Changelog

All notable changes to this project are documented in this file.

[1.3.0] - 2026-10-29

Runtime switched from tflite-runtime to LiteRT (ai-edge-litert). tflite-runtime has not been released since 2023 and has no wheels for Python 3.12+, so the package could not be installed on current systems such as Raspberry Pi OS Trixie (Python 3.13). LiteRT, Google's renamed TensorFlow Lite runtime, provides wheels for Python 3.9–3.14 on 64-bit Raspberry Pi OS, Linux, Windows and macOS. On 32-bit Raspberry Pi OS (armv7l), where LiteRT has no wheels, tflite-runtime is still installed (Python ≤ 3.11). The package imports whichever runtime is available, falling back to tensorflow.lite, and exposes the one in use as teachable_machine_lite.RUNTIME.
NumPy 2 supported: the numpy==1.26.4 pin is replaced by numpy>=1.23.2. The pin was only needed because tflite-runtime is compiled against NumPy 1.x, not because of Teachable Machine models.
Labels: the "<index> " prefix Teachable Machine writes on each line of labels.txt (e.g. 0 Class 1) is now stripped, so result["label"] is "Class 1" instead of "0 Class 1". Labels are read as UTF-8, so non-English class names (e.g. Arabic) work on Windows too.
python_requires is now >=3.9 (it said 3.7 while the docs said 3.9). Requires Pillow>=10.1.
Removed the unused model_type constructor parameter.
Messages go through logging (logger teachable_machine_lite) only; the model-loaded message is no longer printed.

classify_image, show_prediction_on_image and classify_and_show accept a PIL image or a NumPy array (OpenCV / Picamera2 frame) as well as a file path, so camera frames no longer need to be written to disk first. NumPy arrays are BGR by default; pass color_order="RGB" for RGB arrays. Grayscale and 4-channel arrays are accepted.
num_threads constructor parameter, passed to the interpreter (e.g. num_threads=4 on Raspberry Pi 4/5).
top_k parameter on classify_image and classify_and_show; results now include a "predictions" list with the top_k most likely classes.
A warning is logged when the number of labels doesn't match the model's outputs.
Test suite with small generated .tflite models (float, uint8 and int8), plus a CI workflow on Linux x86_64/ARM64, Windows and macOS.
Docs: Raspberry Pi installation guide (virtual environment, Picamera2), Picamera2 example, command-line webcam example, upgrade guide, changelog and releasing pages. The docs site is now deployed automatically by GitHub Actions.

Floating point models gave wrong predictions: raw 0–255 pixels were copied into the float input tensor. Float inputs are now normalized to [-1, 1] as in Teachable Machine's training, so both the "Floating point" and "Quantized" exports work.
int8 quantized outputs were not dequantized (only uint8 was); uint8 dequantization could also overflow with a non-zero zero point.
show_prediction_on_image crashed on Windows and macOS with OSError: cannot open resource, because the DejaVuSans-Bold.ttf font can't be found by name there. It now falls back to Pillow's built-in font.
The label background is now actually translucent (it was drawn opaque black) and covers letters with descenders.
Results contain plain Python int/float instead of NumPy scalars, so they can be serialized with json.dumps.
Image files are closed after reading, so the same file (e.g. screenshot.jpg) can be overwritten every frame on Windows.

Version is defined only in src/__init__.py; setup.py reads it from there (MANIFEST.in includes it in the sdist).
The PyPI publish workflow runs when a GitHub Release is published and builds with python -m build (instead of the deprecated setup.py sdist bdist_wheel).

Pinned numpy==1.26.4 to work around NumPy 2 incompatibility; restructured the documentation.

Initial releases of the TeachableMachineLite class for image classification with TensorFlow Lite models.