init__() missing 1 required positional argument: 'event_size'

Versions: Python 3.9.13 tensorflow-probability==0.17.0 tensorflow==2.9.0

This is also discussed in #742 but I'm starting a new issue to make it more prominent.

[johnsoltis]

I am having the same issues, after (loosely) following the same tutorial.

Versions: Python 3.10.4 Tensorflow 2.9.1 Tensorflow Probability 0.17.0

The command it is failing on is: vae.save(file_name)

The error:

Traceback (most recent call last):
  File "/home/jsoltis2/VAE_LATENT3.py", line 279, in <module>
    vae.save(file_name)
  File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/tmp/__autograph_generated_filete4qkxqr.py", line 13, in tf____call__
    retval_ = ag__.converted_call(ag__.ld(self)._kl_divergence_fn, (ag__.ld(distribution_a),), None, fscope)
  File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1370, in _fn
    kl = kl_divergence_fn(distribution_a, distribution_b_)
  File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1354, in kl_divergence_fn
    distribution_a.log_prob(z) - distribution_b.log_prob(z),
AttributeError: in user code:

    File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1251, in __call__  *
        return self._kl_divergence_fn(distribution_a)
    File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1370, in _fn  **
        kl = kl_divergence_fn(distribution_a, distribution_b_)
    File "/home/jsoltis2/.conda/envs/tf_tfp/lib/python3.10/site-packages/tensorflow_probability/python/layers/distribution_layer.py", line 1354, in kl_divergence_fn
        distribution_a.log_prob(z) - distribution_b.log_prob(z),

    AttributeError: 'Tensor' object has no attribute 'log_prob'

My full model and compilation:

prior = tfd.Independent(tfd.Normal(loc=tf.zeros(latent_dim), scale=1), reinterpreted_batch_ndims=1)

encoder = tfk.Sequential([tfkl.InputLayer(input_shape=(140, 140, 1)), 
    Conv2D(16, kernel_size=(3, 3), activation=None),
    layers.LeakyReLU(),
    Conv2D(32, kernel_size=(3, 3), activation=None),
    layers.LeakyReLU(),
    Conv2D(64, kernel_size=(3, 3), activation=None),
    layers.LeakyReLU(),
    layers.Flatten(),
    tfkl.Dense(tfpl.MultivariateNormalTriL.params_size(latent_dim),activation=None),
    tfpl.MultivariateNormalTriL(latent_dim, activity_regularizer=tfpl.KLDivergenceRegularizer(prior)),]
    ,name='encoder')

encoder.summary()


decoder = tfk.Sequential([tfkl.InputLayer(input_shape=(latent_dim,)),
    Dense(140*140*128, activation=None),
    layers.LeakyReLU(),
    layers.Reshape(target_shape=(140, 140, 128)),
    Conv2DTranspose(64, kernel_size=(3, 3), activation=None, padding='same'),
    layers.LeakyReLU(),
    Conv2DTranspose(32, kernel_size=(3, 3), activation=None, padding='same'),
    layers.LeakyReLU(),
    Conv2DTranspose(16, kernel_size=(3, 3), activation=None, padding='same'),
    layers.LeakyReLU(),
    Conv2DTranspose(1, kernel_size=(3, 3), activation=None, padding='same'),]
    ,name='decoder')
decoder.summary()

vae = tfk.Model(inputs=encoder.inputs, outputs=decoder(encoder.outputs[0]))

vae.compile(optimizer=tf.optimizers.Adam(learning_rate=1e-2), loss='mae', metrics=['mse','mae'])

[jsilter]

Looks like it can be saved using tf.compat.v1.keras.experimental.export_saved_model and loaded with tf.compat.v1.keras.experimental.load_from_saved_model. I'm assuming this doesn't actually save the regularizer and skips over it, which is fine for my purposes. These methods are deprecated, hopefully they won't be removed soon.

[i418c]

I'm running into this on the latest releases as well. It's very confusing to follow a tutorial and have something as basic as model saving not work out of the box.