Skip to content

Latest commit

 

History

History
145 lines (96 loc) · 5.25 KB

CHANGELOG.md

File metadata and controls

145 lines (96 loc) · 5.25 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unrealeased

Changed

  • Migrated to Hydra 1.0 and OmegaConf 2.1 BREAKING any checkpoint created prior to that, in particular the model zoo

Added

  • Mixed precision training support for SparseConv3D models with torchsparse backend (Requires torchsparse >= 1.3.0)

1.3.0

Added

  • MS-SVConv: https://arxiv.org/abs/2103.14533 (thanks @humanpose1)
  • added new data generations techniques for the self-supervised learning (PeriodicSampling, IrregularSampling EllipsoidCrop) (thanks @humanpose1)
  • More ETH benchmark dataset (thanks @humanpose1)

Changed

  • Minkowski 0.5 support

Bug fixes

1.2.0

Added

  • Support for the IRALab benchmark (https://arxiv.org/abs/2003.12841), with data from the ETH, Canadian Planetary, Kaist and TUM datasets. (thanks @simone-fontana)
  • Added Kitti for semantic segmentation and registration (first outdoor dataset for semantic seg)
  • Possibility to load pretrained models by adding the path in the confs for finetuning.
  • Lottery transform to use randomly selected transforms for data augmentation
  • Batch size campling function to ensure that batches don't get too large
  • TorchSparse backend for sparse convolutions
  • Possibility to build sparse convolution networks with Minkowski Engine or TorchSparse
  • PVCNN model for semantic segmentation (thanks @CCInc)

Bug fix

  • Dataset configurations are saved in the checkpoints so that models can be created without requiring the actual dataset
  • Trainer was giving a warning for models that could not be re created when they actually could
  • BatchNorm1d fix (thanks @Wundersam)
  • Fix process hanging when processing scannet with multiprocessing (thanks @zetyquickly)
  • wandb does not log the weights when set in private mode (thanks @jamesjiro)
  • Fixed VoteNet loss definitions and data augmentation parameters (got up to 59.2% mAP25)

Changed

  • More general API for Minkowski with support for Bottleneck blocks and Squeeze and excite.
  • Docker images tags on dockerhub are now latest-gpu and latest-cpu for the latest CPU adn GPU images.

Removed

  • Removed VoteNet from the API because it was not up to date. You can still use the models defined there

1.1.1

Added

  • Teaser support for registration
  • Examples for using pretrained registration models
  • Pointnet2 forward examples for classification, segmentation
  • S3DIS automatic download and panoptic support and cylinder sampling

Changed

  • Moved to PyTorch 1.6 as officialy supported PyTorch version

Bug fix

  • Add context = ssl._create_unverified_context(), data = urllib.request.urlopen(url, context=context) within download_ulr, so ModelNet and ShapeNet can download.

1.1.0

Added

  • Support scannet test dataset and automatic generation of submission files using the eval.py script
  • Full res predictions on Scannet with voting
  • VoteNet model and losses
  • Tracker for object detection
  • Models can specify which attributes they need from the data in order to forward and train properly
  • Full res predictions on ShapeNet with voting
  • Trainer class to handle train / eval
  • Add testing for Trainer:
    • Segmentation: PointNet2 on cap ShapeNet
    • Segmentation: KPConv on scannetV2
    • Object Detection: VoteNet on scannetV2
  • Add VoteNet Paper / Backbones within API
  • Windows support
  • Weights are uploaded to wandb at the end of the run
  • Added PointGroup https://arxiv.org/pdf/2007.01294.pdf
  • Added PretrainedRegistry allowing model weight to be downloaded directly from wandb and DatasetMocking
  • Added script for s3dis cross-validation [scripts/cv_s3dis.py]. 6 different pretrained models will be downloaded, evaluated on full resolution and confusion matrice will be summed to get all metrics.
  • mAP tracker for Panoptic segmentation

Changed

  • evaluation output folder is now a subfolder of the checkpoint it uses
  • saves model checkpoints to wandb
  • GridSampling3D now creates a new attribute coords that stores the non quantized position when the transform is called in quantize mode
  • cuda parameter can be given in command line to select the GPU to use
  • Updated to pytorch geometric 1.6.0

Bugfix

  • LR secheduler resume is broken for update on batch number #328
  • ElasticDistortion transform is now fully functional

Removed

1.0.1

Changed

  • We now support the latest PyTorch
  • Migration to the latest PyTorch Geometric and dependencies

Bugfixes

  • #273 (support python 3.7)

0.2.2

Bugfix

  • Pre transform is being overriden by the inference transform

0.2.1

Added

  • Customizable number of channels at the output of the API models
  • API models expose output number of channels as a property
  • Added Encoder to the API
  • Sampled ModelNet dataset for point clouds