Nanyan "Rosalie" Zhu and Chen "Raphael" Liu

This repository is a child repository of RnR-2018/Deep-learning-with-PyTorch-and-GCP. This serves a primary purpose of facilitating the course BMEN4460 instructed by Dr. Andrew Laine and Dr. Jia Guo at Columbia University. However, it can also be used as a general beginner-level tutorial to implementing deep learning algorithms with PyTorch on Google Cloud Platform.

This repository, Image Classification on MNIST Data, presents an application of simple convolutional neural networks (CNNs) on image classification.

For students in BMEN4460 (or who follow the instructions Step00 through Step02 in the parent repository), please create a Projects folder (if you have not done yet) within your GCP VM and download this repository into that folder.

The notebook BMEN4460-NB2-image_classification_on_MNIST_data.ipynb walks you through the process of building a convolutional neural network (CNN) to classify the handwritten digits.

The notebook with visualization BMEN4460-NB2-image_classification_on_MNIST_data_with_visualization.ipynb intends to visualize how the convolutional kernels/filters evolve over the training process. In our specific case, it seems that these kernels in the first convolutional layer changed minimally throughout the process. Well, don't feel disappointed yet, as there are a lot more you can explore beyond this point. First of all, it is possible that the kernels in deeper layers evolve more drastically over the epochs, which you can check whether it is true or not yourself. Besides, even if this is not the case, at least you get to know this visualization method which you might find useful some day in your own deep learning research.

On the GCP VM Terminal:

cd /home/[username]/
mkdir BMEN4460 # This is only necessary if you have not done this yet
mkdir BMEN4460/MNIST # This is only necessary if you have not done this yet
cd BMEN4460/MNIST
git clone https://github.com/RnR-2018/BMEN4460-NB2-image_classification_on_MNIST_data/

If it says "fatal: could not create work tree dir ...", you may as well try it again with super user permission

sudo git clone https://github.com/RnR-2018/BMEN4460-NB2-image_classification_on_MNIST_data/

You shall then see the following hierarchy of files and folders, hopefully, which matches the hierarchy of the current repository.

BMEN4460-NB2-image_classification_on_MNIST_data
    ├── BMEN4460-NB2-image_classification_on_MNIST_data.ipynb
    ├── BMEN4460-NB2-image_classification_on_MNIST_data_with_visualization.ipynb
    └── data
        └── MNIST_CNN.png

sudo su [username]
cd /opt/anaconda3/env/ # If you installed anaconda in another directory the permission problem is not likely an issue for you.
ls -ll # You should see that you don't have the write permission.
sudo chmod 777 -R ./* # Grants read, write, execute permission to all users.
ls -ll # Now you should have all three permissions.

conda activate BMEN4460
conda install -c anaconda tensorflow pandas seaborn

python -m ipykernel install --user --name BMEN4460 --display-name "Python3.7 BMEN4460"

cd /home/[username]/Projects
sudo chmod 777 -R ./*

This notebook is largely inspired by and adapted from this GitHub repository. The network architecture used in the tutorial comes from this GitHub repository. Please note that the layer dimensions are (highly likely) incorrect in the network architecture illustration given by the cited GitHub repository.