It's been a long time coming but I'm finally getting this post out! I read this paper a couple of years ago and wanted to really understand it because it was state of the art at the time (still pretty close even now). As usual though, once I started down the variational autoencoder line of posts, there was always yet another VAE paper to look into so I never got around to looking at this one.

This post is all about a proper probabilistic generative model called Pixel Convolutional Neural Networks or PixelCNN. It was originally proposed as a side contribution of Pixel Recurrent Neural Networks in [1] and later expanded upon in [2,3] (and I'm sure many other papers). The real cool thing about it is that it's (a) probabilistic, and (b) autoregressive. It's still counter-intuitive to me that you can generate images one pixel at at time, but I'm jumping ahead of myself here. We'll go over some background material, the method, and my painstaking attempts at an implementation (and what I learned from it). Let's get started!

Autoregressive Generative Models

Before we begin, we should review autoregressive generative models. I'll basically summarize what I wrote in one of my previous post: Autoregressive Autoencoders.

An Autoregressive model is usually used in the context of time-series modelling (of random processes) where \(y_n\) depends on \(y_{n-1}\) or some earlier value. In particular, literature usually assume a linear dependence and name these "AR" models of the "ARIMA" notoriety. Here "auto" refers to self, and "regressive" means regressed against.

In the context of deep generative models, we'll drop the condition of linear dependence and formulate our image problem as a random process. In particular, we will:

  1. Use a deep generative models (obviously non-linear), and

  2. Assume the pixels of an image are random variables with a specific ordering (top to bottom, left to right), which formulates it as a random process.

With that in mind, let's review the product rule:

\begin{equation*} p({\bf x}) = \prod_{i=1}^{D} p(x_i | {\bf x}_{<i}) \tag{1} \end{equation*}

where \({\bf x}_{<i} = [x_1, \ldots, x_{i-1}]\). Basically, component \(i\) of \({\bf x}\) only depends on the dimensions of \(j < i\). In your head, you can think of each \(x_i\) as a pixel. So each pixel is going to have a probability distribution that is a function of all the (sub-)pixels that came before it (for RGB images, each of "R", "G", "B" are treated as separate sub-pixels).

The way to generate an image from an autoregressive generative model is as follows:

  1. Naturally, the first (sub-)pixel in our sequence has nothing before it so it's a pure unconditional distribution. We simply sample from this distribution to get a concrete realization for the first (sub-)pixel.

  2. Each subsequent (sub-)pixel distribution is generated in sequence conditioned on all (or a subset of) previously sampled (sub-)pixels. We simply sample from this conditional distribution to get the current pixel value.

  3. Repeat until you have the entire image.

According to my intuition, this is a really weird way to generate an image! Think of it, if you want to generate a picture of a dog, you start at the top left pixel, figure out what it is, then move on to the one beside it, and so on, until somehow you have an image of a dog? This goes against my human sensibilities of generating images where I implicitly have a concept of a dog and recursively fill in broad regions of the image getting to finer detail as I go along. In any case, we can still get good negative log-likelihood (although the quality of the images are another story), so that's all that counts (?).


Now that we covered autoregressive generative models, PixelCNN is not too difficult to understand. We want to build a single CNN that takes as input an image and outputs a distribution for each (sub-)pixel (theoretically, you could have a different network for each pixel but that seems inefficient). There are a couple subtleties when doing that:

  1. Due to the autoregressive nature, output pixel \(i\) should not see any input pixels \(\geq i\), otherwise it wouldn't be autoregressive (you could "see the future").

  2. Selecting a distribution and its corresponding loss function in order to model the output pixels.

Let's take a look at each one separately.

Masked Convolution

The masked convolution is basically the same idea as the masked autoencoder from my post on MADE: Autoregressive Autoencoders. As stated above, we impose (an arbitrary) ordering on the sub-pixels: top to bottom, left to right, R to G to B. Given this, we want to make sure input pixels \(\geq i\) are "hidden" from output pixel \(i\), and we can accomplish this with masks. This is shown on the left side of Figure 1 where output pixel \(i\) only "reads" from its predecessors (the center image is the same thing for a larger multi-scale resolution). This can easily be generated using a "mask" implemented as an element-wise multiplication with a convolution kernel. We'll talk about the implementation of this later on.

