Can you provide me Pascal VOC configs?

[seoha-kim]

Hello.
I'd like to try to replicate your great distortion methods at Pascal VOC.
However, this repository only provides MS-coco configurations.
you used extra 6k iterations for auxiliary task warm-up on the Pascal VOC dataset,
how can I reproduce this training details?

Is it alright to train teacher standalone / and distill students?

teacher standalone train
BASE: "../Base-RCNN-FPN.yaml"
MODEL:
WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
MASK_ON: False
RESNETS:
DEPTH: 101
ROI_HEADS:
NUM_CLASSES: 20
INPUT:
MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
MIN_SIZE_TEST: 800
DATASETS:
TRAIN: ('voc_2007_trainval', 'voc_2012_trainval')
TEST: ('voc_2007_test')
SOLVER:
STEPS: (12000, 16000)
MAX_ITER: 18000 # 17.4 epochs

student distillation
BASE: "../Base-RCNN-FPN.yaml"
MODEL:
WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
MASK_ON: False
RESNETS:
DEPTH: 50
ROI_HEADS:
NUM_CLASSES: 20
DISTILLER:
MODEL_LOAD_OFFICIAL: False
MODEL_DISTILLER_CONFIG: 'PascalVOC-Detection/faster_rcnn_R_101_FPN.yaml'
INS_ATT_MIMIC:
WEIGHT_VALUE: 3.0
INS:
INPUT_FEATS: ['p2', 'p3', 'p4', 'p5', 'p6']
MAX_LABELS: 100
INPUT:
MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
MIN_SIZE_TEST: 800
DATASETS:
TRAIN: ('voc_2007_trainval', 'voc_2012_trainval')
TEST: ('voc_2007_test',)
SOLVER:
STEPS: (12000, 16000)
MAX_ITER: 18000 # 17.4 epochs
CLIP_GRADIENTS: {"ENABLED": True}

[seoha-kim]

If I modify the cfg.If MODEL.DISTILLER.BYPASS_DISTILL = 1000 to 7000, can only decoder be trained during 6000iter?
I wonder if this option is implemented in this repository.

[AmoseKang]

Hello, thanks for your question.
The config you used is basically right (the base voc config is from d2), however, there are few things I need to highlight.

In general, to reproduce the result, there are three steps you need to do:

1. Train a teacher network with R101 backbone, this is basically the teacher standalone train config your provided,
2. Create a new teacher standalone config, replace the 'WEIGHTS' to the checkpoint of step 1,
3. Train and distill the student, you should change MODEL_DISTILLER_CONFIG to the config from step2, and set MODEL.DISTILLER.PRETRAIN_TEACHER_ITERS to 6000, leave BYPASS_DISTILL to 1000. This will append an extra warmup stage before distillation.

[seoha-kim]

Thank you for the detailed instructions.
Thanks to you, I was able to re-produce it well. 😀