Posit AI Weblog: Illustration studying with MMD-VAE

Lately, we confirmed how one can generate photos utilizing generative adversarial networks (GANs). GANs could yield superb outcomes, however the contract there mainly is: what you see is what you get.
Generally this can be all we would like. In different circumstances, we could also be extra excited about really modelling a website. We don’t simply need to generate realistic-looking samples – we would like our samples to be positioned at particular coordinates in area house.

For instance, think about our area to be the house of facial expressions. Then our latent house is perhaps conceived as two-dimensional: In accordance with underlying emotional states, expressions fluctuate on a positive-negative scale. On the identical time, they fluctuate in depth. Now if we educated a VAE on a set of facial expressions adequately masking the ranges, and it did the truth is “uncover” our hypothesized dimensions, we may then use it to generate previously-nonexisting incarnations of factors (faces, that’s) in latent house.

Variational autoencoders are just like probabilistic graphical fashions in that they assume a latent house that’s answerable for the observations, however unobservable. They’re just like plain autoencoders in that they compress, after which decompress once more, the enter area. In distinction to plain autoencoders although, the essential level right here is to plan a loss perform that permits to acquire informative representations in latent house.

In a nutshell

In normal VAEs (Kingma and Welling 2013), the target is to maximise the proof decrease sure (ELBO):

[ELBO = E[log p(x|z)] – KL(q(z)||p(z))]

In plain phrases and expressed when it comes to how we use it in follow, the primary part is the reconstruction loss we additionally see in plain (non-variational) autoencoders. The second is the Kullback-Leibler divergence between a previous imposed on the latent house (usually, a typical regular distribution) and the illustration of latent house as realized from the information.

A significant criticism relating to the standard VAE loss is that it ends in uninformative latent house. Options embrace (beta)-VAE(Burgess et al. 2018), Information-VAE (Zhao, Tune, and Ermon 2017), and extra. The MMD-VAE(Zhao, Tune, and Ermon 2017) applied under is a subtype of Information-VAE that as a substitute of creating every illustration in latent house as comparable as doable to the prior, coerces the respective distributions to be as shut as doable. Right here MMD stands for most imply discrepancy, a similarity measure for distributions based mostly on matching their respective moments. We clarify this in additional element under.

Our goal in the present day

On this submit, we’re first going to implement a typical VAE that strives to maximise the ELBO. Then, we evaluate its efficiency to that of an Information-VAE utilizing the MMD loss.

Our focus shall be on inspecting the latent areas and see if, and the way, they differ as a consequence of the optimization standards used.

The area we’re going to mannequin shall be glamorous (vogue!), however for the sake of manageability, confined to measurement 28 x 28: We’ll compress and reconstruct photos from the Style MNIST dataset that has been developed as a drop-in to MNIST.

A normal variational autoencoder

Seeing we haven’t used TensorFlow keen execution for some weeks, we’ll do the mannequin in an keen manner.
When you’re new to keen execution, don’t fear: As each new approach, it wants some getting accustomed to, however you’ll rapidly discover that many duties are made simpler in the event you use it. A easy but full, template-like instance is out there as a part of the Keras documentation.

Setup and knowledge preparation

As ordinary, we begin by ensuring we’re utilizing the TensorFlow implementation of Keras and enabling keen execution. Moreover tensorflow and keras, we additionally load tfdatasets to be used in knowledge streaming.

By the best way: No must copy-paste any of the under code snippets. The 2 approaches can be found amongst our Keras examples, particularly, as eager_cvae.R and mmd_cvae.R.

The information comes conveniently with keras, all we have to do is the same old normalization and reshaping.

vogue <- dataset_fashion_mnist()

c(train_images, train_labels) %<-% vogue$practice
c(test_images, test_labels) %<-% vogue$take a look at

train_x <- train_images %>%
  `/`(255) %>%
  k_reshape(c(60000, 28, 28, 1))

test_x <- test_images %>% `/`(255) %>%
  k_reshape(c(10000, 28, 28, 1))

What do we want the take a look at set for, given we’re going to practice an unsupervised (a greater time period being: semi-supervised) mannequin? We’ll use it to see how (beforehand unknown) knowledge factors cluster collectively in latent house.

Now put together for streaming the information to keras:

buffer_size <- 60000
batch_size <- 100
batches_per_epoch <- buffer_size / batch_size

train_dataset <- tensor_slices_dataset(train_x) %>%
  dataset_shuffle(buffer_size) %>%
  dataset_batch(batch_size)

test_dataset <- tensor_slices_dataset(test_x) %>%
  dataset_batch(10000)

Subsequent up is defining the mannequin.

Encoder-decoder mannequin

The mannequin actually is 2 fashions: the encoder and the decoder. As we’ll see shortly, in the usual model of the VAE there’s a third part in between, performing the so-called reparameterization trick.

