Created by: iceriver97

我在设置 CROSS_ENTROPY_WEIGHT=“dynamic”后出现了loss开始就一直为0.0001的情况：

epoch=1 step=20 lr=0.00100 loss=0.0001 step/sec=0.632 | ETA 07:54:12
epoch=1 step=30 lr=0.00100 loss=0.0001 step/sec=0.618 | ETA 08:04:20
epoch=1 step=40 lr=0.00100 loss=0.0001 step/sec=0.621 | ETA 08:02:01
epoch=1 step=50 lr=0.00100 loss=0.0001 step/sec=0.619 | ETA 08:03:05
epoch=1 step=60 lr=0.00100 loss=0.0001 step/sec=0.621 | ETA 08:01:21
epoch=1 step=70 lr=0.00100 loss=0.0001 step/sec=0.617 | ETA 08:03:57
epoch=1 step=80 lr=0.00100 loss=0.0001 step/sec=0.616 | ETA 08:04:27
epoch=1 step=90 lr=0.00100 loss=0.0001 step/sec=0.613 | ETA 08:06:52
epoch=2 step=110 lr=0.00099 loss=0.0001 step/sec=0.621 | ETA 08:00:16
epoch=2 step=120 lr=0.00099 loss=0.0001 step/sec=0.619 | ETA 08:01:04
epoch=2 step=130 lr=0.00099 loss=0.0001 step/sec=0.610 | ETA 08:08:02
epoch=2 step=140 lr=0.00099 loss=0.0001 step/sec=0.620 | ETA 08:00:04
epoch=2 step=150 lr=0.00099 loss=0.0001 step/sec=0.618 | ETA 08:01:07
epoch=2 step=160 lr=0.00099 loss=0.0001 step/sec=0.624 | ETA 07:56:38
epoch=2 step=170 lr=0.00099 loss=0.0001 step/sec=0.613 | ETA 08:04:29
epoch=2 step=180 lr=0.00099 loss=0.0001 step/sec=0.621 | ETA 07:58:18

下面是配置文件内容：

# 数据集配置
DATASET:
    DATA_DIR: "./work/dataset/land_train/"
    NUM_CLASSES: 7
    TEST_FILE_LIST: "./work/dataset/land_train/test.txt"
    TRAIN_FILE_LIST: "./work/dataset/land_train/train.txt"
    VAL_FILE_LIST: "./work/dataset/land_train/val.txt"
    VIS_FILE_LIST: "./work/dataset/land_train/test.txt"

# 预训练模型配置
MODEL:
    MODEL_NAME: "ocrnet"
    DEFAULT_NORM_TYPE: "bn"
    # 此处设置backbone的配置
    HRNET:
        STAGE2:
            NUM_CHANNELS: [18, 36]
        STAGE3:
            NUM_CHANNELS: [18, 36, 72]
        STAGE4:
            NUM_CHANNELS: [18, 36, 72, 144]
    # 此处设置OCR HEAD
    OCR:
        OCR_MID_CHANNELS: 512
        OCR_KEY_CHANNELS: 256
    MULTI_LOSS_WEIGHT: [1.0, 1.0]

# 其他配置
TRAIN_CROP_SIZE: (512, 512)
EVAL_CROP_SIZE: (512, 512)
AUG:
    AUG_METHOD: "unpadding"
    FIX_RESIZE_SIZE: (512, 512)
    RICH_CROP:
        ENABLE: True
        BLUR: True
        BLUR_RATIO: 0.2
        MAX_ROTATION: 45
        MIN_AREA_RATIO: 0.5
    MIRROR: True
    FLIP: True
BATCH_SIZE: 8
TRAIN:
    PRETRAINED_MODEL_DIR: "./PaddleSeg/pretrained_model/ocrnet_w18_bn_cityscapes/"
    MODEL_SAVE_DIR: "./PaddleSeg/saved_model/ocrnet_optic/"
    SNAPSHOT_EPOCH: 40
TEST:
    TEST_MODEL: "./PaddleSeg/saved_model/ocrnet_optic/final"
SOLVER:
    NUM_EPOCHS: 200
    LR: 0.001
    LR_POLICY: "poly"
    OPTIMIZER: "adam"
    CROSS_ENTROPY_WEIGHT: 'dynamic'