Circle Predictor

This use the libtorch to create a simple circle predictor just as an example.This is not a perfect network, but feel free to imptove it.. In this example, the network is fed an X value/point from the cartesian coordinate system. The results is the Y point and its opposite point in the circle.

How to build

• Clone the repo:

git clone https://github.com/mjshakir/circle_predictor.git

• Pull the submodules:

git pull --recurse-submodules

• A shortcut is to clone with submodules

git clone https://github.com/mjshakir/circle_predictor.git --recurse-submodules

Get libtorch

• Go to PyTorch to get libtorch
• Unziping

sudo apt-get install -y unzip 
unzip path/to/libtorch.zip

• Copy libtorch to the project folder

cp -r ~/path/to/libtorch ~/path/to/external_library/circle_predictor

• Folder system should look like

circle_predictor/
├── CMakeLists.txt
├── external_library
    └── libtorch
├── include
├── main
└── src

Build

• Create a build folder

~/path/to/circle_predictor/; mkdir build; cd build

• Get Ninja:

sudo apt-get install -y ninja-build

Build Ninja

• Build Ninja:

cmake -DCMAKE_BUILD_TYPE=release .. -G Ninja

• Build Ninja using CLang:

cmake -DCMAKE_CXX_COMPILER=/usr/bin/clang++ -DCMAKE_BUILD_TYPE=release .. -G Ninja

• Build the project:

ninja

TTB issue

• Build Ninja:

cmake -DUSE_TTB=OFF -DCMAKE_BUILD_TYPE=release .. -G Ninja

• Build Ninja using CLang:

cmake -DUSE_TTB=OFF -DCMAKE_CXX_COMPILER=/usr/bin/clang++ -DCMAKE_BUILD_TYPE=release .. -G Ninja

• Build the project:

ninja

Run the program

• Run: (Reminder: be in ~/path/to/circle_predictor/build/)

./bin/circle_predictor

Command line arguments

• -h or --help:
  • Display options and description:
• -s or --filename: default test_results
  • Save the results of the test data in csv file.
• -t or ---training_size: default 100
  • How many different points to train
  • Accepts an integer x > 0
• -g or --generated_size: default 10000
  • How many points generated
  • Accepts an integer x >= 100
• -b or --batch_size: default 20
  • Batch the generated data to train
  • Limitations:
    • Accept an unsigned integer x > 0
    • Must be less then the generated_size
    • Must be less the 1000 this a libtorch limitation
• -u or --use_epoch: default false
  • Train with an epoch iterations or validation precision
• -p or --precision: default 2.5E-1L
  • Determine when to stop the training. This uses a validation set.
• -e or --epoch: default 20
  • How many iteration to train.

Example:

• With precision and validation:

./bin/circle_predictor precision_results 10 1200 false 1E-2

• With Epoch iteration:

./bin/circle_predictor epoch_results 10 1200 true 10