PixelCNN Mask

Figure 1: PixelCNN Mask (source: [1])

However, the mask doesn't just deal with the spatial location of full pixels, it also has to take into account the RGB values, which leads us to two different types of masks: A and B. For the first convolution layer on the original image, the same rule applies as above, never read ahead. For full pixel \(i\)'s mask when reading from pixel \(x_i\), the "B" pixel should only be connected to "G" and "R"; the "G" pixel should only be connected to "R"; and the "R" pixel shouldn't be connected at all to pixel \(i\). You can see the connectivity on the right side of Figure 1 (Note you still have full "read" access to all sub-pixels from predecessors, the differences for these masks only affect the current pixel).

For layers other than the first one, things change. Since any sub-pixel output from a convolution layer has already been masked from it's corresponding input sub-pixel, you are free to use it in another convolution. That might be a bit confusing but you can see this in Mask B in Figure 1. For example, the output of the "G" sub-pixel in Mask B, depends on the "G" sub-pixel output from Mask A, which in turn depends only on the "R". That means "G" from the second layer only depends on "R" from the original input, which is what we wanted. If we didn't do this, the "G" output from Mask B would never be able to "read" from the "R" sub-pixel in the original input. If this is just confusing, just take a minute to study the connectivity in the diagram and I think it should be pretty clear.

Keep in mind Figure 1 really only shows the case where you have three color channels. In all convolution layers beyond the first one, we likely have many different filters per layer. Just imagine instead of "RGB", we have "RRRRGGGGBBBB" as input to Mask B, and you can probably guess what the connectivity should look like. Again this only applies to the current sub-pixels, we should have full connectivity to all spatial predecessors.

Discretized Logistic Mixture Likelihood

I covered this topic in my previous post in detail (Importance Sampling and Estimating Marginal Likelihood in Variational Autoencoders) so I'll just do a summary here.

Modelling pixels is a funny problem. On one hand, we can view them as 8-bit discrete random variables (e.g. 256-way softmax), but on the other hand, we can view them as a sort-of continuous distribution (e.g. real-valued output). The former approach is the way that [1] tackles the problem: just stick a 256-softmax output on each sub-pixel. There are two issues: (1) it uses up a gigantic amount of resources, and (2) there is no relationship between adjacent pixel values (R=100 should be similar to R=101). Mostly due to the first reason, I opted out to not implement it (couldn't fit a full image on my meager GTX 1070).

The other way to go about it is to imagine the pixel generation process as such described in [2]:

  1. For each sub-pixel, generate a continuous distribution \(\nu\) representing the intensity. For example, \(\nu\) could be a logistic distribution parameterized by \(\mu, s\).

  2. Next, "round" each sub-pixel to a discretized distribution over \([0, 255] \in \mathbb{Z}\) by integrating over the appropriate width along the real line (assuming a logistic distribution):

    \begin{equation*} P(x|\mu,s) = \begin{cases} \sigma(\frac{x-\mu+0.5}{s}) & \text{for } x = 0 \\ \sigma(\frac{x-\mu+0.5}{s}) - \sigma(\frac{x-\mu-0.5}{s}) & \text{for } 0 < x < 255 \\ 1 - \sigma(\frac{x-\mu-0.5}{s}) & \text{for } x = 255 \end{cases} \tag{2} \end{equation*}

    where \(\sigma\) is the sigmoid function (recall sigmoid is the CDF of the logistic distribution). Here we basically take the \(\pm 0.5\) interval around each pixel value to compute its discretized probability mass. For the edges, we just integrate to infinity.

However, this will make each pixel uni-modal, which doesn't afford us much flexibility. To improve it, we use a mixture of logistics for \(\nu\):

\begin{equation*} \nu \sim \sum_{i=1}^K \pi_i logistic(\mu_i, s_i) \tag{3} \end{equation*}

where \(\pi_i\) is the categorical weight. To bring it back to neural networks, for each sub-pixel, we want our network to output three things:

  • \(K\) \(\mu_i\)'s representing the centers of our logistic distributions

  • \(K\) \(s_i\)'s representing the scale of our logistic distributions

  • \(K\) \(\pi_i\)'s representing the mixture weights (summing to 1)

We don't need that many mixture components to get a good result ([2] uses \(K=5\)), which is a lot less than a 256-way softmax. Figure 2 shows a realization of a toy discretized mixture I did while testing. You can see clearly that we can model multi-modal distributions, and at the same time have a lot of mass at the 0 and 255 pixels (which [2] claim is a good thing because black and white occur a lot in images?).

