Skip to content

Latest commit

 

History

History
85 lines (56 loc) · 3.15 KB

README.md

File metadata and controls

85 lines (56 loc) · 3.15 KB

Element in NCNN

This project is a fork of the RLBot v4 verison of Element.

It's aim is to port the bot to RLBot v5 and to port the PyTorch model to NCNN so the resulting exe is smaller and faster.

The ported model is already in src/.

Porting from PyTorch to NCNN

The following steps were taken to port the PyTorch model to NCNN. It is assumed your bot has already been ported to RLBot v5 and runs it's PyTorch model.

Prerequisites

Install the normal bot requirements via pip install -r requirements.txt.

Install PyTorch and pip install pnnx. These dependencies are needed to export to NCNN.

Step 1: Export the PyTorch model to TorchScript

This is done in the torch/ directory. agent.py and model.p were files taken almost directly from the original bot.

At the end of the file, you will find this snippet added:

if __name__ == "__main__":
    os.makedirs("../ncnn", exist_ok=True)
    
    # load the original model
    obs_size = 107
    agent = Agent(obs_size, action_categoricals=5, action_bernoullis=3)

    # generate random input tensor
    x = torch.rand(1, obs_size)

    # export model to torchscript
    mod = torch.jit.trace(agent.actor, x)
    mod.save("../ncnn/model.pt")

This creates the ncnn/ directory and saves the TorchScript model there.

Step 2: Exporting the TorchScript model to NCNN

This is done in the ncnn/ directory. Inside the directory, run pnnx model.pt inputshape=[1,107].

Step 3: Using the NCNN model

The resulting ncnn/model.ncnn.bin and ncnn/model.ncnn.param files can be used in the bot. Move them into the src/ directory.

Now, src/agent.py must be remade to use the NCNN model. Example usage can be seen in the generated ncnn/model_ncnn.py. This file includes tenors, but those tensors are immediately converted into numpy arrays.

Common runtime errors

terminate called after throwing an instance of 'std::runtime_error'
  what():  Convert ncnn.Mat to numpy.ndarray. Support only elemsize 1, 2, 4; but given 8

This is caused by the inputted numpy array having a dtype of float64. Convert it to float32 before inputting it into the NCNN model.

Configuring for the v5 botpack

  1. pip install pyinstaller

  2. pyinstaller --onefile src/bot.py --paths src --add-data src/model.ncnn.bin:. --add-data src/model.ncnn.param:. This will create a file called bot.spec - you may have to remove *.spec from .gitignore to commit it.

  3. Create bob.toml in the same directory as the spec file with the following content:

    [[config]]
project_name = "PythonExample"
bot_configs = ["src/bot.toml"]

[config.builder_config]
builder_type = "pyinstaller"
entry_file = "bot.spec"
    • project_name will be the name of your bot's folder in the botpack
    • bot_configs is a list of bot configs that will be included in the botpack
    • builder_type should always be pyinstaller
    • entry_file is the name of the spec file

  4. Commit both bot.spec and bob.toml to your bot's repository. Note that bob.toml CANNOT be renamed, but bot.spec can be anything as long as entry_file is also renamed to reflect the change.