Skip to content

Latest commit

 

History

History
96 lines (60 loc) · 3.71 KB

flow.rst

File metadata and controls

96 lines (60 loc) · 3.71 KB

Work Flow

xmos-ai-tools is available on https://pypi.org/project/xmos-ai-tools/. It includes:

  • the MLIR-based XCore optimizer (xformer) to optimize Tensorflow Lite models for XCore
  • the XCore tflm interpreter to run the transformed models on host
  • the XCore tflm runtime to run transformed models on device

Installation steps

Perform the following steps once

  • Install xmos-ai-tools:

    # Create a virtual environment with
python3 -m venv <name_of_virtualenv>

# Activate the virtual environment
# On Windows, run:
<name_of_virtualenv>\Scripts\activate.bat
# On Linux and MacOS, run:
source <name_of_virtualenv>/bin/activate

# Install xmos-ai-tools from PyPI
pip3 install xmos-ai-tools --upgrade

    Use pip3 install xmos-ai-tools --pre --upgrade instead if you want to install the latest development version.

    Installing xmos-ai-tools will make the xcore-opt binary available in your shell to use directly.

  • Obtain the tool-chain from http://www.xmos.ai/tools and install it according to the platform instructions.

  • Setup XMOS_AITOOLSLIB_PATH environment variable. This is used to identify the installed location of xmos-ai-tools library and headers.

    On Windows, run the following command:

    FOR /F "delims=" %i IN ('python -c "import xmos_ai_tools.runtime as rt; import os; print(os.path.dirname(rt.__file__))"') DO set XMOS_AITOOLSLIB_PATH=%i

    On MacOS and Linux, run the following command:

    export XMOS_AITOOLSLIB_PATH=$(python -c "import xmos_ai_tools.runtime as rt; import os; print(os.path.dirname(rt.__file__))")

    Optionally, you may add the relevant export command in venv/bin/activate (for pip) or a script in ($CONDA_PREFIX)/etc/conda/activate.d/ (for conda), to automatically set the environment variable upon activating your virtual environment.

Example applications

These are 4 example models; in order of complexity

  • app_no_flash - a single model, no flash memory used. This is the fastest but most pressure on internal memory.
  • app_flash_single_model - a single model, with learned parameters in flash memory. This removes a lot of pressure on internal memory.
  • app_flash_two_models - two models, with learned parameters in flash memory.
  • app_flash_two_models_one_arena - two models, with learned parameters in flash memory. The models share a single tensor arena (scratch memory).

For more examples, see the examples folder.

More info regarding the generated C++ model files

The model code is compiled to C++ source and header. The generated header file contains the simple API to interact with the model. Some of the commonly used functions are:

  • model_init(void *flash_data) This takes a single parameter, which is a channel end to the flash server.
  • model_input_ptr(int index) This returns a pointer to the data where the input tensor should be stored; index should be set to zero unless there are multiple inputs.
  • model_invoke() This runs an inference
  • model_output_ptr(int index) This returns a pointer to the data where the output tensor would be stored.

Integration with sensors

In order to integrate with sensors, create separate threads (maybe on a separate tile) that input data from a sensor, camera, microphone, etc, and a thread that performs the signal pre-processing on it, prior to posting the data to the neural network.