v4.5.6

v4.5.6 is the big machine-learning release. The TensorFlow tf C module was replaced by a new, engine-agnostic ml package backed by TensorFlow Lite Micro (TFLM), image→tensor conversion moved to image.to_ndarray(), and dual-core boards moved to Open-AMP. There are several breaking changes — read them below before upgrading ML scripts.

Highlights

  • New ml package — the old tf module was rewritten into an engine-agnostic ml package (ml.Model, ml.preprocessing, ml.utils, ml.apps); tf remains as an alias.

  • TFLM backend — the obsolete libtf was replaced by upstream-based TensorFlow Lite Micro: ~20% faster inference and smaller libraries.

  • Open-AMP dual-core — GIGA and Portenta H7 now use Open-AMP/RPMsg for M7/M4 inter-core communication (the old CM4 mechanism was removed).

  • Onboard Bluetooth — CYW43 Bluetooth firmware enabled on Nicla Vision, Portenta H7, and GIGA.

  • MicroPython 1.23.0, ulab 6.5.2 (with 4-D ndarray support).

  • Breaking: the ML API changed substantially and Haar cascades are disabled by default on most boards — see the breaking changes.

New features

  • ml package — ml.Model with predict() (roi, callback, multi-input list), per-tensor shape/dtype/scale/zero-point attributes, ml.preprocessing.Normalization, ml.utils (NMS), and ml.apps (a pure-Python MicroSpeech keyword spotter with listen() and non-blocking timeout=-1 streaming, plus draw_predictions()).

  • Built-in model system — models listed in models/index.txt are conditionally embedded per board (FOMO on all ML boards, audio models on boards with mics).

  • image.to_ndarray(dtype, buffer=...) — convert an image into a ulab ndarray (optionally in place).

  • Image() from raw arrays — new shape=, strides=, scale= keyword arguments build grayscale/RGB565 images from raw pixel lists.

  • Anti-aliased draw_circle().

  • ssl module frozen into firmware on GIGA, Nicla Vision, Portenta H7, Nano RP2040 Connect, OpenMV 4/4P/PRO/PT, RT1060, and Pico.

  • Audio — audio.init() gained a samples= keyword (PDM samples per channel); gain_db now applies on DFSDM microphones (e.g. Nicla Vision).

  • Debug protocol — a new GET_STATE command returns run/text/JPEG flags, frame geometry, and text in a single packet, cutting host round-trips.

  • Open-AMP vuart example for inter-core communication.

Other changes and improvements

  • MicroPython updated to 1.23.0; ulab to 6.5.2 with 4-D ndarray support.

  • Faster inference — the ML backend keeps persistent state/memory across invocations (~20% faster, supports LSTM-style models).

  • get_similarity() was reimplemented on the draw_image backend (broader format/operation support).

  • morph() and the mean() filter family moved to keyword-argument parsing; mask= now accepts a mutable image.

  • Memory layout reworked — DMA regions power-of-2 aligned, reorderable GC blocks, multiple heaps; less early heap fragmentation on low-RAM boards; RT1060 gains extra GC heap.

  • WiFi (CYW43) is now deinitialized on soft-reset; SPI-TV display flushes only the framebuffer region from cache for smoother updates.

Bug fixes

Camera and sensors:

  • Fixed accidental CPU cache invalidation in the framebuffer/sensor path on STM32 and i.MX RT, which could corrupt image data.

  • Fixed MLX90640/MLX90641 thermal I2C reads on i.MX RT boards (large transfers are now chunked).

Machine learning:

  • Fixed ml.Model load_to_fb handling, input bytearray sizing, ndarray ndim checks, and Normalization image-format/shape validation.

Image / system:

  • Added the missing image.BLACK_BACKGROUND draw-hint constant and an image/tensor buffer-length sanity check.

  • Fixed FIR and audio GC root pointers that could let the collector free in-use buffers.

  • The Servo Shield example now uses SoftI2C so it works on OpenMV RT boards.

