Advanced topics
xmos-ai-tools is available on PyPI.
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
Perform the following steps once:
# 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 beta version.
Some older pre-release wheels may expect the opcode2name helper in the old tflite package layout.
If you use the host interpreter with one of those wheels and see an opcode2name import error,
either upgrade to a newer xmos-ai-tools build containing the compatibility fix,
or constrain tflite with pip3 install "tflite>=2.4.0,<=2.10.0".
from xmos_ai_tools import xformer as xf
# Optimizes the source model for xcore
# The main method in xformer is convert, which requires a path to an input model,
# an output path, and a list of configuration parameters.
# The list of parameters should be a dictionary of options and their values.
#
# Generates -
# * An optimized model which can be run on the host interpreter
# * C++ source and header which can be compiled for xcore target
# * Optionally generates flash image for model weights
xf.convert("source model path", "converted model path", params=None)
# Returns the tensor arena size required for the optimized model
# Only valid after conversion is done
xf.tensor_arena_size()
# Prints xformer output
# Useful for inspecting optimization warnings, if any
# Only valid after conversion is done
xf.print_optimization_report()
# To see all available parameters
# To see hidden options, run `print_help(show_hidden=True)`
xf.print_help()
For example:
from xmos_ai_tools import xformer as xf
xf.convert("example_int8_model.tflite", "xcore_optimised_int8_model.tflite", [
("xcore-thread-count", "5"),
])
To create a parameters file and a tflite model suitable for loading to flash, use the "xcore-weights-file" option.
xf.convert("example_int8_model.tflite", "xcore_optimised_int8_flash_model.tflite", [
("xcore-weights-file ", "./xcore_params.params"),
])
Some of the commonly used configuration options are described here
from xmos_ai_tools.xinterpreters import TFLMHostInterpreter
input_data = ... # define your input data
ie = TFLMHostInterpreter()
ie.set_model(model_path='path_to_xcore_model', params_path='path_to_xcore_params')
ie.set_tensor(ie.get_input_details()[0]['index'], value=input_data)
ie.invoke()
xformer_outputs = []
num_of_outputs = len(ie.get_output_details())
for i in range(num_of_outputs):
xformer_outputs.append(ie.get_tensor(ie.get_output_details()[i]['index'])).
# Note: use ie.close() or "with TFLMHostInterpreter() as ie:" to free resources
ie.close()