2020/6/26 COMP9444 Project 1
COMP9444 Neural Networks and Deep Learning
Term 2, 2020
Project 1 – Japanese Characters and Intertwined Spirals
In this assignment, you will be implementing and training various neural network models for two different classification tasks, and analysing the results.
You are to submit two Python files kuzu.py and spiral.py, as well as a written report hw1.pdf (in pdf format). Provided Files
Copy the archive hw1.zip into your own filespace and unzip it. This should create a directory hw1 with the data file spirals.csv as well as four Python files kuzu.py, spiral.py, kuzu_main.py and spiral_main.py.
Your task is to complete the skeleton files kuzu.py, spiral.py and submit them, along with your report.
Part 1: Japanese Character Recognition
For Part 1 of the assignment you will be implementing networks to recognize handwritten Hiragana symbols. The dataset to be used is Kuzushiji-MNIST or KMNIST for short. The paper describing the dataset is available here. It is worth reading, but in short: significant changes occurred to the language when Japan reformed their education system in 1868, and the majority of Japanese today cannot read texts published over 150 years ago. This paper presents a dataset of handwritten, labeled examples of this old- style script (Kuzushiji). Along with this dataset, however, they also provide a much simpler one, containing 10 Hiragana characters with 7000 samples per class. This is the dataset we will be using.
Text from 1772 (left) compared to 1900 showing the standardization of written Japanese.
1. [1 mark] Implement a model NetLin which computes a linear function of the pixels in the image, followed by log softmax. Run the code by typing:
python3 kuzu_main.py –net lin
Copy the final accuracy and confusion matrix into your report. Note that the rows of the confusion matrix indicate the target character, while the columns indicate the one chosen by the network. (0=”o”, 1=”ki”, 2=”su”, 3=”tsu”, 4=”na”, 5=”ha”, 6=”ma”, 7=”ya”, 8=”re”, 9=”wo”). More examples of each character can be found here.
2. [2 marks] Implement a fully connected 2-layer network NetFull, using tanh at the hidden nodes and log softmax at the output node. Run the code by typing:
python3 kuzu_main.py –net full
Try different values (multiples of 10) for the number of hidden nodes and try to determine a value that achieves high accuracy on the test set. Copy the final accuracy and confusion matrix into your report.
3. [2 marks] Implement a convolutional network called NetConv, with two convolutional layers plus one fully connected layer, all using relu activation function. You are free to choose for yourself the number and size of the filters, metaparameter values, and whether to use max pooling or a fully convolutional architecture. Run the code by typing:
python3 kuzu_main.py –net conv
Your network should consistently achieve at least 93% accuracy on the test set after 10 training epochs. Copy the final accuracy and confusion matrix into your report.
https://www.cse.unsw.edu.au/~cs9444/20T2/hw1/index.html 1/3
2020/6/26 COMP9444 Project 1
4. [7 marks] Discuss what you have learned from this exercise, including the following points: a. the relative accuracy of the three models,
b. the confusion matrix for each model: which characters are most likely to be mistaken for which other characters, and why?
c. you may wish to experiment with other architectures and/or metaparameters for this dataset, and report on your results; the aim of this exercise is not only to achieve high accuracy but also to understand the effect of different choices on the final accuracy.
Part 2: Twin Spirals Task
For Part 2 you will be training on the famous Two Spirals Problem (Lang and Witbrock, 1988). The supplied code spiral_main.py loads the training data from spirals.csv, applies the specified model and produces a graph of the resulting function, along with the data. For this task there is no test set as such, but we instead judge the generalization by plotting the function computed by the network and making a visual assessment.
1. [2 marks] Provide code for a Pytorch Module called PolarNet which operates as follows: First, the input (x,y) is converted to polar co-ordinates (r,a) with r=sqrt(x*x + y*y), a=atan2(y,x). Next, (r,a) is fed into a fully connected neural network with one hidden layer using tanh activation, followed by a single output using sigmoid activation. The conversion to polar coordinates should be included in your forward() method, so that the Module performs the entire task of conversion followed by network layers.
