TSMC course materials of deep learning.

Course slides

Setup your python environment. We recommend to use python3.

sudo pip3 install -r requirements.txt

requirements.txt

Use BMM or GMM to fit MNIST dataset (handwritten digits), and then use the trained model to perform classification tasks.

Demo_BMM.ipynb demos how to use Bernoalli mixture model.
Demo_GMM.ipynb demos how to use Gaussian mixture model.

Please refer to mixture.py, bmm.py, gmm.py for detailed implementations.

kmeans.py perform kmeans clustering on MNIST dataset.

python kmeans.py --path=[Path to MNIST dataset directory. Default to "../MNIST".]
                 --k=[Number of cluster center. Default to 10.]
                 --output=[File path to save cluster centers. Default to "kmeans.dat".]
                 --verbose=[True | False. Default to False.]

Note: Use python3 to run the code.

Neural network for MNIST and embedding visualization

In your code:

tf.summary.FileWriter('your path to log dir', ...)

tensorboard command:

tensorboard --logdir <your path to log dir> --port <your port (defalut:6006)>

In this notebook, we show how to train a neural network to do multiple-class classification in MNIST dataset. Since the url in tesorflow example is broken, please download the dataset from here.

Please change the data directory path to your own.

parser.add_argument('--data_dir', type=str, 
                    default='/home/tommy8054/pythonPlayground/MNIST_data/', # here!
                    help='Directory for storing input data')

Put your log files to desired directory.

parser.add_argument('--log_dir', type=str, 
                    default='/tmp/tensorflow/mnist/logs/mnist_with_summaries', # here!
                    help='Summaries log directory')

Show training detail or embedding visiualization. True for training detail and False for embedding visiualization.

parser.add_argument('--save_log', type=bool, default=False, # here!
                    help='Whether save log file or not')

Sparse coding using neural network

In this section, we demonstrate that sparsity constrain leads to sparse feature while training.

Denoising Autoencoder

This is a simple example of denoising autoencoder on MNIST dataset. Pretrained weights are put in net_pretrained/dae