[ad_1]
About two weeks in the past, we launched TensorFlow Chance (TFP), exhibiting find out how to create and pattern from distributions and put them to make use of in a Variational Autoencoder (VAE) that learns its prior. At present, we transfer on to a unique specimen within the VAE mannequin zoo: the Vector Quantised Variational Autoencoder (VQ-VAE) described in Neural Discrete Illustration Studying (Oord, Vinyals, and Kavukcuoglu 2017). This mannequin differs from most VAEs in that its approximate posterior just isn’t steady, however discrete – therefore the “quantised” within the article’s title. We’ll rapidly take a look at what this implies, after which dive straight into the code, combining Keras layers, keen execution, and TFP.
Many phenomena are finest considered, and modeled, as discrete. This holds for phonemes and lexemes in language, higher-level constructions in pictures (assume objects as a substitute of pixels),and duties that necessitate reasoning and planning.
The latent code utilized in most VAEs, nevertheless, is steady – normally it’s a multivariate Gaussian. Steady-space VAEs have been discovered very profitable in reconstructing their enter, however usually they endure from one thing referred to as posterior collapse: The decoder is so highly effective that it might create life like output given simply any enter. This implies there is no such thing as a incentive to study an expressive latent area.
In VQ-VAE, nevertheless, every enter pattern will get mapped deterministically to one in all a set of embedding vectors. Collectively, these embedding vectors represent the prior for the latent area.
As such, an embedding vector comprises much more info than a imply and a variance, and thus, is far tougher to disregard by the decoder.
The query then is: The place is that magical hat, for us to drag out significant embeddings?
From the above conceptual description, we now have two inquiries to reply. First, by what mechanism can we assign enter samples (that went via the encoder) to acceptable embedding vectors?
And second: How can we study embedding vectors that really are helpful representations – that when fed to a decoder, will end in entities perceived as belonging to the identical species?
As regards project, a tensor emitted from the encoder is just mapped to its nearest neighbor in embedding area, utilizing Euclidean distance. The embedding vectors are then up to date utilizing exponential transferring averages. As we’ll see quickly, which means that they’re really not being realized utilizing gradient descent – a characteristic price stating as we don’t come throughout it day-after-day in deep studying.
Concretely, how then ought to the loss perform and coaching course of look? This may in all probability best be seen in code.
The entire code for this instance, together with utilities for mannequin saving and picture visualization, is obtainable on github as a part of the Keras examples. Order of presentation right here could differ from precise execution order for expository functions, so please to truly run the code take into account making use of the instance on github.
As in all our prior posts on VAEs, we use keen execution, which presupposes the TensorFlow implementation of Keras.
As in our earlier submit on doing VAE with TFP, we’ll use Kuzushiji-MNIST(Clanuwat et al. 2018) as enter.
Now’s the time to take a look at what we ended up producing that point and place your guess: How will that evaluate in opposition to the discrete latent area of VQ-VAE?
np <- import("numpy")
kuzushiji <- np$load("kmnist-train-imgs.npz")
kuzushiji <- kuzushiji$get("arr_0")
train_images <- kuzushiji %>%
k_expand_dims() %>%
k_cast(dtype = "float32")
train_images <- train_images %>% `/`(255)
buffer_size <- 60000
batch_size <- 64
num_examples_to_generate <- batch_size
batches_per_epoch <- buffer_size / batch_size
train_dataset <- tensor_slices_dataset(train_images) %>%
dataset_shuffle(buffer_size) %>%
dataset_batch(batch_size, drop_remainder = TRUE)
Hyperparameters
Along with the “traditional” hyperparameters we’ve in deep studying, the VQ-VAE infrastructure introduces just a few model-specific ones. Initially, the embedding area is of dimensionality variety of embedding vectors instances embedding vector dimension:
# variety of embedding vectors
num_codes <- 64L
# dimensionality of the embedding vectors
code_size <- 16L
The latent area in our instance will likely be of dimension one, that’s, we’ve a single embedding vector representing the latent code for every enter pattern. This will likely be positive for our dataset, however it ought to be famous that van den Oord et al. used far higher-dimensional latent areas on e.g. ImageNet and Cifar-10.
