About R101 corresponding configs and trained models
ShihuaHuang95 opened this issue · 2 comments
Many thanks for this great work. I am wondering if you have any plan to release the R101 corresponding configs and trained models.
I have tried out the following config and just got PQ 45.6.
MODEL:
META_ARCHITECTURE: "PanopticFCN"
WEIGHTS: detectron2://ImageNetPretrained/MSRA/R-101.pkl
MASK_ON: True
PIXEL_MEAN: [123.675, 116.28, 103.53]
PIXEL_STD: [1.0, 1.0, 1.0]
RESNETS:
DEPTH: 101
OUT_FEATURES: ["res2", "res3", "res4", "res5"]
FPN:
IN_FEATURES: ["res2", "res3", "res4", "res5"]
DATASETS:
TRAIN: ("coco_2017_train_panoptic_separated",)
TEST: ("coco_2017_val_panoptic_separated",)
DATALOADER:
FILTER_EMPTY_ANNOTATIONS: True
SOLVER:
BASE_LR: 0.01
WEIGHT_DECAY: 1e-4
LR_SCHEDULER_NAME: "WarmupPolyLR"
POLY_LR_POWER: 0.9
WARMUP_ITERS: 1000
WARMUP_FACTOR: 0.001
WARMUP_METHOD: "linear"
CLIP_GRADIENTS:
ENABLED: True
CLIP_VALUE: 35.0
IMS_PER_BATCH: 16
MAX_ITER: 270000
CHECKPOINT_PERIOD: 10000
INPUT:
MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
MIN_SIZE_TRAIN_SAMPLING: "choice"
MIN_SIZE_TEST: 800
MAX_SIZE_TRAIN: 1333
MAX_SIZE_TEST: 1333
MASK_FORMAT: "bitmask"
VERSION: 2
OUTPUT_DIR: "./PanopticFCN_r101_3x_FAM"
Hi, Thanks for your interest. The config and performance seem right and normal.