https://blog.tensorflow.org/2019/03/variational-autoencoders-with.html?hl=iw

TensorFlow Probability

https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEgaxGqRBosX2MlO2oM9AWpb5d14Ff9DfzfeQCC0-qEVl18Dx4Ieejw0I-NE0xZ5weTlZ5vZb-ltQ2kwm74CUmg96xnL2fvse3o3IZJJPoOmVXQFPM11CcLo3rDK_ZwKK3_6ZHL6ahWAEZk/s1600/1_5GTRWkUWwQm3R-tdgbqcdw.png

מרץ 08, 2019 —
*Posted by Ian Fischer, Alex Alemi, Joshua V. Dillon, and the TFP Team*

At the 2019 TensorFlow Developer Summit, we announced TensorFlow Probability (TFP) Layers. In that presentation, we showed how to build a powerful regression model in very few lines of code. Here, we will show how easy it is to make a Variational Autoencoder (VAE) using TFP Layers.

TensorFlow Probability LayersTFP Layers provide…

Variational Autoencoders with Tensorflow Probability Layers

At the 2019 TensorFlow Developer Summit, we announced TensorFlow Probability (TFP) Layers. In that presentation, we showed how to build a powerful regression model in very few lines of code. Here, we will show how easy it is to make a Variational Autoencoder (VAE) using TFP Layers.

In the traditional derivation of a VAE, we imagine some process that generates the data, such as a latent variable generative model. Consider the process of drawing digits, as in MNIST. Suppose that before you draw the digit, you first decide which digit you will draw, imagining some fuzzy picture in your head. Then, you put pen to paper and try to create the picture in the real world. We can formalize this two step process:

- You sample some
*latent*representation*z*from some prior distribution*z ~ p(z)*. This is the fuzzy picture in your head — let’s say of a “3”. - Based on your sample, you draw the actual picture representation
*x*, modeled itself as a stochastic process*x ~ p(x|z)*. This captures the idea that each time you write a “3”, it looks at least a little different.

Thus, when a handwritten digit is created, we imagine some of the variation is due to some kind of

To train this objective, we maximize the ELBO (Evidence Lower BOund) objective:

Where the three probability density functions are:

*p(z)*, the*prior*on the latent representation*z*,*q(z|x)*, the*variational encoder*, and*p(x|z)*, the*decoder*— how likely is the image*x*given the latent representation*z*.

The ELBO is a lower bound on

```
tfd = tfp.distributions
encoded_size = 16
prior = tfd.Independent(tfd.Normal(loc=tf.zeros(encoded_size), scale=1),
reinterpreted_batch_ndims=1)
```

Here, we have just created a TFP independent Gaussian distribution with no learned parameters, and we have specified that our latent variable, ```
tfpl = tfp.layers
encoder = tfk.Sequential([
tfkl.InputLayer(input_shape=input_shape),
tfkl.Lambda(lambda x: tf.cast(x, tf.float32) - 0.5),
tfkl.Conv2D(base_depth, 5, strides=1,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2D(base_depth, 5, strides=2,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2D(2 * base_depth, 5, strides=1,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2D(2 * base_depth, 5, strides=2,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2D(4 * encoded_size, 7, strides=1,
padding='valid', activation=tf.nn.leaky_relu),
tfkl.Flatten(),
tfkl.Dense(tfpl.MultivariateNormalTriL.params_size(encoded_size),
activation=None),
tfpl.MultivariateNormalTriL(
encoded_size,
activity_regularizer=tfpl.KLDivergenceRegularizer(prior, weight=1.0)),
])
```

The encoder is just a normal Keras Sequential model, consisting of convolutions and dense layers, but the output is passed to a TFP Layer, `MultivariateNormalTril()`

, which transparently splits the activations from the final `Dense()`

layer into the parts needed to specify both the mean and the (lower triangular) covariance matrix, the parameters of a Multivariate Normal. We used a helper, `tfpl.MultivariateNormalTriL.params_size(encoded_size)`

, to make the `Dense()`

layer output the correct number of activations (i.e., the distribution’s parameters). Finally, we said that the distribution should contribute a “regularization” term to the final loss. Specifically, we are adding the KL divergence between the encoder and the prior to the loss, which is the `weight`

argument to something other than 1!)```
decoder = tfk.Sequential([
tfkl.InputLayer(input_shape=[encoded_size]),
tfkl.Reshape([1, 1, encoded_size]),
tfkl.Conv2DTranspose(2 * base_depth, 7, strides=1,
padding='valid', activation=tf.nn.leaky_relu),
tfkl.Conv2DTranspose(2 * base_depth, 5, strides=1,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2DTranspose(2 * base_depth, 5, strides=2,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2DTranspose(base_depth, 5, strides=1,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2DTranspose(base_depth, 5, strides=2,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2DTranspose(base_depth, 5, strides=1,
padding='same', activation=tf.nn.leaky_relu),
tfkl.Conv2D(filters=1, kernel_size=5, strides=1,
padding='same', activation=None),
tfkl.Flatten(),
tfpl.IndependentBernoulli(input_shape, tfd.Bernoulli.logits),
])
```