Encoder mannequin
The encoder makes use of convolutional layers to extract picture options. Its output is a three-D tensor of form batchsize * 1 * code_size.
activation <- "elu"
# modularizing the code just a bit bit
default_conv <- set_defaults(layer_conv_2d, record(padding = "identical", activation = activation))
base_depth <- 32
encoder_model <- perform(identify = NULL,
code_size) {
keras_model_custom(identify = identify, perform(self) {
self$conv1 <- default_conv(filters = base_depth, kernel_size = 5)
self$conv2 <- default_conv(filters = base_depth, kernel_size = 5, strides = 2)
self$conv3 <- default_conv(filters = 2 * base_depth, kernel_size = 5)
self$conv4 <- default_conv(filters = 2 * base_depth, kernel_size = 5, strides = 2)
self$conv5 <- default_conv(filters = 4 * latent_size, kernel_size = 7, padding = "legitimate")
self$flatten <- layer_flatten()
self$dense <- layer_dense(models = latent_size * code_size)
self$reshape <- layer_reshape(target_shape = c(latent_size, code_size))
perform (x, masks = NULL) {
x %>%
# output form: 7 28 28 32
self$conv1() %>%
# output form: 7 14 14 32
self$conv2() %>%
# output form: 7 14 14 64
self$conv3() %>%
# output form: 7 7 7 64
self$conv4() %>%
# output form: 7 1 1 4
self$conv5() %>%
# output form: 7 4
self$flatten() %>%
# output form: 7 16
self$dense() %>%
# output form: 7 1 16
self$reshape()
}
})
}
As all the time, let’s make use of the truth that we’re utilizing keen execution, and see just a few instance outputs.
iter <- make_iterator_one_shot(train_dataset)
batch <- iterator_get_next(iter)
encoder <- encoder_model(code_size = code_size)
encoded <- encoder(batch)
encoded
tf.Tensor(
[[[ 0.00516277 -0.00746826 0.0268365 ... -0.012577 -0.07752544
-0.02947626]]
...
[[-0.04757921 -0.07282603 -0.06814402 ... -0.10861694 -0.01237121
0.11455103]]], form=(64, 1, 16), dtype=float32)
Now, every of those 16d vectors must be mapped to the embedding vector it’s closest to. This mapping is taken care of by one other mannequin: vector_quantizer
.
Vector quantizer mannequin
That is how we are going to instantiate the vector quantizer:
vector_quantizer <- vector_quantizer_model(num_codes = num_codes, code_size = code_size)
This mannequin serves two functions: First, it acts as a retailer for the embedding vectors. Second, it matches encoder output to obtainable embeddings.
Right here, the present state of embeddings is saved in codebook
. ema_means
and ema_count
are for bookkeeping functions solely (be aware how they’re set to be non-trainable). We’ll see them in use shortly.
vector_quantizer_model <- perform(identify = NULL, num_codes, code_size) {
keras_model_custom(identify = identify, perform(self) {
self$num_codes <- num_codes
self$code_size <- code_size
self$codebook <- tf$get_variable(
"codebook",
form = c(num_codes, code_size),
dtype = tf$float32
)
self$ema_count <- tf$get_variable(
identify = "ema_count", form = c(num_codes),
initializer = tf$constant_initializer(0),
trainable = FALSE
)
self$ema_means = tf$get_variable(
identify = "ema_means",
initializer = self$codebook$initialized_value(),
trainable = FALSE
)
perform (x, masks = NULL) {
# to be crammed in shortly ...
}
})
}
Along with the precise embeddings, in its name
technique vector_quantizer
holds the project logic.
First, we compute the Euclidean distance of every encoding to the vectors within the codebook (tf$norm
).
We assign every encoding to the closest as by that distance embedding (tf$argmin
) and one-hot-encode the assignments (tf$one_hot
). Lastly, we isolate the corresponding vector by masking out all others and summing up what’s left over (multiplication adopted by tf$reduce_sum
).
Relating to the axis
argument used with many TensorFlow capabilities, please take into accounts that in distinction to their k_*
siblings, uncooked TensorFlow (tf$*
) capabilities count on axis numbering to be 0-based. We even have so as to add the L
’s after the numbers to adapt to TensorFlow’s datatype necessities.
vector_quantizer_model <- perform(identify = NULL, num_codes, code_size) {
keras_model_custom(identify = identify, perform(self) {
# right here we've the above occasion fields
perform (x, masks = NULL) {
# form: bs * 1 * num_codes
distances <- tf$norm(
tf$expand_dims(x, axis = 2L) -
tf$reshape(self$codebook,
c(1L, 1L, self$num_codes, self$code_size)),
axis = 3L
)
# bs * 1
assignments <- tf$argmin(distances, axis = 2L)
# bs * 1 * num_codes
one_hot_assignments <- tf$one_hot(assignments, depth = self$num_codes)
# bs * 1 * code_size
nearest_codebook_entries <- tf$reduce_sum(
tf$expand_dims(
one_hot_assignments, -1L) *
tf$reshape(self$codebook, c(1L, 1L, self$num_codes, self$code_size)),
axis = 2L
)
record(nearest_codebook_entries, one_hot_assignments)
}
})
}
Now that we’ve seen how the codes are saved, let’s add performance for updating them.
As we mentioned above, they aren’t realized through gradient descent. As an alternative, they’re exponential transferring averages, frequently up to date by no matter new “class member” they get assigned.
