Yige Xu

> Hi, > > first, thank u for having sharing ur cod with us > > I am trying to further pretraining a bert model on my own corpus on...

> Hey men, I encounter the same issue. Are you able to resolve it? > I keep getting this, > OP_REQUIRES failed at cwise_ops_common.cc:70 : Resource exhausted: OOM when allocating...

> Hi, > > thank u for answering > > I reduced the train_batch_size to 8 and max_seq_length to 40 > > but I still get the resource exhausted error...

Thanks for your issue. In run_classifier_single_layer.py (or other fine-tuning scripts), we did not save checkpoints We forgot to remove these unused arguments. If you want to save checkpoints, you can...

thank you for your issue for document classification, we do not split the text into sentences (except the Hierarchical methods) we do not remove punctuation masks. for the whole document,...

hello~ thank you for your interest in our repo. for this problem, do you satisfy the requirements mentioned in README.md? that is, tensorflow==1.1x. or you just used tensorflow 2.x?

_BertWordModel是一个内部调用的类，一般情况下不对外暴露 另外这是为了维护一个从训练测试数据的vocab到bert自身vocab的一个映射，vocab减小只是其中一个收益，另外的优点是这样在调用的时候可以更方便地预处理。如果只是单纯用原生的BertModel的话跟transformers并没有什么区别

用BERT进行文本匹配主要是将两句话拼接在一起，中间用sep分隔符隔开 其余形式（包括模型设置、调参过程）与文本分类类似

如果是相似度匹配则相当于将分类问题转换成一个回归问题，本质上也是近似的

我觉得可以有两种解决办法，第一种是重构数据集，答案正确的为一类，不正确的为另一类 第二种解决办法仍然是回归问题，但是把答案正确的标记相似度为1，答案不正确的标记相似度为0，以此计算相似度