finetune SER model 任務中的 max_seq_len

[bhhsieh]

請問 finetune SER model 任務中的 max_seq_len 預設是 512 https://github.com/PaddlePaddle/PaddleOCR/blob/main/configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml

我們數據集的 token 長度會超過 512，想調整這個值為 1024 但調整後遇到以下 error 想請教如何解決，謝謝

[cuicheng01]

代码里涉及到expand算子的地方需要跟着改动下

[bhhsieh]

請問具體要修改什麼地方? 我最後找到這個function，但沒找到 error 顯示 value (514) 之處， 再麻煩說明一下，感謝！

我的版本為 paddlepaddle-gpu==2.3.1 paddlenlp==2.5.2

[tran601]

我是分批送进去

[cuicheng01]

paddle的版本建议升级下呢？

[ariefwijaya]

@bhhsieh Did you solved the problem? @tran601 How about training model, should we modify the training code to support batches as well?

[ariefwijaya]

@tran601 I've tried you code, but that produce an error:

ValueError: (InvalidArgument) The 1th element of 'shape' for expand_v2 op must be greater than 0, but the value given is 0.
  [Hint: Expected expand_shape[i] > 0, but received expand_shape[i]:0 <= 0:0.] (at /Users/paddle/xly/workspace/3a60970a-8ff4-461a-878a-a5fbdbe8e3e9/Paddle/paddle/phi/infermeta/unary.cc:1215)

[ariefwijaya]

to concat batch : preds = paddle.concat([p1[0],p2[0]],axis=1)