haozech / NN

My implementation of a neuro network framework, which supports customized multi-layered perceptrons.

NN

Deep learning framework written with Python/Numpy for learning purpose.

Devops

This project has a docker image fzxu/nn built for developing environment. To start working on or testing the framework:

• (Prerequisite: docker should be installed) Run the docker_run.sh script in the root directory. Docker will pull the image from docker hub if it's your first time, which might take a while depending on your network.

  # docker run \
  #     -v `pwd`:/nn \
  #     -v `pwd`/bazel-cache:/root/.cache/bazel \
  #     -p 8888:8888 \
  #     --rm -it fzxu/nn /bin/bash
  ./docker_run.sh

  • This docker run command mount the repo dir and the bazel cache dir into the container, and also mapped 8888 port for jupyter notebook usesage.
• Once the container is started, you will be in /nn directory which is the mounting point of your repo dir. Now you can mess around with code, and use bazel:
  • To simply build the source code (which is usually not needed as there's no binary/executables in it)

  bazel build //src:nn_src
  
  # Output should look like:
  
  Extracting Bazel installation...
  Starting local Bazel server and connecting to it...
  INFO: Analyzed target //src:nn_src (6 packages loaded, 17 targets configured).
  INFO: Found 1 target...
  Target //src:nn_src up-to-date (nothing to build)
  INFO: Elapsed time: 13.164s, Critical Path: 0.11s
  INFO: 1 process: 1 internal.
  INFO: Build completed successfully, 1 total action

  • To run the tests (which will build the source code as dependency)

  bazel test //test:all
  
  # Output should look like:
  
  INFO: Analyzed 2 targets (19 packages loaded, 298 targets configured).
  INFO: Found 2 test targets...
  INFO: Elapsed time: 53.557s, Critical Path: 0.99s
  INFO: 10 processes: 6 internal, 4 processwrapper-sandbox.
  INFO: Build completed successfully, 10 total actions
  //test:layer_test                                                        PASSED in 0.8s
  //test:model_test                                                        PASSED in 0.8s
  
  Executed 2 out of 2 tests: 2 tests pass.
  There were tests whose specified size is too big. Use the --test_verbose_timeout_warnings command line option to see which

  • To test the framework by starting jupyter notebook
    • First build custom jupyter entry point with NN as dependency:

    bazel build //examples:jupyter
    
    # Output should look like:
    
    INFO: Analyzed target //examples:jupyter (1 packages loaded, 2 targets configured).
    INFO: Found 1 target...
    Target //examples:jupyter up-to-date:
      bazel-bin/examples/jupyter
    INFO: Elapsed time: 0.566s, Critical Path: 0.03s
    INFO: 4 processes: 4 internal.
    INFO: Build completed successfully, 4 total actions

    • Then run the built jupyter with some extra options:

    ./bazel-bin/examples/jupyter notebook --allow-root --ip 0.0.0.0
    
    # Output should look like:
    
    [I 11:45:45.372 NotebookApp] Writing notebook server cookie secret to /root/.local/share/jupyter/runtime/notebook_cookie_secret
    [I 11:45:46.049 NotebookApp] Serving notebooks from local directory: /nn
    [I 11:45:46.049 NotebookApp] Jupyter Notebook 6.4.0 is running at:
    [I 11:45:46.049 NotebookApp] http://68abe5ca1832:8888/?token=a3fc01601b64f5242b32cb32e254c69456da60aa81ea9b90
    [I 11:45:46.049 NotebookApp]  or http://127.0.0.1:8888/?token=a3fc01601b64f5242b32cb32e254c69456da60aa81ea9b90
    [I 11:45:46.049 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
    [W 11:45:46.063 NotebookApp] No web browser found: could not locate runnable browser.
    [C 11:45:46.064 NotebookApp] 
    
        To access the notebook, open this file in a browser:
            file:///root/.local/share/jupyter/runtime/nbserver-933-open.html
        Or copy and paste one of these URLs:
            http://68abe5ca1832:8888/?token=a3fc01601b64f5242b32cb32e254c69456da60aa81ea9b90
         or http://127.0.0.1:8888/?token=a3fc01601b64f5242b32cb32e254c69456da60aa81ea9b90    # <--- only this link works

TODO

• Transpose layer activation to be more intuitive
• Refactor to introduce operator concept (need design)
• Add more kinds of operators to implement/test more algorithms
  • Other activation functions
  • Other cost functions
  • Convolution layer
  • LSTM (?)
• Performance benchmark
• Better unit/integration testing (how?)
• Introduce Bazel
  • Build Bazel dependencies (e.g. python_rules) into docker image
• Investigate CI and devops
  • Add docker support
  • explore CI options
• Remove torch/torchvision dependency from mnist example