Distribution of Discretized Logistic Mixtures

Figure 2: Distribution of Discretized Logistic Mixtures: Top to bottom represent the "R", "G", "B" components of a single pixel

All of this can be implemented pretty easily by strapping a bunch of masked convolution layers together and setting the last layer to have \(3*3K\) filters (one for each sub-pixel). The really interesting stuff happens in the loss function though which computes the negative log of Equation 2. The nice thing about this loss is that it's fully probabilistic from the start. We'll get to implementing it in a later section but suffice it to say, it's not easy dealing with overflow!

Training and Generating Samples

Training this network is actually pretty easy. All we have to do is make the actual image available on the input and output of the network. Note the input tensor of the network takes an image, while the output of the network outputs a distribution for each sub-pixel. You still use the image on both the input and output but use a custom loss function that computes the negative log-likelihood using the network outputs. Other than the complexity of the loss function, you just set it and go!

Generating images is something that is a bit more complicated but follows the same idea from my post on Autoregressive Autoencoders:

  1. Set \(i=0\) to represent current iteration and pixel (implicitly we would translate it to the row/col/sub-pixel in the image).

  2. Start off with a tensor for your image \(\bf x^0\) with any initialization (it doesn't matter).

  3. Feed \(\bf x^i\) into the PixelCNN network to generate distributional outputs \(\bf y^{i+1}\).

  4. Randomly sample sub-pixel \(u_{i+1}\) from the mixture distribution defined by \(\bf y^{i+1}\) (we only need the subset of values for sub-pixel \(i+1\)).

  5. Set \(\bf x^{i+1}\) as \(\bf x^i\) but replacing the current sub-pixel with \(u_{i+1}\).

  6. Repeat step 2-4 until entire image is generated.

From this algorithm, you can see that we're generating one pixel at a time, a very slow process! For a 32x32x3 image, we basically need to do a forward pass of the network 3072 times for a single image (we have some parallelism because we can do several images in batch but of course we have the "9 women making 9 babies in a month" problem). This is the downside of autoregressive models: training is done in parallel (for convolutional autoregressive models) but generation is sequential (and slow). As a data point, my slow implementation took almost 37 mins to generate 16 images (forward passes were parallelized on the GPU but sampling was sequential in a loop on the CPU).

Implementation Details

So far the theory isn't too bad: some masked layers, extra outputs and some sigmoid losses and we're done, right? Well it's a bit trickier than that, especially the loss function. I'll explain the details (and headaches) that I went through implementing it in Keras. As usual, you can find all my code in this Github repo.

Masked Convolution Layer

The masked convolution layer (which I named PixelConv2D) was actually pretty easy to implement in Keras because I just inherited from the Conv2D layer, build a binary mask and then did an element-wise product with the kernel. There's just a bit of accounting that needs to go on in building the mask such as ensuring that your input is a multiple of 3 and that the right bits are set. This probably isn't the most efficient method of doing it because you literally are multiplying by a binary matrix every time, but it probably is the easiest.

PixelCNN Outputs

The output of the network is a distribution, but how is that realized? The last layer is composed is made up of three sets of PixelConv2D layers representing:

  • Logistic mean values \(\mu\), filters = # of mixture component, no activation

  • Logistic log of inverse scale values \(s\), filters = # of mixture components, "softplus" activation function

  • Pre-softmax mixture inputs, filters = # of mixture components, no activation

The mean is pretty straight forward. We put no restriction on it being in our normalized pixel interval (i.e. \([-1, 1]\)), which seems to work out fine.

