/open3d-3dml-mockup

Primary LanguagePython

Open3D 3DML

This repo is a proposal for the directory structure.

The repo can be used together with the precompiled open3d pip package but will also be shipped with the open3d package. The file examples/inference_segmentation.py contains a working example showing how the repo can be used directly and after it has been integrated in the open3d namespace.

Directories

├─ examples              # place for example scripts and notebooks
├─ docs                  # rst files that can be integrated in the open3d docs
├─ ml3d                  # package root dir; this will become open3d/_ml3d in the package
     ├─ datasets         # generic dataset code; will beopen3d.datasets
     ├─ tf               # directory for tensorflow specific code. same structure as ml3d/torch
     ├─ torch            # directory for pytorch specific code
          ├─ dataloaders # framework specific dataset code, e.g. wrappers that can make use of the generic dataset code.
          │              # This will be open3d.ml.torch.datasets with open3d.ml.torch as ml3d
          ├─ models      # modules with model code and/or packages for more complicated models.
          ├─ pipelines   # pipelines with categories as subdirectories e.g. detection, segmentation

Usage example

Most users will either use tf or torch. The recommended import code is

import open3d as o3d
import open3d.core as o3c
import open3d.ml.torch as ml3d
# or
import open3d.ml.tf as ml3d

# using ml3d
net = ml3d.models.mynet(params)
dataset = ml3d.dataloaders.ScanNet(params)
conv = ml3d.layers.ContinuousConv(params)

When using this repo directly (e.g. for development) the imports are

import open3d as o3d
import open3d.core as o3c
import ml3d.torch as ml3d
# or
import ml3d.tf as ml3d

# using ml3d
net = ml3d.models.mynet(params)
dataset = ml3d.dataloaders.ScanNet(params)
from open3d.ml.tf.layers import ContinuousConv; conv = ContinuousConv(params) # we don't need layers at the level of models and pipelines

Note that in this case ml3d will not behave exactly like the packaged version because ml3d.layers and other functionality from the main repo will be missing. This should not pose a problem because

  1. we could import the missing functionality in ml3d/[tf,torch]/__init__.py
  2. we don't need to use layers and ops for using models and pipelines. (pipelines and models will use absolute imports internally for getting the layers and ops)