2. [1 mark] Run the code by typing
python3 spiral.py –net polar –hid 10
Try to find the minimum number of hidden nodes required so that this PolarNet learns to correctly classify all of the training data within 20000 epochs, on almost all runs. The graph_output() method will generate a picture of the function computed by your PolarNet called polar_out.png, which you should include in your report.
3. [1 mark] Provide code for a Pytorch Module called RawNet which operates on the raw input (x,y) without converting to polar coordinates. Your network should consist of two fully connected hidden layers with tanh activation, plus an output layer. You should not use Sequential but should instead build the network from individual components as shown in the program xor.py from Exercises 5 (repeated in slide 4 of lecture slides 3b on PyTorch). The number of neurons in both hidden layers should be determined by the parameter num_hid.
4. [1 mark] Run the code by typing
python3 spiral.py –net raw
Keeping the number of hidden nodes in each layer fixed at 10, try to find a value for the size of the initial weights (–init) such that this RawNet learns to correctly classify all of the training data within 20000 epochs, on almost all runs. Include in your report the number of hidden nodes, and the values of any other metaparameters. The graph_output() method will generate a picture of the function computed by your RawNet called raw_out.png, which you should include in your report.
5. [1 mark] Provide code for a Pytorch Module called ShortNet which again operates on the raw input (x,y) without converting to polar coordinates. This network should again consist of two hidden layers plus an output layer, but this time should include short-cut connections between every pair of layers (input, hid1, hid2 and output) as depicted on slide 10 of lecture slides 3a on Hidden Unit Dynamics. The number of neurons in both hidden layers should be determined by the parameter num_hid.
6. [1 mark] Run the code by typing
python3 spiral.py –net short
https://www.cse.unsw.edu.au/~cs9444/20T2/hw1/index.html 2/3
2020/6/26 COMP9444 Project 1
You should experiment to find a good value for the initial weight size, and try to find the mininum number of hidden nodes per layer so that this ShortNet learns to correctly classify all of the training data within 20000 epochs, on almost all runs. Include in your report the number of hidden nodes per layer, as well as the initial weight size and any other metaparameters. The graph_output() method will generate a picture of the function computed by your ShortNet called short_out.png, which you should include in your report.
7. [2 marks] Using graph_output() as a guide, write a method called graph_hidden(net, layer, node) which plots the activation (after applying the tanh function) of the hidden node with the specified number (node) in the specified layer (1 or 2). (Note: if net is of type PolarNet, graph_output() only needs to behave correctly when layer is 1).
Hint: you might need to modify forward() so that the hidden unit activations are retained, i.e. replace hid1 = torch.tanh(…) with self.hid1 = torch.tanh(…)
Use this code to generate plots of all the hidden nodes in PolarNet, and all the hidden nodes in both layers of RawNet and ShortNet, and include them in your report.
8. [9 marks] Discuss what you have learned from this exercise, including the following points:
a. the qualitative difference between the functions computed by the hidden layer nodes of the three models, and a brief
description of how the network uses these functions to achieve the classification
b. the effect of different values for initial weight size on the speed and success of learning, for both RawNet and ShortNet
c. the relative “naturalness” of the output function computed by the three networks, and the importance of representation for deep learning tasks in general
d. you may like to also experiment with other changes and comment on the result – for example, changing batch size from 97 to 194, using SGD instead of Adam, changing tanh to relu, adding a third hidden layer, etc.
Submission
You should submit by typing
give cs9444 hw1 kuzu.py spiral.py hw1.pdf
You can submit as many times as you like – later submissions will overwrite earlier ones. You can check that your submission has been received by using the following command:
9444 classrun -check
Additional information may be found in the FAQ and will be considered as part of the specification for the project. You should check this page regularly.
Plagiarism Policy
Group submissions will not be allowed for this assignment. Your program must be entirely your own work. Plagiarism detection software will be used to compare all submissions pairwise and serious penalties will be applied, particularly in the case of repeat offences.
DO NOT COPY FROM OTHERS; DO NOT ALLOW ANYONE TO SEE YOUR CODE
Please refer to the UNSW Policy on Academic Integrity and Plagiarism if you require further clarification on this matter. Good luck!
https://www.cse.unsw.edu.au/~cs9444/20T2/hw1/index.html 3/3