PASCAL VOC dataset settings

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection

https://arxiv.org/abs/2102.09480

MIT License

415 stars 83 forks source link

PASCAL VOC dataset settings #5

Closed prinshul closed 3 years ago

prinshul commented 3 years ago

Can you please mention the settings for PASCAL VOC dataset like Image size, batch size (labeled/unlabeled), total iterations, epochs?

ycliu93 commented 3 years ago

Here are the details of VOC experiments. You could directly use the following parameters in your config files.

INPUT:
  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
  MIN_SIZE_TEST: 800
SOLVER:
  MAX_ITER: 180000 
  IMG_PER_BATCH_LABEL: 32
  IMG_PER_BATCH_UNLABEL: 32
  BASE_LR: 0.001

For other parameters, you could check the last page of the appendix.https://arxiv.org/pdf/2102.09480.pdf

namihagi commented 3 years ago

I have other question about voc setting. BURN_UP_STEP is 45k iterations in your VOC setting?