The encoder is a customized mannequin, comprised of two convolutional layers and a dense layer. It returns the output of the dense layer break up into two components, one storing the imply of the latent variables, the opposite their variance.

latent_dim <- 2

encoder_model <- perform(identify = NULL) {
  
  keras_model_custom(identify = identify, perform(self) {
    self$conv1 <-
      layer_conv_2d(
        filters = 32,
        kernel_size = 3,
        strides = 2,
        activation = "relu"
      )
    self$conv2 <-
      layer_conv_2d(
        filters = 64,
        kernel_size = 3,
        strides = 2,
        activation = "relu"
      )
    self$flatten <- layer_flatten()
    self$dense <- layer_dense(items = 2 * latent_dim)
    
    perform (x, masks = NULL) {
      x %>%
        self$conv1() %>%
        self$conv2() %>%
        self$flatten() %>%
        self$dense() %>%
        tf$break up(num_or_size_splits = 2L, axis = 1L) 
    }
  })
}

We select the latent house to be of dimension 2 – simply because that makes visualization straightforward.
With extra advanced knowledge, you’ll most likely profit from selecting a better dimensionality right here.

So the encoder compresses actual knowledge into estimates of imply and variance of the latent house.
We then “not directly” pattern from this distribution (the so-called reparameterization trick):

reparameterize <- perform(imply, logvar) {
  eps <- k_random_normal(form = imply$form, dtype = tf$float64)
  eps * k_exp(logvar * 0.5) + imply
}

The sampled values will function enter to the decoder, who will try and map them again to the unique house.
The decoder is mainly a sequence of transposed convolutions, upsampling till we attain a decision of 28×28.

decoder_model <- perform(identify = NULL) {
  
  keras_model_custom(identify = identify, perform(self) {
    
    self$dense <- layer_dense(items = 7 * 7 * 32, activation = "relu")
    self$reshape <- layer_reshape(target_shape = c(7, 7, 32))
    self$deconv1 <-
      layer_conv_2d_transpose(
        filters = 64,
        kernel_size = 3,
        strides = 2,
        padding = "identical",
        activation = "relu"
      )
    self$deconv2 <-
      layer_conv_2d_transpose(
        filters = 32,
        kernel_size = 3,
        strides = 2,
        padding = "identical",
        activation = "relu"
      )
    self$deconv3 <-
      layer_conv_2d_transpose(
        filters = 1,
        kernel_size = 3,
        strides = 1,
        padding = "identical"
      )
    
    perform (x, masks = NULL) {
      x %>%
        self$dense() %>%
        self$reshape() %>%
        self$deconv1() %>%
        self$deconv2() %>%
        self$deconv3()
    }
  })
}

Word how the ultimate deconvolution doesn’t have the sigmoid activation you may need anticipated. It is because we shall be utilizing tf$nn$sigmoid_cross_entropy_with_logits when calculating the loss.

Talking of losses, let’s examine them now.

Loss calculations

One solution to implement the VAE loss is combining reconstruction loss (cross entropy, within the current case) and Kullback-Leibler divergence. In Keras, the latter is out there instantly as loss_kullback_leibler_divergence.

Right here, we observe a current Google Colaboratory pocket book in batch-estimating the whole ELBO as a substitute (as a substitute of simply estimating reconstruction loss and computing the KL-divergence analytically):

[ELBO batch estimate = log p(x_{batch}|z_{sampled})+log p(z)−log q(z_{sampled}|x_{batch})]

Calculation of the traditional loglikelihood is packaged right into a perform so we are able to reuse it in the course of the coaching loop.

normal_loglik <- perform(pattern, imply, logvar, reduce_axis = 2) {
  loglik <- k_constant(0.5, dtype = tf$float64) *
    (k_log(2 * k_constant(pi, dtype = tf$float64)) +
    logvar +
    k_exp(-logvar) * (pattern - imply) ^ 2)
  - k_sum(loglik, axis = reduce_axis)
}

Peeking forward some, throughout coaching we are going to compute the above as follows.

First,

crossentropy_loss <- tf$nn$sigmoid_cross_entropy_with_logits(
  logits = preds,
  labels = x
)
logpx_z <- - k_sum(crossentropy_loss)

yields (log p(x|z)), the loglikelihood of the reconstructed samples given values sampled from latent house (a.ok.a. reconstruction loss).

Then,

logpz <- normal_loglik(
  z,
  k_constant(0, dtype = tf$float64),
  k_constant(0, dtype = tf$float64)
)

offers (log p(z)), the prior loglikelihood of (z). The prior is assumed to be normal regular, as is most frequently the case with VAEs.

Lastly,

logqz_x <- normal_loglik(z, imply, logvar)

vields (log q(z|x)), the loglikelihood of the samples (z) given imply and variance computed from the noticed samples (x).

