search

facebookresearch / Detic

Code release for "Detecting Twenty-thousand Classes using Image-level Supervision".
Apache License 2.0
1.86k stars 211 forks source link

Is Deformable-DETR trained for ~48 epochs, not 58 epochs? #64

Open priancho opened 2 years ago

priancho commented 2 years ago

Hi, First, thank you for the great work :-)

While I was reading the paper, I found that the "Generalization to Deformable-DETR" section below mentions that the model is trained for ~48 epochs on LVIS data.

image

But the config file, configs/DeformDETR_L_R50_4x.yaml, shows that it trains 180k iterations with 32 batch size. https://github.com/facebookresearch/Detic/blob/bdf9eb82a9fb8f052b1511fbfc61bcabfc3cde51/configs/Base-DeformDETR_L_R50_4x.yaml#L30

Considering that the LVIS training data consists of about 100k images, the training epoch is 180k batch iterations * 32images/batch / 100k images/epoch should be ~58 epochs.

Am I missing something, or was 48 epochs typo?

Best regards, Han-Cheol

xingyizhou commented 2 years ago

Hi,

You are correct. It should be 58 epochs. We will correct this in future revisions. By 48 epochs we are following the COCO naming convention where 1x = 12 epochs. Apparently this is not true for LVIS.

Best, Xingyi