So here’s a perform update_ema
that can maintain this.
update_ema
makes use of TensorFlow moving_averages to
- first, hold observe of the variety of presently assigned samples per code (
updated_ema_count
), and - second, compute and assign the present exponential transferring common (
updated_ema_means
).
moving_averages <- tf$python$coaching$moving_averages
# decay to make use of in computing exponential transferring common
decay <- 0.99
update_ema <- perform(
vector_quantizer,
one_hot_assignments,
codes,
decay) {
updated_ema_count <- moving_averages$assign_moving_average(
vector_quantizer$ema_count,
tf$reduce_sum(one_hot_assignments, axis = c(0L, 1L)),
decay,
zero_debias = FALSE
)
updated_ema_means <- moving_averages$assign_moving_average(
vector_quantizer$ema_means,
# selects all assigned values (masking out the others) and sums them up over the batch
# (will likely be divided by depend later, so we get a mean)
tf$reduce_sum(
tf$expand_dims(codes, 2L) *
tf$expand_dims(one_hot_assignments, 3L), axis = c(0L, 1L)),
decay,
zero_debias = FALSE
)
updated_ema_count <- updated_ema_count + 1e-5
updated_ema_means <- updated_ema_means / tf$expand_dims(updated_ema_count, axis = -1L)
tf$assign(vector_quantizer$codebook, updated_ema_means)
}
Earlier than we take a look at the coaching loop, let’s rapidly full the scene including within the final actor, the decoder.
Decoder mannequin
The decoder is fairly commonplace, performing a sequence of deconvolutions and eventually, returning a likelihood for every picture pixel.
default_deconv <- set_defaults(
layer_conv_2d_transpose,
record(padding = "identical", activation = activation)
)
decoder_model <- perform(identify = NULL,
input_size,
output_shape) {
keras_model_custom(identify = identify, perform(self) {
self$reshape1 <- layer_reshape(target_shape = c(1, 1, input_size))
self$deconv1 <-
default_deconv(
filters = 2 * base_depth,
kernel_size = 7,
padding = "legitimate"
)
self$deconv2 <-
default_deconv(filters = 2 * base_depth, kernel_size = 5)
self$deconv3 <-
default_deconv(
filters = 2 * base_depth,
kernel_size = 5,
strides = 2
)
self$deconv4 <-
default_deconv(filters = base_depth, kernel_size = 5)
self$deconv5 <-
default_deconv(filters = base_depth,
kernel_size = 5,
strides = 2)
self$deconv6 <-
default_deconv(filters = base_depth, kernel_size = 5)
self$conv1 <-
default_conv(filters = output_shape[3],
kernel_size = 5,
activation = "linear")
perform (x, masks = NULL) {
x <- x %>%
# output form: 7 1 1 16
self$reshape1() %>%
# output form: 7 7 7 64
self$deconv1() %>%
# output form: 7 7 7 64
self$deconv2() %>%
# output form: 7 14 14 64
self$deconv3() %>%
# output form: 7 14 14 32
self$deconv4() %>%
# output form: 7 28 28 32
self$deconv5() %>%
# output form: 7 28 28 32
self$deconv6() %>%
# output form: 7 28 28 1
self$conv1()
tfd$Unbiased(tfd$Bernoulli(logits = x),
reinterpreted_batch_ndims = size(output_shape))
}
})
}
input_shape <- c(28, 28, 1)
decoder <- decoder_model(input_size = latent_size * code_size,
output_shape = input_shape)
Now we’re prepared to coach. One factor we haven’t actually talked about but is the associated fee perform: Given the variations in structure (in comparison with commonplace VAEs), will the losses nonetheless look as anticipated (the same old add-up of reconstruction loss and KL divergence)?
We’ll see that in a second.
Coaching loop
Right here’s the optimizer we’ll use. Losses will likely be calculated inline.
optimizer <- tf$prepare$AdamOptimizer(learning_rate = learning_rate)
The coaching loop, as traditional, is a loop over epochs, the place every iteration is a loop over batches obtained from the dataset.
For every batch, we’ve a ahead go, recorded by a gradientTape
, based mostly on which we calculate the loss.
The tape will then decide the gradients of all trainable weights all through the mannequin, and the optimizer will use these gradients to replace the weights.
Up to now, all of this conforms to a scheme we’ve oftentimes seen earlier than. One level to notice although: On this identical loop, we additionally name update_ema
to recalculate the transferring averages, as these are usually not operated on throughout backprop.
