visual grounding finetune咨询

hello
我有在finetune 模型时，发现模型存在_IncompatibleKeys(主要是fusion_encoder的所有层)， 这个是对训练否有影响。 我按照操作流程进行训练，在epoch=10自动停止，并第十轮结果如下：
{'train_lr1': 1.9676295349747303e-05, 'train_lr2': 4.931851652578033e-06, 'train_loss_seq': 0.08181070101696078, 'miou': 0.6400532953981678, 'accu': 0.7921358685619346, 'epoch': 10}

这个结果与论文的accu值存在差距，两个数值是否一致啊？

Start by reproducing the documented visual grounding finetuning workflow and inspect the reported IncompatibleKeys for the fusion_encoder layers. Compare the epoch 10 accu and miou values with the paper's metric definitions and determine whether the missing keys affect training. Done means clarifying both the checkpoint warning and whether the reported accu values are comparable.

Written by the indexing model from the issue text.