upernet_swin_tiny_patch4_windown7_512x512_160k_ade20k.yaml

DATA:
    BATCH_SIZE: 2  # per GPU [total bs is set to 8 or 16]
    BATCH_SIZE_VAL: 4  # 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_Swin'
    ENCODER:
        TYPE: 'SwinTransformer'
        OUT_INDICES: [0, 1, 2, 3]   # stage_i
    PRETRAINED: './pretrain_models/backbones/vit_large_patch16_224.pdparams'
    DECODER_TYPE: 'UperHead'
    UPERHEAD:
        IN_CHANNELS: [96, 192, 384, 768]
        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
        WINDOW_SIZE: 7
        IN_CHANNELS: 3
        HIDDEN_SIZE: 96 # 768(Base), 1024(Large), 1280(Huge)
        EMBED_DIM: 96
        STAGE_DEPTHS: [2, 2, 6, 2]
        NUM_HEADS: [3, 6, 12, 24]    
        MLP_RATIO: 4
        QKV_BIAS: True
        QK_SCALE: None
        APE: False  # absolute positional embeddings
        PATCH_NORM: True
    AUX:
        AUXIHEAD: True
    AUXFCN:
        IN_CHANNELS: 384
        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: 576
    CROP_SIZE: [512,512]
    STRIDE_SIZE: [341,341]
SAVE_DIR: "./output/UperNet_swin_tiny_patch4_windown7_512x512_160k_ade20k"