search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 653 forks source link

Training Time about ABCNet #82

Closed HumanZhong closed 4 years ago

HumanZhong commented 4 years ago

I'm using 1x 1080ti to train the ABCNet, but it seems quite time-consuming. For example when training only on Total-Text, the eta is up to >2 days. What is the training time in your experiments when using different num of gpus(like 1x, 4x, 8x)? I've noticed that in your ABCNet paper, you said that the overall training time is 3 days, and I want to know is this the training time for both training SynthText150k and Total-Text? THX a lot

stan-haochen commented 4 years ago

Pretraining takes 2 days on 4x1080Ti. And finetuning on TotalText takes 2 hours.

HumanZhong commented 4 years ago

Thanks a lot for your reply. And I have one more question, I noticed that the preformance of ABCNet has increased from 45.2(paper reported) to 53.2(adet version) on CTW1500. Is that a result of just replacing CTCPredictor by AttentionPredictor ? Or is there anything else that contribute to the performance gain?

Yuliang-Liu commented 4 years ago

@HumanZhong That should be one reason. Different training schedule, hyper-parameters, frameworks may also lead to different performance.

innat commented 4 years ago

A small thing is ABCNet tested on all the following sets: ICDAR, CTW1500, and Total-Text? I've found only CTW1500 and Total-Text , pre-trained weights but not for ICDAR!! (Though I didn't go to the directory). And also the training scripts (especially network itself ABCNet is somewhat ambiguous (where does it implement)?

Yuliang-Liu commented 4 years ago

@innat Hi there,

You are right, there is not weights for ICDAR at this version.

What do you mean by training scripts are ambiguous specifically? Only with more details shall we improve the implementation.

innat commented 4 years ago

@Yuliang-Liu san I mean the inferred code scripts from here. My first was something like where is the defined class of ABCNet? Then I read it something integrated with detectron library!! Apart from this, I think it would be great if there is some more detailed information about how one can use it for their custom data set? What are the things that need to be considered (like annotation scripts and anything important basically?

Yuliang-Liu commented 4 years ago

@innat san Yeah, our method is built upon detectron2.

I think it would be great if there is some more detailed information about how one can use it for their custom data set

That is a good suggestion. We will provide more detailed instruction ASAP about how to build the custom data set for training ABCNet.

Yuliang-Liu commented 4 years ago

@innat san The example scripts have been updated. You can find the link in the abcnet homepage or here.

As current discussion is not related to the topic of this issue, I will close it. Please feel free to open a new one.