Skip to content

Latest commit

 

History

History
260 lines (200 loc) · 9.41 KB

README.md

File metadata and controls

260 lines (200 loc) · 9.41 KB

The tflite_micro Python Package

This directory contains the tflite_micro Python package. The following is mainly documentation for its developers.

The tflite_micro package contains a complete TFLM interpreter built as a CPython extension module. The build of simple Python packages may be driven by standard Python package builders such as build, setuptools, and flit; however, as TFLM is first and foremost a large C/C++ project, tflite_micro's build is instead driven by its C/C++ build system Bazel.

Building and installing locally

Building

The Bazel target //python/tflite_micro:whl.dist builds a tflite_micro Python .whl under the output directory bazel-bin/python/tflite_micro/whl_dist. For example:

% bazel build //python/tflite_micro:whl.dist
....
Target //python/tflite_micro:whl.dist up-to-date:
  bazel-bin/python/tflite_micro/whl_dist

% tree bazel-bin/python/tflite_micro/whl_dist
bazel-bin/python/tflite_micro/whl_dist
└── tflite_micro-0.dev20230920161638-py3-none-any.whl

Installing

Install the resulting .whl via pip. For example, in a Python virtual environment:

% python3 -m venv ~/tmp/venv
% source ~/tmp/venv/bin/activate
(venv) $ pip install bazel-bin/python/tflite_micro/whl_dist/tflite_micro-0.dev20230920161638-py3-none-any.whl
Processing ./bazel-bin/python/tflite_micro/whl_dist/tflite_micro-0.dev20230920161638-py3-none-any.whl
....
Installing collected packages: [....]

The package should now be importable and usable. For example:

(venv) $ python
Python 3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tflite_micro
>>> tflite_micro.postinstall_check.passed()
True
>>>  i = tflite_micro.runtime.Interpreter.from_file("foo.tflite")
>>> # etc.

Building and uploading to PyPI

The .whl generated above is unsuitable for distribution to the wider world via PyPI. The extension module is inevitably compiled against a particular Python implementation and platform C library. The resulting package is only binary-compatible with a system running the same Python implementation and a compatible (typically the same or newer) C library.

The solution is to distribute multiple .whls, one built for each Python implementation and platform combination. TFLM accomplishes this by running Bazel builds from within multiple, uniquely configured Docker containers. The images used are based on standards-conforming images published by the Python Package Authority (PyPA) for exactly such use.

Python .whls contain metadata used by installers such as pip to determine which distributions (.whls) are compatible with the target platform. See the PyPA specification for platform compatibility tags.

Building

In an environment with a working Docker installation, run the script python/tflite_micro/pypi_build.sh <python-tag> once for each tag. The script's online help (--help) lists the available tags. The script builds an appropriate Docker container and invokes a Bazel build and test within it. For example:

% python/tflite_micro/pypi_build.sh cp310
[+] Building 2.6s (7/7) FINISHED
=> writing image sha256:900704dad7fa27938dcc1c5057c0e760fb4ab0dff676415182455ae66546bbd4
bazel build //python/tflite_micro:whl.dist \
    --//python/tflite_micro:compatibility_tag=cp310_cp310_manylinux_2_28_x86_64
bazel test //python/tflite_micro:whl_test \
    --//python/tflite_micro:compatibility_tag=cp310_cp310_manylinux_2_28_x86_64
//python/tflite_micro:whl_test
Executed 1 out of 1 test: 1 test passes.
Output:
bazel-pypi-out/tflite_micro-0.dev20230920031310-cp310-cp310-manylinux_2_28_x86_64.whl

By default, .whls are generated under the output directory bazel-pypi-out/.

Uploading to PyPI

Upload the generated .whls to PyPI with the script python/tflite_micro/pypi_upload.sh. This script lightly wraps the standard upload tool twine. A PyPI authentication token must be assigned to TWINE_PASSWORD in the environment. For example:

% export TWINE_PASSWORD=pypi-AgENdGV[....]
% ./python/tflite_micro/pypi_upload.sh --test-pypi bazel-pypi-out/tflite_micro-*.whl
Uploading distributions to https://test.pypi.org/legacy/
Uploading tflite_micro-0.dev20230920031310-cp310-cp310-manylinux_2_28_x86_64.whl
Uploading tflite_micro-0.dev20230920031310-cp311-cp311-manylinux_2_28_x86_64.whl
View at:
https://test.pypi.org/project/tflite-micro/0.dev20230920031310/

See the script's online help (--help) for more.

