Migrate to Python 3.12/3.13, uv, and Keras 3 (.keras format)

[DocGarbanzo]

What this PR does

Migrates donkeycar from Python 3.11 + conda to Python 3.12/3.13 + uv,
upgrades TensorFlow to 2.19, and removes all Keras 2 / SavedModel legacy code
that broke under Keras 3.

Version bump: 5.3.x → 5.4.dev1 (breaking changes — see below).

---

Platform matrix

Platform	Python	TF	Notes
Mac	3.12	2.19 + Metal	tensorflow-metal==1.2.0 has no Python 3.13 wheel
PC / Linux	3.12	2.19	TF SavedModel export broken on Python 3.13
Raspberry Pi	3.13	— (TFLite only)	Must match Debian system Python to reach libcamera/picamera2

---

Install instructions

1 — Install uv (all platforms)

curl -LsSf https://astral.sh/uv/install.sh | sh

---

2a — Raspberry Pi (Python 3.13, Debian Trixie / Pi OS Bookworm+)

Pi uses the Debian system Python 3.13 directly. Camera support
(libcamera, picamera2) is only available as Debian packages installed
into that system Python — they are not on PyPI. The venv must be created
with --system-site-packages so it can see them.

# Install camera system packages first
sudo apt install python3-libcamera python3-picamera2

# Create venv from the system Python 3.13 with access to system packages
uv venv ~/env --python 3.13 --system-site-packages

# Activate on login
echo 'source ~/env/bin/activate' >> ~/.bashrc
source ~/env/bin/activate

# User install (PyPI release):
uv pip install donkeycar[pi]

# Developer install (git clone):
uv pip install -e ".[pi,dev]"

Why not Python 3.12 on Pi?
python3-libcamera and python3-picamera2 are Debian packages installed
under /usr/lib/python3/dist-packages. A custom Python 3.12 venv's
--system-site-packages only reads /usr/local/lib/python3.12/site-packages
— it cannot see the Debian packages. Python 3.13 is the Debian system
Python, so its venv resolves both locations correctly.

Why no full TensorFlow on Pi?
Pi uses ai-edge-litert (Google's official TFLite successor, replacing the
dead tflite-runtime package). Inference-only; no training on Pi.

---

2b — Mac (Python 3.12)

# Create a persistent named venv
uv venv ~/.venvs/donkeycar --python 3.12

# Activate on login
echo 'source ~/.venvs/donkeycar/bin/activate' >> ~/.zshrc
source ~/.venvs/donkeycar/bin/activate

# User install (PyPI release):
uv pip install donkeycar[macos]

# Developer install (git clone):
uv pip install -e ".[macos,dev]"

Why Python 3.12, not 3.13?
tensorflow-metal==1.2.0 (required for GPU acceleration on Apple Silicon)
has no Python 3.13 wheel. TF 2.19 is the last version compatible with
tensorflow-metal==1.2.0. Additionally, TF's SavedModel export
(model.export()) is broken on Python 3.13 due to inspect.getattr_static
behaviour changes.

---

2c — PC / Linux (Python 3.12)

# Create a persistent named venv
uv venv ~/.venvs/donkeycar --python 3.12

# Activate on login (bash)
echo 'source ~/.venvs/donkeycar/bin/activate' >> ~/.bashrc
source ~/.venvs/donkeycar/bin/activate

# User install (PyPI release):
uv pip install donkeycar[pc]

# Developer install (git clone):
uv pip install -e ".[pc,dev]"

---

Breaking changes

Change	Action required
Python 3.11 no longer supported	Recreate your venv with Python 3.12 (Mac/PC) or 3.13 (Pi)
conda removed	Uninstall old conda env; follow uv install steps above
.savedmodel models cannot be loaded	Retrain your pilots — Keras 3 cannot load SavedModel directories. New models save as .keras automatically
picamera2 removed from pip extras	Install via sudo apt install python3-picamera2 on Pi
tflite-runtime replaced by ai-edge-litert	Run uv pip install donkeycar[pi] — handled automatically
RPi.GPIO replaced by gpiozero	Run uv pip install donkeycar[pi] — handled automatically

---

Key code changes

• pyproject.toml: replaces setup.cfg + MANIFEST.in; all metadata in one file
• donkeycar/__init__.py: version 5.4.dev1; Python floor check updated to 3.12 (Mac/PC) / 3.13 (Pi)
• parts/interpreter.py: removed broken TensorRT.load() branch that called keras.models.load_model() on a .savedmodel dir (not supported in Keras 3); added tensor-API fallback in TfLite interpreter for models without serving_default signature
• pipeline/training.py: TRT export unconditionally uses model.export() (dead .savedmodel guard removed)
• pipeline/database.py: .savedmodel removed from name-stripping and prefix map
• management/ui/: all file filters and sync buttons updated from .savedmodel → .keras
• scripts/convert_to_tflite.py: accepts .keras input instead of .savedmodel
• templates/complete.py, simulator.py: .savedmodel removed from model-path detection
• parts/keras.py, pipeline/training.py, management/: tensorflow.python.keras.* → tensorflow.keras.* (removed in TF 2.16)
• .github/workflows/python-package.yml: replaced conda workflow; explicitly creates venv with matrix Python version before install

Test plan

• pytest donkeycar/tests/test_keras.py — 8 pass, 1 xfail (KerasLSTM, CudnnRNNV3 unsupported in TFLite on macOS)
• Train a model → verify saved as .keras
• Load .keras model in UI (pilot screen, train screen transfer selector)
• Sync screen shows buttons: h5, keras, tflite, trt
• On Pi (Python 3.13 + --system-site-packages): camera works, TFLite inference runs, donkey drive completes

🤖 Generated with Claude Code