scannet200_swin.sh

set -e

export DETECTRON2_DATASETS="/projects/katefgroup/language_grounding/SEMSEG_100k"
SCANNET200_DATA_DIR="/path/to/train_validation_database.yaml"

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0,1 python train_odin.py  --dist-url='tcp://127.0.0.1:6578' --num-gpus 2 --config-file configs/scannet_context/swin_3d.yaml \
OUTPUT_DIR /projects/katefgroup/language_grounding/bdetr2/arxiv_reproduce/scannet_swin200_1e5 SOLVER.IMS_PER_BATCH 4 \
SOLVER.CHECKPOINT_PERIOD 4000 TEST.EVAL_PERIOD 4000 \
INPUT.FRAME_LEFT 7 INPUT.FRAME_RIGHT 7 INPUT.SAMPLING_FRAME_NUM 15 \
MODEL.WEIGHTS '/projects/katefgroup/language_grounding/odin_arxiv/m2f_coco_swin.pkl' \
SOLVER.BASE_LR 1e-4 \
INPUT.IMAGE_SIZE 512 \
MODEL.CROSS_VIEW_CONTEXTUALIZE True \
INPUT.CAMERA_DROP True \
INPUT.STRONG_AUGS True \
INPUT.AUGMENT_3D True \
INPUT.VOXELIZE True \
INPUT.SAMPLE_CHUNK_AUG True \
MODEL.MASK_FORMER.TRAIN_NUM_POINTS 50000 \
MODEL.CROSS_VIEW_BACKBONE True \
DATASETS.TRAIN "('scannet200_context_instance_train_200cls_single_highres_100k',)" \
DATASETS.TEST "('scannet200_context_instance_val_200cls_single_highres_100k','scannet200_context_instance_train_eval_200cls_single_highres_100k')" \
MODEL.PIXEL_DECODER_PANET True \
MODEL.SEM_SEG_HEAD.NUM_CLASSES 200 \
MODEL.MASK_FORMER.TEST.SEMANTIC_ON True \
SKIP_CLASSES "[119, 200]" \
USE_GHOST_POINTS True \
MODEL.FREEZE_BACKBONE False \
SOLVER.TEST_IMS_PER_BATCH 2 \
SAMPLING_STRATEGY "consecutive" \
USE_SEGMENTS True \
SOLVER.MAX_ITER 100000 \
DATALOADER.NUM_WORKERS 8 \
DATALOADER.TEST_NUM_WORKERS 2 \
MAX_FRAME_NUM -1 \
MODEL.MASK_FORMER.DICE_WEIGHT 6.0 \
MODEL.MASK_FORMER.MASK_WEIGHT 15.0 \
USE_WANDB True \
USE_MLP_POSITIONAL_ENCODING True \
INPUT.MIN_SIZE_TEST 512 \
INPUT.MAX_SIZE_TEST 512 \
HIGH_RES_SUBSAMPLE True \
SCANNET200_DATA_DIR $SCANNET200_DATA_DIR

# MODEL.WEIGHTS '/projects/katefgroup/language_grounding/odin_arxiv/m2f_coco.pkl' \
# reduce lr at 76k iterations to 1e-5 and get the best checkpoint at 5.5k for instance \
# segmentation and 3k for semantic segmentation