The network output corresponding to scale is set to be an inverse because we never want to divide by 0. As for modelling the output as the logarithm, I suspect (but haven't observed) that it's just a better match for neural network output ranges. For example, your network needs to output \(6\) instead of \(e^6\), where the latter will have to have huge weights on the last layer. The "softplus" seems to be the best fit here because it's very smooth (unlike "ReLU"), and the non-negative logarithm values ensure \(s < 1\). Since we're dealing with normalized pixels between \([-1, 1]\), we would never want a shape parameter wider than half the interval (that would just put almost all the mass on the end points). In fact, I've observed log inverse scale to be in the 3-5 range, meaning the shape parameter is as small as \(\frac{1}{e^5}\).

Finally, the network's output corresponding to the mixture components are the inputs to the softmax, essentially the "pre-softmax". This is done because in the loss function we compute the \(\log\) of the softmax, which is numerically more stable to compute if we have the raw pre-softmax inputs rather than the direct softmax outputs. It's a small change and just requires a few extra processing steps when we're actually generating images to get the mixture weights.

Network Architecture

I used the same architecture as the PixelCNN paper [1] except for the outputs where I used logistic mixture outputs instead of a softmax (as described above):

  • 5x5 Conv Layer, Mask A, ReLU

  • 15 - 3x3 Resnet Blocks, Mask B, \(h=128\) (shown in Figure 3)

  • 2 - 1x1 Conv layers, Mask B, ReLU, \(filters=1024\)

  • Output layers (as described above)

PixelCNN Resnet Block

Figure 3: PixelCNN Resnet Block (source [1])

The network widths above are for each colour channel, which are concatenated after each operation and fed into the next PixelConv2D which know how to deal with them via masks.

Loss Function

The loss function was definitely the hardest part of the entire implementation. There are so many subtleties, I don't know where to begin. And this was after I heavily referenced the PixelCNN++ code [4]. Let's start with computing the log-likelihood shown in Equation 2. There are actually 4 different cases (actual condition in parenthesis when scaled to pixel range \([-1,1]\)).

Note: I'm using \(0.5\) in the equations below but substitute \(\frac{1}{2(127.5)}\) for the rescaled pixel range \([-1,1]\).

Case 1 Black Pixel: \(x\leq 0\) (\(x < -0.999\))

Here we just need to do a bit of math to simplify the expression:

\begin{equation*} \log\big( \sigma(\frac{x-\mu+0.5}{s}) \big) = \frac{x-\mu+0.5}{s} - \text{softplus}(\frac{x-\mu+0.5}{s}) \tag{4} \end{equation*}

where softplus is defined as \(\log(e^x+1)\), see this Math Stack Exchange Question for more details.

Case 2 White Pixel: \(x\geq 255\) (\(x > 0.999\))

Again, simply a simplification of Equation 2:

\begin{equation*} \log\big(1 - \sigma(\frac{x-\mu-0.5}{s}) \big) = -\text{softplus}(\frac{x-\mu-0.5}{s}) \tag{5} \end{equation*}

If you just expand out the sigmoid into exponentials, this should be a pretty easy to derive.

Case 3 Overflow Condition: \(\sigma(\frac{x-\mu+0.5}{s}) - \sigma(\frac{x-\mu-0.5}{s}) < 10^{-5}\)

This is the part where we have to be careful. Even if we don't have a black or white pixel, we can still overflow. Imagine the case where the centre of our logistic distribution is way off from our pixel range e.g. \(\mu=1000, s=1\). This means that any pixel \(x\) within the \([-1, 1]\) range will be incredibly close to \(0\) (remember we don't have infinite precision) since it's so far out in the tail of the distribution. As such, the difference will also be nearly zero and when we try to take the logarithm, we get \(-\infty\) or NaNs.

Interestingly enough, the code from PixelCNN++ [4] says that this condition doesn't occur in their code, but for me it definitely happens. I took this condition out and I started getting NaNs everywhere. It's possible with their architecture it doesn't happen but I suspect either their comment is misleading or they just didn't do a lot of checks.

