Science

ML and MLOps

3D shape classifier with data versioning, experiment tracking and CI-gated deployment.

Done Sprint S8+S16 · 13 points

In short, no jargon for non-developersRead the summaryClose

What it is: An artificial intelligence that recognises the shape of a 3D object (cube, sphere, cylinder…), trained on thousands of examples.
What it is for: The principle of automatic recognition, applied to 3D shapes: sorting parts, recognising scanned objects.
What it shows: The whole cycle of an artificial intelligence project: data, training, measured accuracy (over 95 %) and controlled release.

Extends the 3D CNN from my PhD

Source code

How the pieces fit together

DVC pipeline (params.yaml, dvc.lock) → Data (6 shapes, 512 points): generates
Data (6 shapes, 512 points) → Models (scikit-learn, PyTorch): train
Models (scikit-learn, PyTorch) → MLflow (local SQLite store)
Models (scikit-learn, PyTorch) → ONNX (pointnet.onnx, 300 KB): checked export
ONNX (pointnet.onnx, 300 KB) → Python API (POST /v1/mesh/classify)
DVC pipeline (params.yaml, dvc.lock) → Project page (results read at build): metrics, matrices

Results

Measured on 600 test point clouds (100 per class), never seen in training; recomputed by CI on every change to the project.

Model	Accuracy	Macro F1	Training (CPU)
Baseline: invariant descriptors + boosting (scikit-learn)	97.2 %	0.971	33 s
PointNet (PyTorch)	95.2 %	0.951	259 s
PointNet exported to ONNX (served by the API)	95.2 %	–	–

CI fails if the best accuracy, or the exported model's, drops below 90.0 %.

Confusion matrix, baseline
True class \ predicted	sphere	torus	box	cylinder	cone	capsule
sphere	100	0	0	0	0	0
torus	0	100	0	0	0	0
box	0	0	89	9	1	1
cylinder	0	0	4	94	0	2
cone	0	0	0	0	100	0
capsule	0	0	0	0	0	100

Confusion matrix, PointNet
True class \ predicted	sphere	torus	box	cylinder	cone	capsule
sphere	100	0	0	0	0	0
torus	0	100	0	0	0	0
box	0	0	88	11	0	1
cylinder	0	0	12	83	0	5
cone	0	0	0	0	100	0
capsule	0	0	0	0	0	100

Rows: the true class; columns: the predicted class; the diagonal counts the right answers. Nearly all errors come from boxes and cylinders being confused.

Tech stack

PyTorch
scikit-learn
DVC
MLflow
ONNX
GitHub Actions

Who does what

PyTorch: Trains the PointNet, a network that classifies a point cloud.
scikit-learn: The baseline: rotation-invariant descriptors and boosting, which beats the network here (97.2% against 95.2%).
DVC: Describes the pipeline (data, training, export, check) so that it replays identically; the data stays out of git.
MLflow: Keeps each training run's parameters, metrics and loss curve.
ONNX: The PointNet's export format, checked identical to PyTorch, which the API serves without PyTorch.
GitHub Actions: Replays the pipeline in CI and fails if accuracy drops below 90%.

Definition of Done

Code reviewed, diff by diff
Unit and integration tests green
Green CI on Linux, Windows and macOS when compiled
README, REVIEW.md and DECISIONS.md up to date
Texts in French and English
Demo deployed and tested on mobile