shimin
shimin
me to ，how to solve this
模型用from PIL import ImageFont来读取字体的，然后转换embedding形式的张量， embeddings = position_embeddings + glyph_embeddings 直接和位置embed相加 再和下游的bert一起训练，训练保存词向量，完成训练后，应该可以直接就使用glece-bert层之后的进行预测推理 > 理解错了：（这篇论文不是一个预训练任务？没有利用Glyce生成一个embedding词表么？然后再下游任务的时候，不是查这个词表的来获取权重值的？）
@czhang99 Could you please share the test data?
plus 1
Can you please help me with the script how you split in train test and validation.
you need to install opencv 3.0.0
> Yes, we are training just on the 7000 examples of train_spider.json to get to SOTA. I was meaning to get to adding spider_others.json examples and I was really busy...
Hi ,I have tried which just simply modify the default.jsonet from ` "train_data_path": dataset_path + "train_spider.json",` to `"train_data_path": [dataset_path + "train_spider.json", dataset_path + "train_others.json"],` and the spider.py file: ``` def...
> Great. let me know how it does on the dev set. > There's a bit of code duplication here, but we can solve this later. @OhadRubin Hi， I just...