Hardware and board support

  • Arduino GIGA — HM01B0 / HM0360 sensor support; Open-AMP dual-core.

  • Bluetooth — CYW43 BT firmware on Nicla Vision, Portenta H7, and GIGA.

  • Portenta H7 — Open-AMP M7/M4 RPMsg.

  • OpenMV RT1060 — additional GC heap blocks (more Python memory).

Breaking API changes

User-visible API breaks between v4.5.5 and v4.5.6. Scope: Python C-modules in modules/ and Python libraries in scripts/libraries/.

Each change is tagged with its impact:

  • major — most scripts that used it need edits.

  • minor — narrow API; only affects scripts that used it.

  • behavior — same API, different results; re-check tuned scripts.

  • tooling — affects dual-core / board mechanisms, not the Python API.

Changes are grouped by impact in that order. If you just want to port your code, jump to the migration checklist at the end. Each commit hash links to its diff on GitHub.

tf module replaced by the ml package (major)

The TensorFlow tf C module was rewritten into an engine-agnostic ml package (tf is kept as a backwards-compatible alias, but new code should use ml). The module-level functions tf.load, tf.load_builtin_model and the detect()/segment()/classification methods were removed — construct a ml.Model and call predict(). ml.Model(path) now returns just the model (no (labels, model) tuple); labels are a model.labels attribute. predict() now returns ulab ndarrays (not float tuples), supports multi-input models (pass a list), and NMS moved to ml.utils; input normalization moved to ml.preprocessing.Normalization. The old ml.py is now ml.apps.

Commits: c7228cbb4, 6c212409c, 3e37f46db, 9a186f4e2, 70b89f474, 3f8491cb0, 4506682c2, 8b38f3837

image.unpack() removed — use to_ndarray() (major)

The short-lived image.unpack() was removed; convert an image to a tensor with image.to_ndarray(dtype, buffer=...) instead, and apply scale/mean/ stdev normalization with ml.preprocessing.Normalization rather than the old built-in image-scaling path.

Commits: 9848eed12, de0d46fa6

Image() raw-array scale argument (minor)

When constructing an Image from a raw pixel array, the scale argument now takes a (min, max) range instead of (scale, add).

Commit: 7b79fb4c7

Haar cascades disabled by default on most boards (behavior)

To free flash space, Haar-cascade face detection (find_features() / image.HaarCascade()) is now disabled by default on Arduino GIGA, Nicla Vision, Portenta H7, OpenMV 3, OpenMV 4 / 4 Plus / PRO, and OpenMV Pure Thermal. Scripts using Haar cascades on those boards must rebuild the firmware with the feature enabled.

Commit: 6ce27c910

Script interruption and tensor scaling (behavior)

The USB debugger now interrupts a running script via MicroPython’s VM-abort feature instead of a forced PendSV jump (cleaner, but the interruption point differs). The SCALE_S128_127 input scaling was corrected to map 0–255 to −128–127 without an erroneous gain — models relying on the old (incorrect) scaling will produce different results.

Commits: e758a0f95, a4d97c572

CM4 coprocessor replaced by Open-AMP (tooling)

The deprecated CM4 coprocessor firmware mechanism on GIGA / Nicla Vision / Portenta H7 was removed and replaced by Open-AMP/RPMsg. Dual-core code must move to the Open-AMP model (a vuart example is provided).

Commits: 3cc57fea4, 93f2d4c41

Migration checklist

For a clean port to v4.5.6 the typical work is:

  1. Port ML code to ml: construct ml.Model(path), call predict(), read model.labels, expect ndarray outputs, and move normalization to ml.preprocessing.Normalization and NMS to ml.utils (the ml package change).

  2. Replace image.unpack() with image.to_ndarray() (the unpack removal).

  3. Update any Image() raw-array scale argument to a (min, max) range (the Image scale change).

  4. If you use Haar cascades on an affected board, rebuild with the feature enabled (the Haar cascade change).

  5. Re-validate models that depended on SCALE_S128_127 (the scaling change).

  6. Move dual-core code to Open-AMP (the CM4 change).