Machine Problem #10 Solution

Note: The assignment will be autograded.  It is important that  you do not use additional libraries, or change the provided functions’ input  and output.

 

 

Part 1:  Setup

 

• Connect to a EWS machine.

 

ssh (netid)@remlnx.ews.illinois.edu

 

 

• Load python  module, this will also load pip and virtualenv.  We use python  3.4 in this assignment.

 

module load python/3.4.3

 

 

• Reuse the virtual  environment from the previous mps.

 

source ~/cs446sp_2018/bin/activate

 

 

• Go to your svn directory.  Copy mp10 into your svn directory,  and change directory  to mp10.

 

cd ~/(netid)

svn cp https://subversion.ews.illinois.edu/svn/sp18-cs446/_shared/mp10 . cd mp10

 

 

• Install  the requirements through  pip.

 

pip install --upgrade pip

pip install -r requirements.txt

 

 

• Create  the  data  directory  and  prevent svn from checking in the  data  directory.   Remember not to submit  any training  data  to your svn directory.

 

mkdir MNIST_data

svn propset svn:ignore MNIST_data .

 

 

 

 

 

 

 

 

 

1

 

 

Part 2:  Exercise

 

In this exercise, we will learn how to generate images from the data  distribution using VAE. Review the lecture on Variational  AutoEncoder  for formulations.  You are asked to implement a VAE.

 

Part 2.1  VAE

 

You are asked to implement a VAE to learn the data distribution of the MNIST dataset. For the ease of visualization,  you can use 2-dimensional latent space.  Use fully connected layers for both  the  encoder and  decoder in the  VAE.  More details  are provided  in the  function docstring.

In  main vae.py , the training procedure, data reader and visualization are provided for you. Fill in the functions  in the file  vae.py .  To test your program,  run  main vae.py and see the generated  images stored in ‘latent space vae.png’.

 

The visualized results will look something similar to the figure below.

 

 

 

Figure 1: Generated  images by VAE

 

 

Hints:

 

• You need to modify/implement the following functions in  vae.py :

 

1)  sample z(self, z mean, z log var):  Sample z using the reparametrization trick given the outputs  of the encoder: z mean and z log var.

 

2)   encoder(self, x):  Build a two-layer  network.   Use two  fully-connected  layers with

100 nodes and then 50 nodes. Two output branches for the two ( nlatent)-dim vectors representing  the z mean and z log var.

 

3)  decoder(self, z): Build a two-layer network.  Use two fully-connected layers with 50 nodes and then  100 nodes. The output is a tensor containing  the decoded images.

 

4)  latent loss(self, z mean, z log var):  Calculate  the latent loss.

 

5)  reconstruction loss(self, f, x gt):  Calculate  the reconstruction loss.

 

 

6) loss(self, f, x gt, z mean, z var):  The total  loss of your VAE. You can call the above loss functions.

 

7) update  op(self,  loss, learning rate):   Perform  parameter update.    New an  Adam optimizer  here using tf.train.AdamOptimizer to minimize the  loss of your VAE and update  the parameters.

 

8) generate  samples(self,  z np):  Generate  random  samples from the  provided  latent samples z np (z np is a numpy array).

 

• You can use tensorflow.contrib.slim.fully  connected  to implement the fully-connected layers in the network.

 

• The  autograder   will test  your  implementation  in   vae.py only.    You  can  change

main vae.py .

 

• Remember  to make the input  dimension (ndims,  the argument of VariationalAutoen- coder’s constructor) and the latent dimension (nlatent, the argument of VariationalAu- toencoder’s constructor) flexible. The auto-grader  will test  it.

 

• Make sure the reparameterization trick works correctly.

 

• Make sure the shape of the network follows the instruction.

 

• Make sure the losses are calculated  correctly.

 

• The generated  images should look similar to Figure 1.

 

• For your information,  it took no more than  one minute to train  the VAE from scratch to get the  result  in Figure  1 in TA’s implementation.  We won’t time your program, but  you should keep this in mind.

 

 

 

Part 3:  Writing Tests

In  test.py we have provided  basic test-cases.   Make sure your code can pass the  simple

testing  cases provided in test.py.  Feel free to write more.  To test  the code, run

 

nose2

 

 

 

 

Part 4:  Submit

 

Submit  your python  code to the  svn repository.   You don’t  have to submit  the  visualized images.  Submitting  the code is equivalent to committing  the code.  This can be done with the following command:

 

svn commit -m "Some meaningful comment here."

 

 

Lastly, double check on your browser.  You can see your code at

 

https://subversion.ews.illinois.edu/svn/sp18-cs446/(netid)/mp10/

 

 

Make sure that  the file vae.py is up-to-date in your svn repository by checking on browsers after you commit your code. No  late submission will  be  accepted.