From these three parts, we are going to compute the ultimate loss as

loss <- -k_mean(logpx_z + logpz - logqz_x)

After this peaking forward, let’s rapidly end the setup so we prepare for coaching.

Ultimate setup

Moreover the loss, we want an optimizer that may try to decrease it.

optimizer <- tf$practice$AdamOptimizer(1e-4)

We instantiate our fashions …

encoder <- encoder_model()
decoder <- decoder_model()

and arrange checkpointing, so we are able to later restore educated weights.

checkpoint_dir <- "./checkpoints_cvae"
checkpoint_prefix <- file.path(checkpoint_dir, "ckpt")
checkpoint <- tf$practice$Checkpoint(
  optimizer = optimizer,
  encoder = encoder,
  decoder = decoder
)

From the coaching loop, we are going to, in sure intervals, additionally name three features not reproduced right here (however obtainable within the code instance): generate_random_clothes, used to generate garments from random samples from the latent house; show_latent_space, that shows the whole take a look at set in latent (2-dimensional, thus simply visualizable) house; and show_grid, that generates garments in keeping with enter values systematically spaced out in a grid.

Let’s begin coaching! Really, earlier than we do this, let’s take a look at what these features show earlier than any coaching: As an alternative of garments, we see random pixels. Latent house has no construction. And various kinds of garments don’t cluster collectively in latent house.

Coaching loop

We’re coaching for 50 epochs right here. For every epoch, we loop over the coaching set in batches. For every batch, we observe the same old keen execution circulation: Contained in the context of a GradientTape, apply the mannequin and calculate the present loss; then outdoors this context calculate the gradients and let the optimizer carry out backprop.

What’s particular right here is that we’ve got two fashions that each want their gradients calculated and weights adjusted. This may be taken care of by a single gradient tape, offered we create it persistent.

After every epoch, we save present weights and each ten epochs, we additionally save plots for later inspection.

num_epochs <- 50

for (epoch in seq_len(num_epochs)) {
  iter <- make_iterator_one_shot(train_dataset)
  
  total_loss <- 0
  logpx_z_total <- 0
  logpz_total <- 0
  logqz_x_total <- 0
  
  until_out_of_range({
    x <-  iterator_get_next(iter)
    
    with(tf$GradientTape(persistent = TRUE) %as% tape, {
      
      c(imply, logvar) %<-% encoder(x)
      z <- reparameterize(imply, logvar)
      preds <- decoder(z)
      
      crossentropy_loss <-
        tf$nn$sigmoid_cross_entropy_with_logits(logits = preds, labels = x)
      logpx_z <-
        - k_sum(crossentropy_loss)
      logpz <-
        normal_loglik(z,
                      k_constant(0, dtype = tf$float64),
                      k_constant(0, dtype = tf$float64)
        )
      logqz_x <- normal_loglik(z, imply, logvar)
      loss <- -k_mean(logpx_z + logpz - logqz_x)
      
    })

    total_loss <- total_loss + loss
    logpx_z_total <- tf$reduce_mean(logpx_z) + logpx_z_total
    logpz_total <- tf$reduce_mean(logpz) + logpz_total
    logqz_x_total <- tf$reduce_mean(logqz_x) + logqz_x_total
    
    encoder_gradients <- tape$gradient(loss, encoder$variables)
    decoder_gradients <- tape$gradient(loss, decoder$variables)
    
    optimizer$apply_gradients(
      purrr::transpose(listing(encoder_gradients, encoder$variables)),
      global_step = tf$practice$get_or_create_global_step()
    )
    optimizer$apply_gradients(
      purrr::transpose(listing(decoder_gradients, decoder$variables)),
      global_step = tf$practice$get_or_create_global_step()
    )
    
  })
  
  checkpoint$save(file_prefix = checkpoint_prefix)
  
  cat(
    glue(
      "Losses (epoch): {epoch}:",
      "  {(as.numeric(logpx_z_total)/batches_per_epoch) %>% spherical(2)} logpx_z_total,",
      "  {(as.numeric(logpz_total)/batches_per_epoch) %>% spherical(2)} logpz_total,",
      "  {(as.numeric(logqz_x_total)/batches_per_epoch) %>% spherical(2)} logqz_x_total,",
      "  {(as.numeric(total_loss)/batches_per_epoch) %>% spherical(2)} complete"
    ),
    "n"
  )
  
  if (epoch %% 10 == 0) {
    generate_random_clothes(epoch)
    show_latent_space(epoch)
    show_grid(epoch)
  }
}

Outcomes

How effectively did that work? Let’s see the varieties of garments generated after 50 epochs.

Additionally, how disentangled (or not) are the completely different courses in latent house?

And now watch completely different garments morph into each other.

How good are these representations? That is laborious to say when there may be nothing to check with.

