delveintodetail
delveintodetail
> > In the competition, we used our own internal tool (FastOCR) to implement our algorithm. In the mmocr framework, we did not implement it. If you fully understand the...
> Hi，@JiaquanYe @delveintodetail，我测试了一下这个[大模型](https://github.com/JiaquanYe/TableMASTER-mmocr/blob/master/configs/textrecog/master/table_master_ConcatLayer_ResnetExtract_Ranger_0721.py)的推理速度： > 在max_len=500，遇到EOS不提前终止的情况下： > > 1. 没有去除Code line 42-43的冗余计算时，大概是11s/img； > 2. 去除冗余计算后，大概是6s/img； > 3. 再加上memory-cached inference的话，大概是4.5s/img。 > > 请问一下在相同config模型下，你们内部实现的memory-cached inference的推理速度最快能到多少？我profile了一下目前的memory-cached inference的实现，主要瓶颈都在KQV的matrix计算上了，应该不是pytorch具体实现的问题。目前都是采用的原生Pytorch进行的计算，没有经过torchscript或者其他工程化操作。 In the decoder, the computation of K...
> > > > In the competition, we used our own internal tool (FastOCR) to implement our algorithm. In the mmocr framework, we did not implement it. If you fully...
There are several papers that adopt transformer on OCR. 1. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition 2. MASTER: Multi-Aspect Non-local Network for Scene Text Recognition 3. A...
> Hi! > > Thanks so much for strengthening the open-source community with this project! > > Will you also share the config and the checkpoint file? Seems like it...
> Hi! Yes, this computing power is beyond my budget. Hahaha Are you not sure about the release date or not sure if you are going to release it? Anyway,...
> @delveintodetail > Hi, thanks for the great repo! > Will you continue to have a plan to release a pre-trained model? Thanks in advance. Sure and definitely, @JiaquanYe will...
@victor-ab @huyhoang17 We have released the table structure reconstruction pre-trained models. Please check it. We will release text line recognition model soon.
hello, 实际上是可以的， 就是预测出来了框， 如果框的准确性非常高（实际上对于pubtab这个数据， 是可以很准的），是可以不需要用我们那个复制的后处理的， 目前那个post-processing是我们为了比赛刷点搞出来的， 比较复杂， 我们最初的想法也是直接抠出预测的区域做ocr， 后面觉得目前这个处理算法可以提升指标所以采用了， 尽管这个策略有点复杂。。。
Tablemaster is a method designed for recognizing table scientific literature. Yes it can be adopted to this task, but you should collect enough training data.