upernet_cswin_small_patch4_512x512_160k_ade20k.yaml

DATA:
    BATCH_SIZE: 1  # per GPU [total bs is set to 8 or 16]
    BATCH_SIZE_VAL: 1  # per GPU
    DATASET: 'ADE20K' # dataset name
    DATA_PATH: '/home/ssd3/wutianyi/datasets/ADEChallengeData2016'
    CROP_SIZE: (512,512)  # input_size (training)
    NUM_CLASSES: 150
MODEL:
    NAME: 'UperNet_CSwin'
    DROPOUT: 0.0   # dropout rate for linear projection
    ATTENTION_DROPOUT: 0.0  # dropout rate for attention
    DROP_PATH: 0.2
    ENCODER:
        TYPE: 'CSwinTransformer'
        OUT_INDICES: [0, 1, 2, 3]   # stage_i
    PRETRAINED: './pretrain_models/backbones/cswin_small_224.pdparams'
    DECODER_TYPE: 'UperHead'
    UPERHEAD:
        IN_CHANNELS: [64, 128, 256, 512]
        IN_INDEX: [0, 1, 2, 3]
        POOL_SCALES: [1, 2, 3, 6]
        CHANNELS: 512
        DROP_RATIO: 0.1
        ALIGN_CORNERS: False
    TRANS:
        PATCH_SIZE: 4
        IN_CHANNELS: 3
        HIDDEN_SIZE: 64  # 64(tiny, small), 96(base), 144(large)
        EMBED_DIM: 64
        STAGE_DEPTHS: [2, 4, 32, 2]
        NUM_HEADS: [2, 4, 8, 16]
        SPLIT_SIZES: [1, 2, 7, 7]       # cswin
        MLP_RATIO: 4
        QKV_BIAS: True
        QK_SCALE: None
        APE: False  # absolute positional embeddings
        PATCH_NORM: True
    AUX:
        AUXIHEAD: True
    AUXFCN:
        IN_CHANNELS: 256   # channel of the 1/16 resolution features
        UP_RATIO: 16

TRAIN:
    BASE_LR: 0.00006
    END_LR: 1e-4
    DECODER_LR_COEF: 10.0
    ITERS: 160000
    POWER: 0.9
    DECAY_STEPS: 160000
    LR_SCHEDULER:
        NAME: 'PolynomialDecay'
    OPTIMIZER:
        WEIGHT_DECAY: 0.0
        GRAD_CLIP: 1.0
        NAME: 'SGD'
        MOMENTUM: 0.9
VAL:
    MULTI_SCALES_VAL: False
    SCALE_RATIOS: [0.5, 0.75, 1.0]
    IMAGE_BASE_SIZE: 512
    CROP_SIZE: [512,512]
    STRIDE_SIZE: [341,341]
SAVE_DIR: "./output/UperNet_cswin_small_patch4_512x512_160k_ade20k"