Skip to content

Logical batch size. #1177

New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Jump to bottom
Closed
StuLiu opened this issue Jan 4, 2022 · 3 comments
Closed

Logical batch size. #1177

StuLiu opened this issue Jan 4, 2022 · 3 comments
Assignees
@MengzhangLI

Comments

@StuLiu
Copy link

StuLiu commented Jan 4, 2022

There is one GPU only in my computer. Therefore, I can't reproduce the metrix produced by 4 GPU as shown in your readme.md.
Could you please implement the function of 'logical batch size'. So I can train with only one GPU at the cost of training time, and reproduce the similar metrix.
@MengzhangLI
Copy link
Contributor

MengzhangLI commented Jan 4, 2022 •
edited
Loading

Hi, you can train with one gpu instead of four gpu by changing corresponding samples_per_gpu to four times as before, and set only one gpu to train. The related batch size could be change here:

mmsegmentation/configs/bisenetv1/bisenetv1_r18-d32_4x4_1024x1024_160k_cityscapes.py

Line 9 in ff9d695

samples_per_gpu=4,

@MengzhangLI MengzhangLI self-assigned this Jan 4, 2022
@StuLiu
Copy link
Author

StuLiu commented Jan 4, 2022

Hi, you can train with one gpu instead of four gpu by changing corresponding samples_per_gpu to four times as before, and set only one gpu to train. The related batch size could be change here:

mmsegmentation/configs/bisenetv1/bisenetv1_r18-d32_4x4_1024x1024_160k_cityscapes.py

Line 9 in ff9d695

samples_per_gpu=4,

Thanks for your instruction. However, the OOM Error was caused by limited memory (24GB is still not enough).

@MengzhangLI
Copy link
Contributor

Try to use FP16, have a look on this config:

https://github.com/open-mmlab/mmsegmentation/blob/master/configs/bisenetv2/bisenetv2_fcn_fp16_4x4_1024x1024_160k_cityscapes.py

aravind-h-v pushed a commit to aravind-h-v/mmsegmentation that referenced this issue Mar 27, 2023
@anton-l
Fix cpu offloading (open-mmlab#1177)
24895a1 
* Fix cpu offloading

* get offloaded devices locally for SD pipelines
wjkim81 pushed a commit to wjkim81/mmsegmentation that referenced this issue Dec 3, 2023
@ly015 @jin-s13
@luminxu @zengwang430521 @mzr1996
[Feature] Add Webcam API (open-mmlab#1178)
fa07051 
* add cam runner [wip]

* update webcam runner and node api

* reorganize module structure

* add model node

* add pose model node [wip]

* update message

* add monitor node

* rename config

* refactor webcam util

* add update_route_info

* add input_check_interval

* add fps for one-euro in mmpose node

* refactor visualition nodes

* fix bug

* modify runner argument and fix bug

* add doc

* add doc

* add webcam_runner docstrings

* add docstrings

* update interface and docstring

* add RecorderNode

* * add syncrhonous result binding mode
* refactor node event handler

* fix typing

* fix typing (open-mmlab#1082)

* * Add RecorderNode for saving video to files
* Add exiting handler method to nodes
* Add VideoEndingMessage

* rename _on_exit ot on_exit

* fix typing

* Fix a few issues

* register node event handler in __init__
* change default pose model to vipnas_mblv3
* Node._on_exit -> Node.on_exit

* fix typing

* fix typing

* change from thread based to process based

* solve cuda distributed [wip]

* fix process creating method and buffer size

* thread based buffer

* node based on thread

* node set daemon=False as default

* fix logging level and add docstrings

* modify register_event

* * reorganize webcam tool configs
* modify frame rate calculation
* rename BillboardNode as NoticeBoardNode

* [WebcamAPI] Add demo README (open-mmlab#1133)

* add README.md

* update gif

* [WebcamAPI] Add facwswap node and config (open-mmlab#1094)

* add facwswap node

* adapt to webcam_api

* adapt to webcam_api

* move config

* add sync mode in runner

* fix bugs and update nodes

* [WebcamAPI] Supersaiyan (open-mmlab#1098)

* add supersaiyan effects

* fix init

* apply background

* update config

* apply saiyan effects

* rebase

* rebase and fix

* update doc

* update docs

* add load image from disk or url

* make the resources to url links

* mv load_image_from_disk_or_url to utils/misc

* move effect funcs to webcam_api

* rebase

* fix pytest

* fix docs

* add get_local_path_given_url

* add get_local_path_given_url

* rename to get_cached_file_path

* update readme

* update readme

* fix bug that only one bbox is detected for each class

* fix faceswap config

* add newyear demo (open-mmlab#1143)

Co-authored-by: zengwang <zengwang430521@gmail.com>

* [Feature] Add a `SpriteNode` for Webcam API (open-mmlab#1160)

* Fix missing flag in `load_image_from_disk_or_url`

* Add SpriteNode

* Meow dwen dwen (open-mmlab#1173)

* fix mmpose nodes

* add dummy video capture

* add xdwendwen node

* fix configs

* Update README.md

* [WebcamAPI] Update XDwenDwen example (open-mmlab#1177)

* Update XDwenDwen example

* add rigid transform
* update configs and readme files

* extend meow dwen dwen

* update config

* support cfg-options

* turn-off default help instruction

* fix typo

Co-authored-by: Jas <jinsheng@sensetime.com>
Co-authored-by: Lumin <30328525+luminxu@users.noreply.github.com>
Co-authored-by: zengwang <zengwang430521@gmail.com>
Co-authored-by: Ma Zerun <mzr1996@163.com>
# for free to join this conversation on GitHub. Already have an account? # to comment
Assignees

@MengzhangLI MengzhangLI

Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@StuLiu @MengzhangLI