Anyways, to solve this problem, we actually approximate the integral (area under the PDF) by taking the centered PDF of the logistic and multiply it by a pixel width interval (\(\frac{1}{127.5}\) is a pixel width interval in our [-1, 1] range):

\begin{align*} \log(\text{PDF} \cdot \frac{1}{127.5}) &= \log\big(\frac{e^{-(x-m)/s}}{{s(1 + e^{-(x-m)/s})^2}}\big)-\log(127.5) \\ &= -\frac{x-m}{s} - \log(s) - 2\log(1+e^{-(x-m)/s}) - \log(127.5) \\ &= -\frac{x-m}{s} - \log(s) - 2\cdot\text{softplus}(-\frac{x-m}{s}) - \log(127.5) \\ \tag{6} \end{align*}

If that weren't enough, I did some extra work to make it even more precise! This was not in the original implementation [4], and actually in retrospect, I'm not sure if it even helps but I'm going to explain it anyways since I spent a bunch of time on it.

For Equation 6, it obviously is not equivalent to the actual expression \(\log\big(\sigma(\frac{x+0.5-\mu}{s}) - \sigma(\frac{x-0.5-\mu}{s})\big)\). So at the cross over point of \(10^{-5}\), there must be some sort of discontinuity in the loss. This is shown in Figure 4.

PixelCNN Loss Discontinuity for Edge Case

Figure 4: PixelCNN Loss Discontinuity for Edge Case

For various values of \(\text{invs}=\log(\frac{1}{e^s})\), I plotted the discontinuity as a function of the centered x values. The dotted line represents the cross over point. When you are very far away from the center (larger x values), the exception case kicks in, but when we get closer to being in the right region, we get the normal case. As we get a tighter distribution (larger invs), the point at which the exception case kicks in is sooner. The problem is the exception case has a smaller loss than the normal case, which means it's possible that it might flip/flop between the two cases.

To adjust for it, I added a line of best fit as a function of invs. Figure 5 shows this line of best fit (never mind the axis label where I'm using inconsistent naming).

PixelCNN Loss Adjustment for Edge Case

Figure 5: PixelCNN Loss Adjustment for Edge Case

As you can see, the line of best fit is pretty good up to \(\text{invs}\approx 6\). In most cases, I haven't observed such a tight distribution on the outputs so it probably will serve us pretty well. As I mentioned above, I'm actually kind of skeptical that it makes a difference but here it is anyways.

Case 4 Standard Scenario: Otherwise

We simply just directly compute the logarithm of the middle case of Equation 2 with a protection on the difference:

\begin{equation*} \log\big(\text{max}(\sigma(\frac{x-\mu+0.5}{s}) - \sigma(\frac{x-\mu-0.5}{s}), e^{-12})\big) \tag{7} \end{equation*}

I'm not sure if the max operation is needed here but the code in [4] says it helps deal with problems computing the gradients in Tensorflow. For example, when you are close to 0, even though this case shouldn't occur because it's protected using a tf.where statement for Case 3, the logarithm in this branch may still be computed and generate a NaN, which could screw up the computation (remember tf.where statements still have a gradient!). Anyways, didn't test it out much, just left it as is since it's pretty harmless.

Implementation Strategy

One thing about these types of models is that they're complex! It's actually very easy to have a small bug that takes a long time to debug. After many false starts trying to get it working in one shot, I decided to actually do best practice and start small. The way I did this was incrementally test out new parts and then slowly put the pieces together. These are the iterations that I did (which correspond to the different notebooks you'll see on Github):

  • PixelConv2D layer: making sure that I got all the masking right. I made several toy networks ("unit tests") just to make sure the input was masked properly.

  • Next, I generated 2x2 RGB images from a known logistic distribution for each sub-pixel. Using this approach I could actually take a look at the distributional parameter outputs, plot the distribution for each pixel and then compare to actuals. The network I used for this was essentially just an output layer because I was testing how the loss function behaved. I spent a lot of time here trying to understand what was going on in the loss function. It was very helpful.

  • As a continuation, I generated the same RGB images but with a mixture of logistics. This was a natural extension of the previous iteration and allowed me to test out the logic I had for mixtures.

  • As a further continuation, I put together the actual network I wanted to use (with fewer Resnet layers). I had a couple of bugs here too.

  • After I was more or less sure that things were working on the toy example, I moved to the actual CIFAR10 images. Here I started out with a single image and tiny slices of it (e.g. 2x2, 4x4, 8x8, etc.) working my way up to the entire image. I wanted to see if I could overfit on a single example, which would give me some indication that I was on the right track.

