GitLab

* add cfg

* add eps format

* update README

* update model zoo

* add paconv readme

* add metafile

* hack drop_ratio

* use base schedule cfg

* remove paconv cuda

* monor fix

* remove paconv cuda config related

* update paconv benchmark

* update model link

* add PAConv decode head

* add config files

* add paconv's correlation loss

* support reg loss in Segmentor class

* minor fix

* add augmentation to configs

* fix ed7 in cfg

* fix bug in corr loss

* enable syncbn in paconv

* rename to loss_regularization

* rename loss_reg to loss_regularize

* use SyncBN

* change weight kernels to kernel weights

* rename corr_loss to reg_loss

* minor fix

* configs fix IndoorPatchPointSample

* fix grouped points minus center error

* update transform_3d & add configs

* merge master

* fix enlarge_size bug

* refine config

* remove cfg files

* minor fix

* add comments on PAConv's ScoreNet

* refine comments

* update compatibility doc

* remove useless lines in transforms_3d

* rename with_loss_regularization to with_regularization_loss

* revert palette change

* remove xavier init from PAConv's ScoreNet

* group-free-3d head

* GroupFree3DNet->VoteNet

* modify docstring

* bugfix: calculate pts_instance_label, decoder self/cross posembed init

* support GroupFree3DNet, modify scannnet train config

* support point cloud input features dim = 0

* add groupfree3dnet test case

* bugfix: softmax in decode boxes

* support multi-stage predictions

* modify GroupFree3DMultiheadAttention input parameters

* refactor: support sunrgbd-based train

* refactor: support sunrgbd-based train

* fix parts of bug

* modify multi-stage prediction

* fixbug: conv_channels

* bugfix: permute

* bugfix: permute

* bugfix: expand

* fix MAX_NUM_OBJ=64

* 4 gpu training, score_thr = 0.0

* modify config, repeattime=1

* bigfix: expand

* modify: GroupFree3DMHA, build_positional_encoding

* modify: GroupFree3DMHA, build_positional_encoding

* bugfix: torch.nn

* bugfix: mean loss

* residual -> identity

* fix name: DropOut -> Dropout

* delete sunrgbd-based congfig

* Fix: trailing whitespace

* suffix -> prefix

* bugfix: groupfree3d config
Co-authored-by: jinhui <PJLAB\jinhui@shai14001019l.pjlab.org>

* refactor K and Rt to depth2img for SUN RGB-D

* fix lint

* update 3 tests

* fix extra calib key and comments

* remove calib from browse_dataset

* fix cam to depth; rename return_z

* first try of kitti

* fix python3

* imvoxelnet is ready for open-mmlab/mmdetection3d

* apply pre-commit

* update to ConvModule and AlignedAnchor3DGenerator

* add unit tests

* fix torch.Tensor in docstrings

* revert anchor ranges

* refactor PN SA module code

* refactor PAConv op & add PAConvSAModule

* add unit test

* add PAConvCUDA SA module

* add unit test

* minor fix

* merge master

* move paconv sa modules to a new file

* rename to BasePointSAModule

* rename to PAConvCUDASAModule

* delete typing comment

* refine docs

* rename mlp_spec to mlp_channel

* fix typos in comment

* Fixed wrong config paths and fixed a bug in test

* Fixed metafile

* support knn query in QueryAndGroup

* add cuda implemented assign_scores op

* add unit test for paconv assign_score op

* refactor op

* add non-cuda & cuda version PAConv

* minor fix

* fix cuda-9.0 compatibility

* add weight init of paconv

* fix typos

* refactor paconv

* build BaseSegmentor for point sem seg

* add encoder-decoder segmentor

* update mmseg dependency

* fix linting errors

* warp predicted seg_mask in dict

* add unit test

* use build_model to wrap detector and segmentor

* fix compatibility with mmseg

* faster sliding inference

* merge master

* configs for training on ScanNet

* fix CI errors

* add comments & fix typos

* hard-code class_weight into configs

* fix logger bugs

* update segmentor unit test

* logger use mmdet3d

* use eps to replace hard-coded 1e-3

* add comments

* replace np operation with torch code

* add comments for class_weight

* add comment for BaseSegmentor.simple_test

* rewrite EncoderDecoder3D to avoid inheriting from mmseg

* support PN2 decode head

* add mmseg dependency in github workflow

* complete PN2 decode head

* modify backbone pn2 to support seg task & its unit test

* add unit test for PN2 decode_head

* Support base mono3d dense head and anchor free mono3d head

* Support FCOS3D head

* Support FCOS3D baseline on nuScenes

* Fix an import error caused by update of mmcv/mmdet

* Change img_scale to scale_factor in the MultiScaleFlipAug in the config

* Add pred_bbox2d in the params of anchor_free_mono3d_head

* Add unit test for fcos3d head

* Fix a minor bug when setting img_metas in the unit test

* Add unit test for fcos3d detector

* Simplify the logic of weights initialization

* Add comments to specify the reason of cloning features

* Update head config

- fix 'invalid configuration argument' error triggered by empty point input. test cases covering similar situations are added to test_dynamic_scatter.py as well.
trivial changes:
- switch to using torch::unique_dim to generate reduce mapping instead of calculating it from scratch.

