$$\newcommand{\dint}{\mathrm{d}} \newcommand{\vphi}{\boldsymbol{\phi}} \newcommand{\vpi}{\boldsymbol{\pi}} \newcommand{\vpsi}{\boldsymbol{\psi}} \newcommand{\vomg}{\boldsymbol{\omega}} \newcommand{\vsigma}{\boldsymbol{\sigma}} \newcommand{\vzeta}{\boldsymbol{\zeta}} \renewcommand{\vx}{\mathbf{x}} \renewcommand{\vy}{\mathbf{y}} \renewcommand{\vz}{\mathbf{z}} \renewcommand{\vh}{\mathbf{h}} \renewcommand{\b}{\mathbf} \renewcommand{\vec}{\mathrm{vec}} \newcommand{\vecemph}{\mathrm{vec}} \newcommand{\mvn}{\mathcal{MN}} \newcommand{\G}{\mathcal{G}} \newcommand{\M}{\mathcal{M}} \newcommand{\N}{\mathcal{N}} \newcommand{\S}{\mathcal{S}} \newcommand{\diag}[1]{\mathrm{diag}(#1)} \newcommand{\diagemph}[1]{\mathrm{diag}(#1)} \newcommand{\tr}[1]{\text{tr}(#1)} \renewcommand{\C}{\mathbb{C}} \renewcommand{\R}{\mathbb{R}} \renewcommand{\E}{\mathbb{E}} \newcommand{\D}{\mathcal{D}} \newcommand{\inner}[1]{\langle #1 \rangle} \newcommand{\innerbig}[1]{\left \langle #1 \right \rangle} \newcommand{\abs}[1]{\lvert #1 \rvert} \newcommand{\norm}[1]{\lVert #1 \rVert} \newcommand{\two}{\mathrm{II}} \newcommand{\GL}{\mathrm{GL}} \newcommand{\Id}{\mathrm{Id}} \newcommand{\grad}[1]{\mathrm{grad} \, #1} \newcommand{\gradat}[2]{\mathrm{grad} \, #1 \, \vert_{#2}} \newcommand{\Hess}[1]{\mathrm{Hess} \, #1} \newcommand{\T}{\text{T}} \newcommand{\dim}[1]{\mathrm{dim} \, #1} \newcommand{\partder}[2]{\frac{\partial #1}{\partial #2}} \newcommand{\rank}[1]{\mathrm{rank} \, #1}$$

# Generative Adversarial Networks (GAN) in Pytorch

This week is a really interesting week in the Deep Learning library front. There are two new Deep Learning libraries being open sourced: Pytorch and Minpy.

Those two libraries are different from the existing libraries like TensorFlow and Theano in the sense of how we do the computation. In TensorFlow and Theano, we have to symbolically construct our computational graph first before running it. In a sense, it is like writing a whole program before running it. Hence, the degree of freedom that we have in those libraries are limited. For example, doing loop, one need to use tf.while_loop() function in TensorFlow or scan() in Theano. Those approaches are less intuitive compared to imperative programming.

Enter Pytorch. It is a Torch’s port for Python. The programming style of Pytorch is imperative, meaning that if we’ve already familiar using Numpy to code our alogrithm up, then jumping to Pytorch should be a breeze. One does not need to learn symbolic mathematical computation, like in TensorFlow and Theano.

With that being said, let’s try Pytorch by implementing Generative Adversarial Networks (GAN). As a reference point, here is the TensorFlow version.

Let’s start by importing stuffs:

Now let’s construct our Generative Network $G(z)$:

It is awfully similar to the TensorFlow version, what is the difference then? It is subtle without more hints, but basically those variables Wzh, bzh, Whx, bhx are real tensor/ndarray, just like in Numpy. That means, if we evaluate it with print(Wzh) the value is immediately shown. Also, the function G(z) is a real function, in the sense that if we input a tensor, we will immediately get the return value back. Try doing those things in TensorFlow or Theano.

Next is the Discriminator Network $D(X)$:

Attentive readers will notice that unlike in TensorFlow or Numpy implementation, adding bias to the equation is non-trivial in Pytorch. It is a workaround since Pytorch has not implemented Numpy-like broadcasting mechanism yet. If we do not use this workaround, the X @ W + b will fail because while X @ W is mb_size x h dimensional tensor, b is only 1 x b vector!

Now let’s define the optimization procedure:

While at this point, in TensorFlow we just need to run the graph with G_solver and D_solver as the entry points, in Pytorch we need to tell the program what to do with those instances. So, just like in Numpy, we run the “forward-loss-backward-update” loop:

So first, let’s define the $D(X)$’s “forward-loss-backward-update” step. First, the forward step:

Nothing fancy, it is just a Numpy-like operations. Next, the backward and update step:

That is it! Notice, when we were constructing all the Ws and bs, we wrapped them with Variable(..., requires_grad=True). That wrapping is basically telling Pytorch that we cares about the gradient of those variables, and consequently pytorch.autograd module will calculate their gradients automatically, starting from D_loss. We could inspect those gradients by inspecting grad instance of the variables, e.g. Wxh.grad.

Of course we could code up our own optimizer. But Pytorch has built-in optimizers ready in pytorch.optim module. What it does is to abstract the update process and at each iteration, we just need to call D_solver.step() to update our variables, now that grad instance in those variables has been computed by backward() function.

As we have two different optimizers, we need to clear up the computed gradient in our computational graph as we do not need it anymore. Also, it is necessary so that the gradients won’t mix up with the subsequent call of backward() as D_solver shares some subgraphs with G_solver.

We do similar things to implement the “forward-loss-backward-update” for $G(z)$:

And that is it, really.

But we might ask, why do all of those things matter? Why not to just use TensorFlow or Theano? The answer is when we want to inspect or debug inside the computation graph, thing could be hairy in symbolic computation. Think of it like this: we are given a compiled program and what we can do is to run it. How do we debug a specific suboperation inside that program? Granted in TensorFlow we could inspect any variable by returning it once the computation is done, but still, we could only inspect it at the end of the computation not before.

In contrast, in imperative computation, we could just use print() function basically anywhere and anytime we want and immediately it will display the value. Doing other “non-trivial” operations like loop and conditional are also become much more easier in Pytorch, just like the good old Python. Hence, one could argue that this way of programming is more “natural”.

The full code is available in my Github repo: https://github.com/wiseodd/generative-models.