ssygeneral

ReCTS模型确实可以同时识别英文和中文，但是我对它进行了微调，使用totaltext在rects的模型权重上进行了再训练，得到的模型却识别出来的却是各种中文乱码，这是为什么？下面是配置文件。 _BASE_: "Base-SWINTS_swin.yaml" MODEL: WEIGHTS: "/swin-transformer/SwinTextSpotter-main/output/rects_model_final.pth" SWINTS: NUM_PROPOSALS: 300 NUM_CLASSES: 2 REC_HEAD: POOLER_RESOLUTION: (16,40) RESOLUTION: (32, 80) BATCH_SIZE: 128 NUM_CLASSES: 5463 DATASETS: TRAIN: ("rects",) TEST: ("totaltext_test",) SOLVER: STEPS: (140000,160000) MAX_ITER:...