Machine Problem #3 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. Also, you should include your best-trained ‘trained  weights.np’ under   codefromscratch when you make submission  to SVN.

 

 

Part 1:  Setup

 

•  Remote connect to an EWS machine.

 

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

 

 

•  Load python  module, this will also load pip and virtualenv

 

module load python/3.4.3

 

 

•  Reuse the virtual  environment from mp0.

 

source ~/cs446sp_2018/bin/activate

 

 

•  Copy mp3 into your svn directory,  and change directory  to mp3.

 

cd ~/(netid)

svn cp https://subversion.ews.illinois.edu/svn/sp18-cs446/_shared/mp3 . cd mp3

 

 

•  Install  the requirements  through  pip.

 

pip install -r requirements.txt

 

 

•  Create  data  directory  and Download the data  into the data  directory.

 

mkdir data

wget --user (netid) --ask-password \

http://courses.engr.illinois.edu/cs446/secure/assignment3_data.zip \

-O data/assignment3_data.zip

 

 

•  Unzip assignment3  data.zip

 

unzip data/assignment3_data.zip -d data/

 

 

 

1

 

 

•  Prevent svn from checking in the data  directory.

 

svn propset svn:ignore data .

 

Part 2:  Exercise

In this exercise, we will build a logistic model based binary  classifier that  can separate  two-

class audio datasets.  To be specific, the  datasets  consist of features  extracted  from audio waveforms of letters  of the alphabet.   The two classes represent the ee-set (b,c,d,e,g,p,t,v,z) and the eh-set (f,l,m,n,s,x),  respectively.  The binary  classifier will be based on the logistic regression model from our lecture as well as the sigmoid function from the written  homework.

The overall flow of the binary classifier is illustrated in Fig. 1 and we will implement each of the blocks in both “code-from-scratch”  and “TensorFlow”  parts.  In  main.py , the overall program structure is provided for you.

 

 

 

 

Training


Forward Operation (Compute Score/Loss)

 

 

Backward Operation (Gradient Update)

 

 

 

 

Data

(x,y)                   Preprocessing                 Model                          Score                          Loss

 

 

 

Output

 

Testing


Forward Operation (Classification)

 

 

 

Figure 1: High-level Flow

 

 

 

Part 2.1  Code-From-Scratch Implementation

 

•  Reading and Preprocessing Data.

 

In  io tools.py , we will fill in one function for reading in the dataset as well as pre- processing it.  The downloaded dataset has only “trainset” because the real “testset” belongs to autograder, which is used to evaluate  the actual  accuracy of your classifier. No matter “trainset” or “testset”, they share the same format.  Specifically, the “train- set” has an “indexing.txt” and a folder “samples”, where the former file lists all samples line-by-line in the form of “class label sample file.txt” and the latter  folder contains all sample files in ASCII text  format.  For more details, check the “README.txt”.  After

 

 

the  dataset is read,  it should be preprocessed  to match  the  specified form in return arguments  of the function.  More details are provided in the function docstring.

 

•  Logistic Model

 

In  logistic model.py , we will implement  a class for logistic models and use it for binary  classification.  The class will support  the following operations:

 

–  Forward operation. Forward  operation  is the function  which takes  input  fea- tures  and  generate  a  probability  score.   In  this  case,  for logistic models,  it  is



1+e−w   x
 
score = g(w| x) =       1    |


, where the sigmoid function is used from the written

homework.  Note that,  we redefine x = [1, x] and w = [b, w].

 

–  Backward operation. Backward operation  is for computing  the gradient of the loss function  with respect  to the model parameters.  This is computed  after  the forward operation to update  the model. Note that, in this part, loss function

should use  the Pi log  1 + exp(−y(i)wT φ(x(i)))  as  in  the lecture slides.

–  Classify operation.   Classify operation  performs  the  binary  classification  on input  dataset.  Each  of input  samples is computed  based on current update-to- date  model and then  is predicted  as one of the two classes {+1/ − 1}. Since all samples are taken  in as input,  the final output will be a vector of predicted  class labels.

 

–  Fit operation.  It  is the  training  function  which utilizes  the  above  methods as necessary to fit the  logistic model to the  input  dataset through  the  gradient descent approach.  As in the lecture, gradient descent is an optimization algorithm, where the model adjusts its parameters in direction of the negative gradient of the loss value.  The gradient descent should be repeated  iteratively  until convergence.

 

–  Save/Load operation.  After  training  with  hundreds  or thousands   of itera- tions,  the  final model should  be saved  into  a binary  file using  the  methods  – np.tofile(‘trained weights.np’) and np.fromfile(‘trained weights.np’), so that, in future,  the well-trained  model can be restored  immediately  and per- form classification tasks.   Note that, you   should also  include your best- trained ‘trained  weights.np’ under  codefromscratch when submitting your code to SVN.

 

•  Running Experiments. The  main.py shows the overall flow containing all previous methods.  We will not grade the  main.py file, feel free to modify it.

 

 

Relevant Files: everything in  the folder codefromscratch

 

 

 

Part 2.2  TensorFlow Implementation

In this part, we will implement the model in the previous section using TensorFlow (TF).  The



2
 
TF part focuses more on training method using the quadratic loss function Pi  yi  − g(w xi ) 

 

 

where  y  ∈  {0, 1},  instead  of  the  previous  loss  function  Pi log  1 + exp(−y(i) wT φ(x(i) )) 

where y ∈ {−1, 1}. Also, for simplicity, less operations  are required to implement in the TF

part.

 

•  Reading and Preprocessing Data.  In general, everything  is the same except the class labels.

 

•  Logistic Model The class in TF part  only needs to support  these operations:

 

–  Build Graph operation.  This  operation  builds  the  entire  training  and  test- ing TF  graph  for our logistic classifier, which includes model creation,  forward score/loss computation, optimizer definition, as well as training  accuracy compu- tation.   It is better  to put  the entire  TF  graph  in a single method,  according to our experience.

–  Fit operation. The training  operation  in TF utilizes the quadratic  loss as men- tioned earlier, which also includes iteratively  gradient descent update  on weights as well as computation of forward score and training  accuracy  during  each iter- ation.  After the model is trained,  this operation  also returns  the score vector of last iteration.

 

•  Running Experiments. The  main.py shows the overall flow containing all previous methods.  We will not grade the  main.py file, feel free to modify it.

 

 

Relevant Files: everything in  the folder codefromtf

 

 

 

Part 3:  Writing Tests

In  test scratch.py and  test tf.py , we have provided basic test-cases.  Feel free to write

more.  To test  the code, run

 

python test_*.py -v or

nose2

 

 

Part 4:  Submit

Submitting the code is equivalent to committing  the code. This can be done with the follow

command:

 

svn commit -m "Some meaningful comment here."

 

 

Lastly, double check on your browser that  you can see your code at

 

https://subversion.ews.illinois.edu/svn/sp18-cs446/(netid)/mp3/