Numeric modules (ulab + scipy + torch_c + matplotlib)

This directory bundles micropython-ulab (NumPy-like arrays + embedded SciPy subset) with C-ML (import cml / import torch, torch_c-compatible API) and a frozen matplotlib pyplot subset for a single USER_C_MODULES path plus training plots.

Git submodules (fetched automatically by the build script):

• micropython-ulab — https://github.com/OpenSourceJesus/micropython-ulab
• lib/C-ML — https://github.com/OpenSourceJesus/C-ML
• scipy — https://github.com/OpenSourceJesus/scipy (upstream reference; runtime SciPy is from ulab import scipy)
• matplotlib — https://github.com/OpenSourceJesus/matplotlib (upstream reference; runtime plotting is import matplotlib.pyplot)

Symlinks expose both build trees under one USER_C_MODULES root (required by the Make build):

ulab/ -> ../../micropython-ulab/code
cml/  -> ../cml-micropython/cml

Modules at runtime

Import	Role
from ulab import numpy	NumPy-like ndarray
from ulab import scipy	SciPy subset (integrate, linalg, optimize, signal, special)
import cml / import torch	torch_c tensors, autograd, nn
import matplotlib.pyplot	SVG line plots (see lib/matplotlib-micropython)

The top-level scipy/ directory in the repo root is upstream CPython SciPy (reference only). On MicroPython, use from ulab import scipy.

The top-level matplotlib/ directory is upstream CPython Matplotlib (reference only). On MicroPython, use import matplotlib.pyplot as plt from the frozen package in lib/matplotlib-micropython.

Build (unix)

Recommended — run from anywhere; the script resolves paths from the repo root:

/path/to/micropython/build-ml.sh

Manual equivalent (note: cd ../../ports/unix, not cd ports/unix, if you are still inside lib/numeric-modules):

REPO=/path/to/micropython
ln -sfn ../../micropython-ulab/code "$REPO/lib/numeric-modules/ulab"
ln -sfn ../cml-micropython/cml "$REPO/lib/numeric-modules/cml"
make -C "$REPO/ports/unix" VARIANT=ml submodules
make -C "$REPO/ports/unix" VARIANT=ml
"$REPO/ports/unix/build-ml/micropython" "$REPO/tests/ml/integration.py"
"$REPO/ports/unix/build-ml/micropython" "$REPO/tests/ml/matplotlib_integration.py"

Individual stacks:

make -C /path/to/micropython/ports/unix VARIANT=cml
make -C /path/to/micropython/ports/unix USER_C_MODULES=../../micropython-ulab

CMake ports (rp2, esp32, …)

make USER_C_MODULES=../../lib/numeric-modules/micropython.cmake

Requires MICROPY_PY_FFI=1 when using the cml FFI shims (as in the ml unix variant).

ulab ↔ torch_c interop

After cml.init(), cml.tensor() accepts ulab ndarrays directly (via the buffer protocol):

import cml
from ulab import numpy as np
from ulab import scipy

cml.init()
x = np.array([[1.0, 2.0], [3.0, 4.0]])
t = cml.tensor(x)          # ndarray → Tensor (float32)
y = np.array(t.tolist())   # Tensor → ndarray
z = scipy.special.erf(np.array([1.0, 2.0]))
cml.cleanup()

C-ML defaults to float32; ulab follows the port float width (often float64 on unix). Values are converted when importing ndarrays.