Connect to the server via ssh

Given the account information as below,

Id	Name	Username	Password	IP	Port-22	Port-6006	Port-8888
0	Alice	alice	password	166.111.69.245	22000	23000	24000

ssh to the server with

ssh -p 22000 alice@166.111.69.245

For Mac / Linux, you can use terminal. For Windows, you can use MobaXterm

Jupyter

Jupyter notebook is recommented to used in this homework.

If you are not familar with Jupyter notebook, see Jupyter notebook basics.

Run the jupyter notebook on the server,

git clone https://github.com/wujian752/MLP
cd MLP
jupyter notebook --ip 0.0.0.0

Jupyter would return you a url like

http://0.0.0.0:8888/?token=70f9d9457fec7045f5f991ad9d896060122b13d60cafbbc0

Replace the 0.0.0.0 with 166.111.69.245 and the port 8888 with 24000.

Then copy/paste this URL into your browser.

What are these files ?

Jupyter notebook files

Problem.ipynb: build and train your model in this notebook.

Demo.ipynb: a demo of this framework in which a MLP is built with builtin FCLayer, ReLULayer.

Python files

layers/
    - fc_layer.py: implement the forward and backward of fully connected layer in this file.
    - sigmoid_layer.py: implement the forward and backward of ReLU layer in this file.
    - relu_layer.py: implement the forward and backward of Sigmoid layer in this file.
    - __init__.py: mark layers/ as Python package directory.

criterion/
    - softmax_cross_entropy.py: the softmax cross entropy criterion.
    - __init__.py: mark criterion/ as Python package directory.

builtin/
    - Some pyc files that can be executed.

MNIST_data/
    - the directory in which MNIST dataset is stored.

network.py: the class network is defined here.

optimizer.py: the SGD is defined here and you need to implement the SGD with momentum in this file.

solver.py: the train function and test function are defined here.

You can train a MLP in the Demo.ipynb with the builtin layers.

And after you implement your FCLayer, SigmoidLayer and ReLULayer in layers, you can set use_builtin to False and run the program again to test your implementation.

Then, build and train the SigmoidMLP and ReLUMLP in Problem.ipynb.

Finally, implement SGDwithMomentum in optimizer.py and train ReLUMLP_momentum with SGDwithMomentum in Problem.ipynb.