Skip to content

fix: lr_multiplier is unused and scaled for 100e instead of 50e - #73

Open
andrewwhitecdw wants to merge 1 commit into
NVlabs:mainfrom
andrewwhitecdw:bugfix/odise-caption-coco-50e-7764b6cd
Open

fix: lr_multiplier is unused and scaled for 100e instead of 50e#73
andrewwhitecdw wants to merge 1 commit into
NVlabs:mainfrom
andrewwhitecdw:bugfix/odise-caption-coco-50e-7764b6cd

Conversation

@andrewwhitecdw

Copy link
Copy Markdown

Problem

fix: lr_multiplier is unused and scaled for 100e instead of 50e

Fix

Replace:

lr_multiplier = L(WarmupParamScheduler)(
    scheduler=L(MultiStepParamScheduler)(
        values=[1.0, 0.1, 0.01],
        # assume 100e with batch-size 64 as original LSJ
        # Equivalent to 100 epochs.
        # 100 ep = 184375 iters * 64 images/iter / 118000 images/ep
        milestones=[163889, 177546],
        num_updates=184375,
    ),
    # for warmup length we adopted COCO LSJ setting
    warmup_length=500 / 184375,
    warmup_factor=0.067,
)

optimizer.lr = 1e-4
optimizer.weight_decay = 0.05

with:

lr_multiplier = L(WarmupParamScheduler)(
    scheduler=L(MultiStepParamScheduler)(
        values=[1.0, 0.1, 0.01],
        # 50e schedule scaled from the 100e LSJ setting
        milestones=[81944, 88773],
        num_updates=92188,
    ),
    # for warmup length we adopted COCO LSJ setting
    warmup_length=500 / 92188,
    warmup_factor=0.067,
)

train.lr_multiplier = lr_multiplier

optimizer.lr = 1e-4
optimizer.weight_decay = 0.05

Files changed

  • configs/Panoptic/odise_caption_coco_50e.py

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant