CECS 551 Assignment 9

CECS 551
Assignment 9
Total: 30 Points
General Instruction
• Submit uncompressed file(s) in the Dropbox folder via BeachBoard (Not email).
1. Design LSTM network and implement it using Keras library to learn simple arithmetic
operations. The objective of the network is estimating result of addition or subtraction
of two numbers.
(a) (5 points) Implement a function to generate all pairs of query and answer. The
query includes two integer numbers (0˜99) and the an operation (+ or -), and the
answer includes correct results of the queries. You should have 100 × 100 × 2 =
20, 000 pairs of queries and answers. Please note that the lengths of queries and
answers are fixed as 5 and 4, respectively.
Data set
Query(X): ‘0+0 ’, ‘0-0 ’, ‘0+1 ’, ‘0-1 ’, ..., ‘99+99’, ‘99-99’
Answer(Y): ‘+0 ’, ‘+0 ’, ‘+1 ’, ‘-1 ’ ..., ‘+198’, ‘+0 ’
(b) (5 points) Implement a function to encode a string into one-hot-encoding scheme.
Please note that the dimensions of a queries and answers are fixed as 5 × 13 and
4 × 13, respectively.
alphabet = [‘0’,‘1’,‘2’,‘3’,‘4’,‘5’,‘6’,‘7’,‘8’,‘9’,‘+’,‘-’,‘ ’]
Encoding exmaple
‘4+27 ’
[[0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
[0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]]
‘+31 ’
[[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
[0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1]]
(c) (5 points) Implement a Encoder-Decoder LSTM network using the following code.
Explain how this code implements Encoder-Decoder scheme.
model = Sequential()
model.add(LSTM(?, input_shape=(5, 13), return_sequences=False))
model.add(RepeatVector(4))
model.add(LSTM(?, return_sequences=True))
model.add(Dense(13, activation=‘softmax’))
CECS 551 Assignment 9 - Page 2 of 2
(d) (5 points) Shuffle the data set, and use 70% samples as the training set, 15% as
the validation set, and 15% as the test set. Train the network and tune the hyperparameters, then report the best test accuracy and its setting.
(e) (5 points) Reverse the query and answer strings in the data set, then repeat training
with the settings of best test accuracy of (d).
‘4+27 ’ -> ‘ 72+4’
‘+31 ’ -> ‘ 13+’
(f) (5 points) With the setting of the best test(valid) accuracy, draw the chart of valid
accuracy vs. epoch for both non-reverse(baseline) and reverse data set as shown
in Figure 1. You can set validataion data=val data on model.fit() for this.
(Actual numbers might be different.)
Figure 1: An example of accuracy vs. epoch