Commits

Breadcrumbs

History for

onmain

User selector

All users

Datepicker

All time

Commit History

Commits on Jul 15, 2024

• Optimization of large dataset loading solved: processes should be decided according to #cpu/#gpu. Now, we use multiprocess image loading by default. By the way, some names of variables in loss_dist…

  Show description for d3fe6ce
  

  Jianhao Huang

  committed
  d3fe6ce

Commits on Jul 14, 2024

• code should be formatted by black & add gsplat in environment.yml

  

  Jianhao Huang

  committed
  1c07351

Commits on Jul 13, 2024

• datasets sharded and local sampling is supported

  

  Jianhao Huang

  committed
  6b75bfd

• Reduce cpu memory consumption

  TarzanZhao

  committed
  cb161e6

Commits on Jun 30, 2024

• delete redundant image.close() for matrixcity dataset.

  TarzanZhao

  committed
  fcd7076

Commits on Jun 9, 2024

• refactor code.

  TarzanZhao

  committed
  7fb6230

Commits on Jun 5, 2024

• (1) create environment.yml for this rrepo (2) some minor changes.

  TarzanZhao

  committed
  3e1a138

Commits on May 29, 2024

• intermediate checkpoint for code clean up. TODO: check the render, visualization_generator.py, readme.md, and all todo in this repo.

  TarzanZhao

  committed
  fb448d7

Commits on May 19, 2024

• [IMPORTANT] all neurips experiments are using this commit and I keep all .sh launching scripts in this commit. I have fixed some bugs in this commit. And this commit need to be cleaned code(Idid no…

  Show description for 7e18dce
  TarzanZhao

  committed
  7e18dce

Commits on May 10, 2024

• fix the resolution mismatch between image and camera_inf. TODO: fix this from begin.

  

  Hexu Zhao

  committed
  48313c9

• support final_train2.py and add experiments flags and do many experiments.

  

  Jinyang Li

  committed
  a1bace4

Commits on May 8, 2024

• Implement final system and final system 2: support load balance in this system.

  

  Jinyang Li

  committed
  f309f84

Commits on May 6, 2024

• neurips init: use my codebase to train on large-scale scenes successfully. (1) add many debug flags (2) fix many small bugs. We should keep these fixs in all following code.

  

  Jinyang Li

  committed
  d785baa

Commits on Apr 30, 2024

• hack for 4 gpu per node cross-node training.

  

  Jinyang Li

  committed
  1e14e40

Commits on Apr 25, 2024

• Code for rubble experiment. (1) densify until memory > 17.5 (2) log memory usage even after densification finishes (3) distributed_load_ply (4) render with distributed_load

  

  Jinyang Li

  committed
  ea85c17

Commits on Apr 23, 2024

• (1) add distributed_save flag (2) add check_cpu_memory flag: release pil image memory after changing it to torch.tensor (3) modify render.py to fit the new code; however, I have not implemented the…

  Show description for 05c6821
  

  Jinyang Li

  committed
  05c6821

Commits on Apr 22, 2024

• add args.num_train_cameras

  

  Jinyang Li

  committed
  f57dd99

• add multiprocesses_image_loading; However, it does not have speed up now. I do not know the reason.

  

  Jinyang Li

  committed
  bf52649

Commits on Apr 21, 2024

• (1) add args.distributed_dataset_storage: load all images to gpu0 and then scatter to all gpu. (2) not check memory yet. (3) not do any overlapping yet.

  

  Jinyang Li

  committed
  2092de4

Commits on Apr 19, 2024

• add end2end timer. avoid measuring the saving checkpoint time.

  

  Jinyang Li

  committed
  d414b94

Commits on Apr 14, 2024

• Merge branch 'dist' into HEAD

  

  Hexu Zhao

  committed
  0d23cc7

• implemented different resolution for train and test

  Show description for 6823fcc
  

  Hexu Zhao

  committed
  6823fcc

Commits on Apr 13, 2024

• avoid using device_mesh because it requires torch 2.2;

  

  Hexu Zhao

  committed
  ac0efa4

Commits on Mar 29, 2024

• use handwriten allgather to replace torch.distributed.all_gather_object to speed up strategy.update_stats.

  

  Hexu Zhao

  committed
  7d56641

Commits on Mar 26, 2024

• 1.Update the rank and world size because we add DP and run across nodes and therefore more types of ranks and corresponding groups. 2. delete the timer output to screen.

  

  Hexu Zhao

  committed
  5ef6f35

Commits on Mar 24, 2024

• finished debug dp_system. the loss is correct now. we are able to use mp and dp together. dp could speed up the train scene training a lot. I also update the README.md. TODO: 1. run it on multiple …

  Show description for cc83604
  

  Hexu Zhao

  committed
  cc83604

• implement dp_system scratch. I have not debug for it now.

  TarzanZhao

  committed
  8ac1c16

Commits on Mar 19, 2024

• start to code for data parallel.

  

  Hexu Zhao

  committed
  8b6b1ff

Commits on Mar 14, 2024

• refactor code.

  

  Hexu Zhao

  committed
  ad6100c

Commits on Mar 12, 2024

• clean logging on screeen: add print_rk_k and etc.

  

  Hexu Zhao

  committed
  73e7c5f

Commits on Mar 4, 2024

• add img_dist_compile_mode flag: support faster modes for border pixel communication to speed up end2end training. consideration: functional communication primitive or non-functional, the former one…

  Show description for 7f8e7a1
  

  Hexu Zhao

  committed
  7f8e7a1

Commits on Mar 1, 2024

• enable change block size. TODO: fix the bug that I have to rebuild from scratch to make the change take effect.

  

  Hexu Zhao

  committed
  98ad688

Commits on Feb 19, 2024

• 1.support bsz>1 2. support args.performance_stats: make render.py and metrics.py run successfully again; change the stats logging and arguments checking code.

  

  Hexu Zhao

  committed
  6ee8e5e

Commits on Feb 12, 2024

• support redistribute 3dgs after few times densification to provide locality of 3dgs during rendering. it could reduce communication volumne, but increase unbalance of memory distribution. TODO: L1 …

  Show description for 21cd5ce
  

  Hexu Zhao

  committed
  21cd5ce

Commits on Feb 11, 2024

• refactor code for experiments: 1)move local_loss_computation before allreduce to correctly measure the local time for loss computation. 2) support experiments setting for fixed_training_image: only…

  Show description for 2a0aacb
  

  Hexu Zhao

  committed
  2a0aacb

Pagination

Previous

Next