  • Naturally, I extended this to 2 images, then multiple images, finally the entire dataset.

One thing that I found incredibly useful is to take notes for each set of experiments... duhhh! I know it's obvious but it's easy to be lazy when you're working on your own. You'll see at the bottom of the notebook I put some notes for each time I was working on it. You'll see some of the frustration and false starts that I went through.

Besides the obvious reasons why it's good to document progress, it was extra helpful because I only get to work on this stuff so sporadically that it was sometimes a month or two between sessions. Try remembering what tweaks you were doing to your loss function from two months ago! Anyways, it was really helpful because I could re-read my previous train of thought and then move on to my next experiment. Highly recommend it.

Another small thing that I did was that I prototyped everything in a notebook first and then as I was more confident that it worked, I moved it into a Python file. This was helpful because each notebook didn't have it's own (perhaps out of date) copy of a function. It also made the notebooks a bit nicer to read. I would recommend doing the first prototype in a notebook though, you want the agility of modifying things on the fly with the least friction.


Based on the implementation above, I was able to train a PixelCNN network. Some images generated are shown in Figure 6. Not very impressive... One thing that I think would make it better is to actually increase the convolution size. I'm just using a 5x5 window where with masking, only results in roughly a 3x5 window. I suspect going a bit bigger might help the image look better, similarly with the internal 3x3 convolutions. PixelCNN++ actually has varying internal kernel sizes to capture different resolutions.

One thing that I struggled with for a long time was that I was generating crappy images like the ones below. I tried really hard to get images that looked like the original ones. However, when I looked closer at the generated PixelCNN images, they also had a mishmash of things. They also weren't generating anything intelligible either (although theirs did look much better than mine). In any case, since we're not expecting photo-realistic images like GANs, I decided to move on.

PixelCNN Generated Images

Figure 6: PixelCNN Generated Images

What I was relatively happy with are my loss performance results shown in Table 1. Using the bits/pixel metric (see my discussion of this metric from my Importance Sampling post), I was able to achieve somewhat close results to the paper: 3.4 (mine) vs 3.14 ([1]).

It's not too surprising the results are different. I tried to use the same architecture as PixelCNN [1] except for how they model their output pixels. For that I used the PixelCNN++ [2] paper. PixelCNN has more flexibility in the output layer being a 256-way softmax so you would expect it do a bit better. PixelCNN++ does do better overall but I think it's because of their architectural changes with varying resolutions of conv layers compared with a Resnet-like architecture of PixelCNN.

Table 1: PixelCNN Loss on CIFAR10 (bits/pixel)


Training Loss

Validation Loss

My Implementation



PixelCNN [1]



PixelCNN++ [2]



I trained the PixelCNN to reduce the learning rate on a loss plateau 1 . Interestingly, I couldn't get the architecture to overfit. Somehow the validation loss is incredibly close to the training loss. This is in contrast to the PixelCNN++ paper, which states that overfitting is a major problem (which they address via dropout). This kind of suggests that the vanilla Resnet architecture probably isn't the most efficient for this task, which the authors of [1] change up in their extension paper in [3]. Still, I'm pretty happy with the results, which is my best CIFAR-10 loss to date.


I'm so glad that I finally got to implement and write about this post. It's been a long time coming and was particularly hard this past year where I've been so incredibly busy with work both at the company and university. In the two years since I wanted to write a post on this paper, there have been at least another half dozen papers that have come out that I want to write posts on. It's a bit of a never-ending cycle but that's what makes it fun! Hopefully you won't have to wait another two years for me to get to those papers!

Further Reading


I needed to start relatively small in the learning rate at 0.002 and then gradually reduce the learning rate or else the loss plateaus. One thing that I think happens is because we're modifying the scale parameter \(s\), you never want big jumps or else the loss changes dramatically. Thus, as you get in the right area with \(\mu\), you want to learn slower to "fine-tune" \(s\).

Hi, I'm Brian Keng. This is the place where I write about all things technical.

Twitter: @bjlkeng

Signup for Email Blog Posts