The form here is essentially the same as the encoder, but now we are using transposed convolutions to take our latent representation, which is a 16 dimensional vector, and turn it into a 28 x 28 x 1 tensor. That final tensor parameterizes the pixel-independent Bernoulli distribution.```
vae = tfk.Model(inputs=encoder.inputs,
outputs=decoder(encoder.outputs[0]))
```

Our model is just a Keras Model where the outputs are defined as the composition of the encoder and the decoder. Since the encoder already added the KL term to the loss, we need to specify only the reconstruction loss (the first term of the ELBO above).```
negative_log_likelihood = lambda x, rv_x: -rv_x.log_prob(x)
vae.compile(optimizer=tf.optimizers.Adam(learning_rate=1e-3),
loss=negative_log_likelihood)
```

The loss function takes two arguments — the original input, x, and the output of the model. We call that `rv_x`

because it is a random variable. This example demonstrates some of the core magic of TFP Layers — even though Keras and Tensorflow view the TFP Layers as outputting tensors, TFP Layers are actually `-rv_x.log_prob(x)`

.```
x = eval_dataset.make_one_shot_iterator().get_next()[0][:10]
xhat = vae(x)
assert isinstance(xhat, tfd.Distribution)
```

But if a TFP Layer returns a Distribution, what happens when we compose the decoder with the output of the encoder: `decoder_model(encoder_model.outputs[0]))`

? Well, in order for Keras to view the encoder distribution as a Tensor, TFP Layers actually “reifies” the distribution as a sample from that distribution, which is just a fancy way of saying that Keras sees the Distribution object as the Tensor we would have gotten, had we called `encoder_model.sample()`

. But, when we need to access the Distribution object directly, we can — just like we do in the loss function when we call `rv_x.log_prob(x)`

. TFP Layers provides the distribution-like and Tensor-like behaviors automatically, so you don’t need to worry about Keras getting confused.`vae_model.fit()`

:```
vae.fit(train_dataset,
epochs=15,
validation_data=eval_dataset)
```

With this model, we are able to get an ELBO of around 115 nats (the Decoder modes generated by encoding images from the MNIST test set. |

Decoder modes generated by sampling from the prior. |

We utilized the tensor-like and distribution-like semantics of TFP layers to make our code relatively straightforward.

Next post

TensorFlow Probability
**·**

Variational Autoencoders with Tensorflow Probability Layers

מרץ 08, 2019
—
*Posted by Ian Fischer, Alex Alemi, Joshua V. Dillon, and the TFP Team*

At the 2019 TensorFlow Developer Summit, we announced TensorFlow Probability (TFP) Layers. In that presentation, we showed how to build a powerful regression model in very few lines of code. Here, we will show how easy it is to make a Variational Autoencoder (VAE) using TFP Layers.

TensorFlow Probability LayersTFP Layers provide…

Build, deploy, and experiment easily with TensorFlow