Description
Enhance p1 with support for back propagation, training, and prediction.
This new program will take 8 command-line arguments:
neural_net_descr_file
like p1, this describes the NN; sample is in p2NET0;
use the simple sigmoid function at each neuron
input_filename
file containing the inputs to the NN; this file will be in csv format,
e.g. see p2SEEDS.csv
low_train_rng
index of first record in the input file to be used to train the NN
hi_train_rng
index just PAST last record in the input file to be used to train the NN;
note that this acts like python’s low:high indexes for lists etc
low_test_rng
index of first record in the input file to be used to test the NN
hi_test_rng
index just PAST last record in the input file to be used to test the NN;
note that this acts like python’s low:high indexes for lists etc
num_epochs
number of iterations to perform over the train input records before
beginning the testing phase
print_internals_flag
this flag will be 0 or 1; if 1, print the internal values of each neuron
in the NN before training and after training (before testing); print one
line per neuron, containing this info:
bias
weights
output (if one has been computed by training)
delta (if one has been computed by training)
We will back-propagate changes to weights but will leave biases un-altered.
Our discussions about backprop may include comments about values for momentum
and learning-rate. We will NOT be employing those values in our nets.
They are not represented in our command-line args. If you want to code a
value for learning-rate, just use 1.0 which should cause it to have no effect.
We will pre-process the data for our network before feeding it into the net:
Normalize the values in each attribute to work well with the sigmoid
function. We will discuss details of how to do that in class.
Labels can be represented as a one-hot version of their index into a
list of valid labels. Our network will have one neuron in the output
layer for each possible label in the list of labels. This will match
with the one-hot representation. Again, we will discuss details in class.
At the end of each training epoch which is a multiple of 10, print this kind of
information in this format:
0 70 of 210 0.3333
10 96 of 210 0.4571
20 139 of 210 0.6619
where in the last line, 20 is the epoch number, 139 is the number correct
of 210 which is the total number tested; and 0.6619 is 139/210 rounded to
4 decimal places.
At the end of the test phase, print the result in this form:
test result: 141 of 210 0.6714
Use turnin to submit a tar file containing all of your project files, including
a makefile that will build the executable program which MUST be named p2.
To build the project, I will cd to my directory containing your files and
simply type:
rm -rf p2
rm -f *.o
make
compute new weight:
w(n + 1) = w(n) + 0 + 1.0 * (d(n) * y)
