NTU H.Y. Lee Machine Learning Homework

Prepare

virtual environment and dependencies

method 1: virtualenv

create virtual environment

$ virtualenv ./ENV

enter virtual environment

$ source ./ENV/bin/activate

if you want to exit virual environment,

$ deactivate

install dependencies under virtual environment

$ pip2.7 install -r requirements.txt

method 2: pyenv + pipenv

use pyenv to local python version to this project,

$ pyenv install 2.7.15
$ pyenv local

use pipenv to set up dependencies,

$ pipenv --python 2.7.15
$ pipenv install

enter virtual environment

$ pipenv shell

GPU support on tensorflow

please install NVIDIA graphics driver, CUDA 9.0 and cuDNN 7.0 first. (ref: https://medium.com/@zhanwenchen/install-cuda-and-cudnn-for-tensorflow-gpu-on-ubuntu-79306e4ac04e )

otherwise, please choose cpu version tensorflow in requirements.txt

HW00

Original: https://docs.google.com/presentation/d/1VCnqWX469V4Qi_dHJIXxWMhvtg3h6pyBuRYPao-GZ1k/edit#slide=id.g17e2c04840_1_14

Q1. File given/hw0_data.dat has 11 columns splitted by single space. Please choose specific column i from this file, sort the sequence from small to large, and output to result/ans1.txt. Given i = 1.

$ cd hw00
$ ./Q1.sh 1 given/hw0_data.dat

Q2. Input the picture given/Lena.png. Please let this picture upside down and then left/right reversed (rotate 180 degree). Output to result/ans2.png

$ cd hw00
$ ./Q2.sh given/Lena.png

HW01

Original: http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2016/Lecture/hw1.pdf

data set: This data is an observation from 豐原 station, recorded weather parameters each hour at one day.

• given/train.csv: Choose first 20 days in each month to be a training set.
• given/test.csv: Choose last 10 days in each month to be a testing set. From testing set, select data among continuous 10 hours as a batch. Use first 9 hours data in this batch as a feature and PM2.5 at 10 hour as answer. And we hide this answer.

Please train a linear model to predict the answer in given/test.csv (format of output reference given/sampleSubmission.csv).

Ans:

Here presents two methods to implement linear regresssion.

1. using pseudo inverse,

$ cd hw01
$ python main.py --method pseudo_inverse --output result/pseudo_inverse.csv

2. using gradient descent,

$ cd hw01
$ python main.py --method gradient_descent --output result/gradient_descent.csv

HW02

Orignal: http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2016/Lecture/hw2.pdf

Q1: Implement logistic regression to detect spams: We have some labeled emails. In given/spam_train.csv, first severval columns present features of words and the last column presents spam/not spam labels. Those features are explained in given/spambase.names. Please implement logistic regression to predict whether letters are spams or not in given/spam_test.csv?

train:

$ cd hw02
$ python logistic_regression.py --type train --model result/model1.p

test:

$ python logistic_regression.py --type test --model result/model1.p --output result/result1.csv

Q2: Implement DNN to detect spams

train:

$ cd hw02
$ python dnn.py --type train --model result/model2.h5

test:

$ python dnn.py --type test --model result/model2.h5 --output result/result2.csv

HW03

Original: http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2016/Lecture/ML%20HW3.pdf

The CIFAR-10 dataset (https://www.cs.toronto.edu/~kriz/cifar.html) consists of 60000 32x32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images. The 10 classes include airplane, automobile, bird, cat, deer, dog, frog, horse, ship and truck, labeled 0-9 in order.

In this problem, we picked 500 images in each class from training set as labeled data, and hidden the other 45000 images' label from training set as unlabeled data.

Please use below code to prepare data as above instruction.

$ cd hw03
$ python prepare_data.py

• given/all_label.p contains ten classes (0-9), each class has 500 images.
• given/all_unlabel.p contains 45000 images
• given/test.p contains 10000 images
• given/test_ans.txt are labels of given/test.p

Q1. Supervised Learning: Use given/all_label.p data and CNN to predict given/test.p

$ cd hw03
$ python supervised_cnn.py --type train --model_config ycnet3 --model_name 002

The best result on training as following,

At epoch 22,
training loss: 0.9808
training accuracy: 75.87 %
validation loss: 1.0050
validation accuracy: 68.20 %

the log is at ./hw03/result/log/LOG_spv-cnn_ycnet3_002.logg

the model is at ./hw03/result/model/MODEL_spv-cnn_ycnet3_002.hdf5

The best validation accuracy can reach 68.20%.

Chart of training and validation loss:

Evaluation Result:

$ python supervised_cnn.py --type eval --model_config ycnet3 --model_name 002
Test: loss=1.1057958264350891, acc=64.47 %

Q2. Semi-supervised Learning Method 1: Self-training method. Try to use trained supervised model to label unlabeled data above specific reliablity threshold. Add those trusted data into labeled data and then use the augmented data to update CNN model.

$ cd hw03
$ python self_train_cnn.py --type train --model_config ycnet3 --model_name 002

Go from above Q1 trained supervised CNN model and use unlabeled data to update it.

In my observation, the key point is that reliablity threshold must be high enough. Otherwise the CNN model would get worst because adding incorrect labeled data causes more noise.

The best result on training as following

At Round 2 and epoch  7,
training loss: 0.4838
training accuracy: 89.31 %
validation loss: 0.6703
validation accuracy: 79.60 %

the log is at ./hw03/result/log/LOG_st-cnn_ycnet3_002.logg

the model is at ./hw03/result/model/MODEL_st-cnn_ycnet3_002.hdf5

Hence, the best validation accuracy can reach 79.60 %. Self-training method is work.

Evaluation Result:

$ python self_train_cnn.py --type eval --model_config ycnet3 --model_name 002
Test: loss=0.958553598595, acc=69.39 %

Q3. Semi-supervised Learning Method 2: Use all data (labeled data + unlabeled data) to pre-train autoencoder and extract some features of data. And use encoder in this autoencoder to do supervised learning on labeled data.

$ cd hw03
$ python cnn_autoencoder.py --type train --model_config AutoencoderClassifier02 --model_name 002

The best result on training as following,

At epoch 45,
training loss: 1.1201
training accuracy: 66.74 %
validation loss: 0.9521
validation accuracy: 67.80 %

the log is at ./hw03/result/log/LOG_ae-cnn_AutoencoderClassifier02_002.logg

the model is at ./hw03/result/model/MODEL_ae-cnn_AutoencoderClassifier02_002.hdf5

The best validation accuracy can reach 67.680. It is at same level with supervised CNN. But this model is healthier than supervised CNN, beacause traning lass and validation loss are closer each other in this model. So the pretrain process helps this model healthier.

The line of validation loss is more smoothing than supervised CNN.

Evaluation Result:

$ python cnn_autoencoder.py --type eval --model_config AutoencoderClassifier02 --model_name 002
Test: loss=1.06174045963, acc=64.85 %

HW04

Original: http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2016/Lecture/ML%20HW4.pdf

這個部分還沒完成，架構反覆檢查是沒有問題，但是不知道為什麼沒有做出Data分離成群的效果。