segformer_mit-b0_512x512_160k_ade20k.yaml

DATA:
    BATCH_SIZE: 2  # per GPU [total bs is set to 8 or 16]
    BATCH_SIZE_VAL: 1  # per GPU
    DATASET: 'ADE20K' # dataset name
    DATA_PATH: '/home/aistudio/data/ADEChallengeData2016'
    CROP_SIZE: (512,512)  # input_size (training)
    NUM_CLASSES: 150
MODEL:
    NAME: 'Segformer'
    ENCODER:
        TYPE: 'MixVisionTransformer'
        OUT_INDICES: [0,1,2,3]
    PRETRAINED: None
    DECODER_TYPE: 'SegformerHead'
    SEGFORMER:
        IN_CHANNELS: [32, 64, 160, 256]    # BO is half of B1-B5
        CHANNELS: 256
        ALIGN_CORNERS: False
    TRANS:
        IN_CHANNELS: 3
        EMBED_DIM: 32               # BO is half of B1-B5
        NUM_STAGES: 4 
        NUM_LAYERS: [2, 2, 2, 2]    # BO-B5 differs
        NUM_HEADS: [1, 2, 5, 8]
        PATCH_SIZE: [7, 3, 3, 3]
        STRIDES: [4, 2, 2, 2]
        SR_RATIOS: [8, 4, 2, 1]
        HIDDEN_SIZE: 768 
        MLP_RATIO: 4             
        QKV_BIAS: True
    DROPOUT: 0.0
    ATTENTION_DROPOUT: 0.0
    DROP_PATH: 0.1

TRAIN:
    BASE_LR: 0.00006
    END_LR: 0.0
    DECODER_LR_COEF: 10.0
    ITERS: 2000
    POWER: 1.0
    DECAY_STEPS: 2000
    LR_SCHEDULER:
        NAME: 'PolynomialDecay'
    OPTIMIZER:
        WEIGHT_DECAY: 0.01
        GRAD_CLIP: 1.0
        NAME: 'AdamW'
        MOMENTUM: 0.9
VAL:
    MULTI_SCALES_VAL: False
    IMAGE_BASE_SIZE: 512                                                                                                                                                                                            
    CROP_SIZE: [512, 512]
    STRIDE_SIZE: [512, 512]
SAVE_DIR: "./output/segformer_mit-b0_512x512_160k_ade20k"