Using tflite_micro from within the TFLM source tree

🚧 The remainder of this document is under construction and may contain some obsolete information. 🚧

The only package that needs to be included in the BUILD file is //python/tflite_micro:runtime. It contains all the correct dependencies to build the Python interpreter.

Examples

Depending on the workflow, the package import path may be slightly different.

A simple end-to-end example is the test python/tflite_micro/runtime_test.py:testCompareWithTFLite(). It shows how to compare inference results between TFLite and TFLM.

A basic usage of the TFLM Python interpreter looks like the following. The input to the Python interpreter should be a converted TFLite flatbuffer in either bytearray format or file format.

# For the Bazel workflow
from tflite_micro.python.tflite_micro import runtime


# If model is a bytearray
tflm_interpreter = runtime.Interpreter.from_bytes(model_data)
# If model is a file
tflm_interpreter = runtime.Interpreter.from_file(model_filepath)

# Run inference on TFLM using an ndarray `data_x`
tflm_interpreter.set_input(data_x, 0)
tflm_interpreter.invoke()
tflm_output = tflm_interpreter.get_output(0)

Input and output tensor details can also be queried using the Python API:

print(tflm_interpreter.get_input_details(0))
print(tflm_interpreter.get_output_details(0))

Technical Details

The Python interpreter uses pybind11 to expose an evolving set of C++ APIs. The Bazel build leverages the pybind11_bazel extension.

The most updated Python APIs can be found in python/tflite_micro/runtime.py.

Custom Ops

The Python interpreter works with models with custom ops but special steps need to be taken to make sure that it can retrieve the right implementation. This is currently compatible with the Bazel workflow only.

  1. Implement the custom op in C++

Assuming that the custom is already implemented according to the linked guide,

// custom_op.cc
TfLiteRegistration *Register_YOUR_CUSTOM_OP() {
    // Do custom op stuff
}

// custom_op.h
TfLiteRegistration *Register_YOUR_CUSTOM_OP();
  1. Implement a custom op Registerer

A Registerer of the following signature is required to wrap the custom op and add it to TFLM's ops resolver. For example,

#include "custom_op.h"
#include "tensorflow/lite/micro/all_ops_resolver.h"

namespace tflite {

extern "C" bool SomeCustomRegisterer(tflite::PythonOpsResolver* resolver) {
    TfLiteStatus status = resolver->AddCustom("CustomOp", tflite::Register_YOUR_CUSTOM_OP());
    if (status != kTfLiteOk) {
        return false;
    }
    return true;
}
  1. Include the implementation of custom op and registerer in the caller's build

For the Bazel workflow, it's recommended to create a package that includes the custom op's and the registerer's implementation, because it needs to be included in the target that calls the Python interpreter with custom ops.

  1. Pass the registerer into the Python interpreter during instantiation

For example,

interpreter = runtime.Interpreter.from_file(
    model_path=model_path,
    custom_op_registerers=['SomeCustomRegisterer'])

The interpreter will then perform a dynamic lookup for the symbol called SomeCustomRegisterer() and call it. This ensures that the custom op is properly included in TFLM's op resolver. This approach is very similar to TFLite's custom op support.

Print Allocations

The Python interpreter can also be used to print memory arena allocations. This is very helpful to figure out actual memory arena usage.

For example,

tflm_interpreter.print_allocations()

will print

[RecordingMicroAllocator] Arena allocation total 10016 bytes
[RecordingMicroAllocator] Arena allocation head 7744 bytes
[RecordingMicroAllocator] Arena allocation tail 2272 bytes
[RecordingMicroAllocator] 'TfLiteEvalTensor data' used 312 bytes with alignment overhead (requested 312 bytes for 13 allocations)
[RecordingMicroAllocator] 'Persistent TfLiteTensor data' used 224 bytes with alignment overhead (requested 224 bytes for 2 tensors)
[RecordingMicroAllocator] 'Persistent TfLiteTensor quantization data' used 64 bytes with alignment overhead (requested 64 bytes for 4 allocations)
[RecordingMicroAllocator] 'Persistent buffer data' used 640 bytes with alignment overhead (requested 608 bytes for 10 allocations)
[RecordingMicroAllocator] 'NodeAndRegistration struct' used 440 bytes with alignment overhead (requested 440 bytes for 5 NodeAndRegistration structs)

10016 bytes is the actual memory arena size.

During instantiation via the class methods runtime.Interpreter.from_file or runtime.Interpreter.from_bytes, if arena_size is not explicitly specified, the interpreter will default to a heuristic which is 10x the model size. This can be adjusted manually if desired.