So let’s dive into MMD-VAE and see the way it does on the identical dataset.

MMD-VAE

MMD-VAE guarantees to generate extra informative latent options, so we’d hope to see completely different habits particularly within the clustering and morphing plots.

Information setup is similar, and there are solely very slight variations within the mannequin. Please take a look at the whole code for this instance, mmd_vae.R, as right here we’ll simply spotlight the variations.

Variations within the mannequin(s)

There are three variations as regards mannequin structure.

One, the encoder doesn’t must return the variance, so there is no such thing as a want for tf$break up. The encoder’s name methodology now simply is

perform (x, masks = NULL) {
  x %>%
    self$conv1() %>%
    self$conv2() %>%
    self$flatten() %>%
    self$dense() 
}

Between the encoder and the decoder, we don’t want the sampling step anymore, so there is no such thing as a reparameterization.
And since we gained’t use tf$nn$sigmoid_cross_entropy_with_logits to compute the loss, we let the decoder apply the sigmoid within the final deconvolution layer:

self$deconv3 <- layer_conv_2d_transpose(
  filters = 1,
  kernel_size = 3,
  strides = 1,
  padding = "identical",
  activation = "sigmoid"
)

Loss calculations

Now, as anticipated, the massive novelty is within the loss perform.

The loss, most imply discrepancy (MMD), relies on the concept that two distributions are equivalent if and provided that all moments are equivalent.
Concretely, MMD is estimated utilizing a kernel, such because the Gaussian kernel

[k(z,z’)=frac{e^z-z’}{2sigma^2}]

to evaluate similarity between distributions.

The thought then is that if two distributions are equivalent, the common similarity between samples from every distribution ought to be equivalent to the common similarity between combined samples from each distributions:

[MMD(p(z)||q(z))=E_{p(z),p(z’)}[k(z,z’)]+E_{q(z),q(z’)}[k(z,z’)]−2E_{p(z),q(z’)}[k(z,z’)]]
The next code is a direct port of the writer’s authentic TensorFlow code:

compute_kernel <- perform(x, y) {
  x_size <- k_shape(x)[1]
  y_size <- k_shape(y)[1]
  dim <- k_shape(x)[2]
  tiled_x <- k_tile(
    k_reshape(x, k_stack(listing(x_size, 1, dim))),
    k_stack(listing(1, y_size, 1))
  )
  tiled_y <- k_tile(
    k_reshape(y, k_stack(listing(1, y_size, dim))),
    k_stack(listing(x_size, 1, 1))
  )
  k_exp(-k_mean(k_square(tiled_x - tiled_y), axis = 3) /
          k_cast(dim, tf$float64))
}

compute_mmd <- perform(x, y, sigma_sqr = 1) {
  x_kernel <- compute_kernel(x, x)
  y_kernel <- compute_kernel(y, y)
  xy_kernel <- compute_kernel(x, y)
  k_mean(x_kernel) + k_mean(y_kernel) - 2 * k_mean(xy_kernel)
}

Coaching loop

The coaching loop differs from the usual VAE instance solely within the loss calculations.
Listed here are the respective strains:

 with(tf$GradientTape(persistent = TRUE) %as% tape, {
      
      imply <- encoder(x)
      preds <- decoder(imply)
      
      true_samples <- k_random_normal(
        form = c(batch_size, latent_dim),
        dtype = tf$float64
      )
      loss_mmd <- compute_mmd(true_samples, imply)
      loss_nll <- k_mean(k_square(x - preds))
      loss <- loss_nll + loss_mmd
      
    })

So we merely compute MMD loss in addition to reconstruction loss, and add them up. No sampling is concerned on this model.
After all, we’re curious to see how effectively that labored!

Outcomes

Once more, let’s take a look at some generated garments first. It looks like edges are a lot sharper right here.

The clusters too look extra properly unfold out within the two dimensions. And, they’re centered at (0,0), as we’d have hoped for.

Lastly, let’s see garments morph into each other. Right here, the graceful, steady evolutions are spectacular!
Additionally, practically all house is crammed with significant objects, which hasn’t been the case above.

MNIST

For curiosity’s sake, we generated the identical sorts of plots after coaching on authentic MNIST.
Right here, there are hardly any variations seen in generated random digits after 50 epochs of coaching.

Additionally the variations in clustering should not that huge.

However right here too, the morphing seems to be way more natural with MMD-VAE.

Conclusion

To us, this demonstrates impressively what huge a distinction the price perform could make when working with VAEs.
One other part open to experimentation would be the prior used for the latent house – see this discuss for an summary of different priors and the “Variational Combination of Posteriors” paper (Tomczak and Welling 2017) for a preferred current method.

For each price features and priors, we anticipate efficient variations to grow to be manner greater nonetheless after we go away the managed setting of (Style) MNIST and work with real-world datasets.