Right here is the important performance:
num_epochs <- 20
for (epoch in seq_len(num_epochs)) {
iter <- make_iterator_one_shot(train_dataset)
until_out_of_range({
x <- iterator_get_next(iter)
with(tf$GradientTape(persistent = TRUE) %as% tape, {
# do ahead go
# calculate losses
})
encoder_gradients <- tape$gradient(loss, encoder$variables)
decoder_gradients <- tape$gradient(loss, decoder$variables)
optimizer$apply_gradients(purrr::transpose(record(
encoder_gradients, encoder$variables
)),
global_step = tf$prepare$get_or_create_global_step())
optimizer$apply_gradients(purrr::transpose(record(
decoder_gradients, decoder$variables
)),
global_step = tf$prepare$get_or_create_global_step())
update_ema(vector_quantizer,
one_hot_assignments,
codes,
decay)
# periodically show some generated pictures
# see code on github
# visualize_images("kuzushiji", epoch, reconstructed_images, random_images)
})
}
Now, for the precise motion. Contained in the context of the gradient tape, we first decide which encoded enter pattern will get assigned to which embedding vector.
codes <- encoder(x)
c(nearest_codebook_entries, one_hot_assignments) %<-% vector_quantizer(codes)
Now, for this project operation there is no such thing as a gradient. As an alternative what we are able to do is go the gradients from decoder enter straight via to encoder output.
Right here tf$stop_gradient
exempts nearest_codebook_entries
from the chain of gradients, so encoder and decoder are linked by codes
:
codes_straight_through <- codes + tf$stop_gradient(nearest_codebook_entries - codes)
decoder_distribution <- decoder(codes_straight_through)
In sum, backprop will maintain the decoder’s in addition to the encoder’s weights, whereas the latent embeddings are up to date utilizing transferring averages, as we’ve seen already.
Now we’re able to deal with the losses. There are three elements:
- First, the reconstruction loss, which is simply the log likelihood of the particular enter underneath the distribution realized by the decoder.
reconstruction_loss <- -tf$reduce_mean(decoder_distribution$log_prob(x))
- Second, we’ve the dedication loss, outlined because the imply squared deviation of the encoded enter samples from the closest neighbors they’ve been assigned to: We would like the community to “commit” to a concise set of latent codes!
commitment_loss <- tf$reduce_mean(tf$sq.(codes - tf$stop_gradient(nearest_codebook_entries)))
- Lastly, we’ve the same old KL diverge to a previous. As, a priori, all assignments are equally possible, this part of the loss is fixed and may oftentimes be disbursed of. We’re including it right here primarily for illustrative functions.
prior_dist <- tfd$Multinomial(
total_count = 1,
logits = tf$zeros(c(latent_size, num_codes))
)
prior_loss <- -tf$reduce_mean(
tf$reduce_sum(prior_dist$log_prob(one_hot_assignments), 1L)
)
Summing up all three elements, we arrive on the general loss:
beta <- 0.25
loss <- reconstruction_loss + beta * commitment_loss + prior_loss
Earlier than we take a look at the outcomes, let’s see what occurs inside gradientTape
at a single look:
with(tf$GradientTape(persistent = TRUE) %as% tape, {
codes <- encoder(x)
c(nearest_codebook_entries, one_hot_assignments) %<-% vector_quantizer(codes)
codes_straight_through <- codes + tf$stop_gradient(nearest_codebook_entries - codes)
decoder_distribution <- decoder(codes_straight_through)
reconstruction_loss <- -tf$reduce_mean(decoder_distribution$log_prob(x))
commitment_loss <- tf$reduce_mean(tf$sq.(codes - tf$stop_gradient(nearest_codebook_entries)))
prior_dist <- tfd$Multinomial(
total_count = 1,
logits = tf$zeros(c(latent_size, num_codes))
)
prior_loss <- -tf$reduce_mean(tf$reduce_sum(prior_dist$log_prob(one_hot_assignments), 1L))
loss <- reconstruction_loss + beta * commitment_loss + prior_loss
})
Outcomes
And right here we go. This time, we are able to’t have the 2nd “morphing view” one usually likes to show with VAEs (there simply isn’t any 2nd latent area). As an alternative, the 2 pictures beneath are (1) letters generated from random enter and (2) reconstructed precise letters, every saved after coaching for 9 epochs.
Two issues soar to the attention: First, the generated letters are considerably sharper than their continuous-prior counterparts (from the earlier submit). And second, would you’ve been capable of inform the random picture from the reconstruction picture?
At this level, we’ve hopefully satisfied you of the facility and effectiveness of this discrete-latents strategy.
Nonetheless, you may secretly have hoped we’d apply this to extra advanced information, akin to the weather of speech we talked about within the introduction, or higher-resolution pictures as present in ImageNet.
The reality is that there’s a steady tradeoff between the variety of new and thrilling strategies we are able to present, and the time we are able to spend on iterations to efficiently apply these strategies to advanced datasets. Ultimately it’s you, our readers, who will put these strategies to significant use on related, actual world information.
[ad_2]