* modify vote_head to support 3dssd

* delete .keys()

* add 3dssd unittest

* 3dssd->ssd3d

* Added image loading in SUNRGB-D dataset (#195)

* image loading

* format and docstring fix

* removed irrelevant files

* removed irrelevant files

* load image only if modality is pc+img

* added modality like nuscenes

* Added imvotenet image branch pretrain (#217)

* image loading

* naive commit

* format and docstring fix

* removed irrelevant files

* removed irrelevant files

* load image only if modality is pc+img

* added modality like nuscenes

* pretrain_2d_model

* finetune sunrgbd

* finetune sunrgbd

* deleted unused code

* fixed a bug

* resolve conflict

* update config file

* fix docstring and configs

* integrated vote fusion

* coords transform and unit test

* Update docstring

* refactor and add unit test

* fix bug caused by mmcv upgrade; delete pdb breakpoint

* add point fusion unittest

* remove unused file

* fix typos

* updates

* add assertion info

* update

* add unittest

* add vote fusion unittest

* add vote fusion unittest

* [Refactor] VoteNet refactor (#322)

* votenet refactor

* remove file

* minor update

* docstring

* initial update of imvotenet

* [Feature] Support vote fusion (#297)

* integrated vote fusion

* coords transform and unit test

* Update docstring

* refactor and add unit test

* add point fusion unittest

* remove unused file

* updates

* add assertion info

* update

* add unittest

* add vote fusion unittest

* add vote fusion unittest

* minor update

* docstring

* change np ops to torch

* refactor test

* update

* refactor of image mlp and np random ops to torch

* add docstring

* add config and mod dataset

* fix bugs

* add_comments

* fix bugs

* fix_bug

* fix bug

* fix bug

* fix bug

* fix bug

* final fix

* fix bug

* ?

* add docstring

* move train/test cfg

* change img mlp default param

* rename config

* minor mod

* change config name

* move train/test cfg

* some fixes and 2d utils

* fix config name

* fix config override issue

* config simplify & reformat

* explicitly set eval mode->override train()

* add fix_img_branch to config

* remove set_img_branch_eval_mode

* temporal fix, change calibs to calib

* more docstring and view/reshape, expand/repeat change

* complete imvotenet docstring

* fix docstring

* add config and some minor fix

* rename config
Co-authored-by: ZwwWayne <wayne.zw@outlook.com>

* support knn gpu op

* made it more robust and fixed comments

* bugfix and add warning in docstring

* support knn gpu op

* made it more robust and fixed comments

* a faster & more memory-efficient implementation of DynamicScatter

* fix format issues and add pytest skip code for tests on machines without cuda support

* some trivial changes:

decrease the number of kernel threads per block to 512, to enable inference on GPUs with computing capability lower than 2.0

change the backpropagation behavior of max-reduction. when there are multiple points shares the same maximum feature value, only the first point (with lowest row index) among them is chosen to propagate the output gradient back. before this change, all points with the same maximum feature value can propagate the output gradient back. this change makes the max-reduction behaves in consistence with torch.max. this change may cause gradcheck failure in test_dynamic_scatter.py. please do not worry about it because torch.max fails the gradcheck too.

* fix typo
Co-authored-by: zhanggefan <1152009@tongji.edu.cn>

* fix bug caused by mmcv upgrade; delete pdb breakpoint

* fix typos

* use torch.cuda api

* update unittest for points_in_bbox_gpu and points_in_boxes_batch

* Added comments for explanation

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg/test_cfg to model

* Move train_cfg and test_cfg into model

* modify centerpoint configs

* Modify docs

* modify build_detector

* modify test_config_build_detector

* modify build_detector parameters

* Adopt the same strategy in build_detector

* fix bug caused by mmcv upgrade; delete pdb breakpoint

* fix typos

* Adjust the directory structure of tests

* Adopted the new directory via _get_config_directory

* Adopted the new directory via _get_config_directory

* Adopted the new directory via _get_config_directory

* modified the wrong code comments