psalm/mask_config/Base-segmention.yaml

MODEL:
  BACKBONE: 
    TYPE: 'swin' # 'resnet' or 'swin'
  PRETRAINED_WEIGHTS: 
  IS_TRAINING: True    
  RESNETS:
    DEPTH: 50
    STEM_OUT_CHANNELS: 64
    STRIDE_IN_1X1: False    
    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
  SWIN:
    TYPE: "base" # "tiny" or "small" or "base" or "large"
    EMBED_DIM: 96
    DEPTHS: [2 2 6 2]
    NUM_HEADS: [3 6 12 24]
    PATCH_SIZE: 4
    WINDOW_SIZE: 7
    MLP_RATIO: 4.
    QKV_BIAS: True
    QK_SCALE: 
    DROP_RATE: 0.
    ATTN_DROP_RATE: 0.
    DROP_PATH_RATE: 0.3
    APE: False
    PATCH_NORM: True
    OUT_INDICES: (0 1 2 3)
    PRETRAIN_IMG_SIZE: 384
    USE_CHECKPOINT: False
    OUT_FEATURES: ["res2", "res3", "res4", "res5"]
DATASETS:
  TRAIN: 'dataset/training.odgt'
  VALID: 'dataset/validation.odgt'
  ROOT_DIR: 'nuImages/ImageData/nuimages-v1.0-all-samples/'
  PIXEL_MEAN: [0.485, 0.456, 0.406]
  PIXEL_STD: [0.229, 0.224, 0.225]

SOLVER:
  IMS_PER_BATCH: 16
  BASE_LR: 0.0001
  MAX_ITER: 160000
  WARMUP_FACTOR: 1.0
  WARMUP_ITERS: 0
  WEIGHT_DECAY: 0.05
  OPTIMIZER: "ADAMW"
  LR_SCHEDULER_NAME: "WarmupPolyLR"
  BACKBONE_MULTIPLIER: 0.1
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 0.01
    NORM_TYPE: 2.0
  AMP:
    ENABLED: True
INPUT:
  MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 640) for x in range(5, 21)]"]
  MIN_SIZE_TRAIN_SAMPLING: "choice"
  CROP:
    ENABLED: True
    TYPE: "absolute"
    SIZE: [224, 320, 480, 512] # [640, 800, 960, 1120]
    MAX_SIZE: [1024, 576] # [width, height]
    SINGLE_CATEGORY_MAX_AREA: 1.0
  COLOR_AUG_SSD: True
  SIZE_DIVISIBILITY: 640  # used in dataset mapper
  FORMAT: "RGB"
  DATASET_MAPPER_NAME: "mask_former_instance"
TRAIN:
  LOG_DIR: 'logs'
  CKPT_DIR: 'ckpt'  
  BATCH_SIZE: 9
  WORKERS: 8
  EPOCH: 300
SOLVER:
  LR: 0.00006
  OPTIMIZER: "ADAMW"
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 0.01
    NORM_TYPE: 2.0
TEST:
  EVAL_PERIOD: 5000
  TEST_DIR: 'test'
  SAVE_DIR: 'output'
  AUG:
    ENABLED: False
    MIN_SIZES: [320, 480, 640, 800, 960, 1120]
    MAX_SIZE: 4480
    FLIP: True
DATALOADER:
  FILTER_EMPTY_ANNOTATIONS: True
  NUM_WORKERS: 4
VERSION: 2