search

maplefallfrost / GAT-Stroke

Graph attention network for stroke classification
16 stars 4 forks source link

Questions about the data? #1

Open Keneyr opened 5 years ago

Keneyr commented 5 years ago

What's the meaning of the folders 'data' and 'edge'? How did you get the files such as 0.label2,0.stroke_feature???? Can you upload the raw dataset ?

maplefallfrost commented 5 years ago

Thanks for attention. The 'data' directory stores the unary features extracted from strokes and ground truth(.label). The 'edge' directory stores the pairwise features and relationship between strokes(.edge). The raw dataset is called IAMonDo dataset which is organized in XML form and the preprocess process is a little bit complicated. I will try to fix this issue soon.

marcoleewow commented 5 years ago

Hi @maplefallfrost , very interested on how you preprocess the inkml data from IAMonDo to the uploaded data. It would be nice if you can upload the code or send me the paper so that I learn more from it.

maplefallfrost commented 5 years ago

Hi @maplefallfrost , very interested on how you preprocess the inkml data from IAMonDo to the uploaded data. It would be nice if you can upload the code or send me the paper so that I learn more from it.

Thank you for your interest and sorry for the late reply. My feature extraction method follows the same procedure as in my previous paper https://ieeexplore.ieee.org/document/7900138 and also similar in these two papers: https://www.sciencedirect.com/science/article/pii/S0031320313001878 http://web.tuat.ac.jp/~nakagawa/pub/2015/pdf/Truyen_et_al_Pattern_Recognition.pdf Recently I am struggling for my graduation and maybe I have no much time to reorganize my C++ code. I will try to release this part as soon as possible and the time might be at the beginning of October, soon after ICDAR2019.

sheirving commented 4 years ago

Thankyou very much for share! I had train your model by using features, which extracted by my code. The result is lower than using your features data, so could you share the code of feature extraction When you are convenient.

maplefallfrost commented 4 years ago

Thankyou very much for share! I had train your model by using features, which extracted by my code. The result is lower than using your features data, so could you share the code of feature extraction When you are convenient.

I feel very sorry that I still don't have time to release my code for feature extraction as I am still preparing for graduation now. So how is your result? Is there a serious drop with accuracy? I think the accuracy is accepted if it is > 98.4

sheirving commented 4 years ago

Thankyou very much for your reply! On the IAM database, I got 95.498%. And for Diagram-FA, the accuracy is 98.426%, for Diagram-FC, the accyracy is 97.431%. I trained the three model respectively！I have tried different data preprocessing such as resampling and normalization, which have a little influence on the performance(about 0.8% one IAM). So Maybe we should check the feature extraction section?

maplefallfrost commented 4 years ago

Thankyou very much for your reply! On the IAM database, I got 95.498%. And for Diagram-FA, the accuracy is 98.426%, for Diagram-FC, the accuracy is 97.431%. I trained the three models respectively！I have tried different data preprocessing such as resampling and normalization, which have a little influence on the performance(about 0.8% one IAM). So maybe we should check the feature extraction section?

Is it the text/non-text classification accuracy mentioned here? Or the general multi-class accuracy?

sheirving commented 4 years ago

Thankyou very much for your reply! On the IAM database, I got 95.498%. And for Diagram-FA, the accuracy is 98.426%, for Diagram-FC, the accuracy is 97.431%. I trained the three models respectively！I have tried different data preprocessing such as resampling and normalization, which have a little influence on the performance(about 0.8% one IAM). So maybe we should check the feature extraction section?

Is it the text/non-text classification accuracy mentioned here? Or the general multi-class accuracy?

yes, text/non-text classification !

maplefallfrost commented 4 years ago

Thankyou very much for your reply! On the IAM database, I got 95.498%. And for Diagram-FA, the accuracy is 98.426%, for Diagram-FC, the accuracy is 97.431%. I trained the three models respectively！I have tried different data preprocessing such as resampling and normalization, which have a little influence on the performance(about 0.8% one IAM). So maybe we should check the feature extraction section?

Is it the text/non-text classification accuracy mentioned here? Or the general multi-class accuracy?

yes, text/non-text classification !

Ok, that's a little bit weird as I think it is not hard for any structured prediction model to get over 97% accuracy on the IAMOnDo dataset. I need to finish my thesis this month so maybe I can upload my code in the next month. Sorry for the inconvience= =

Marcovaldong commented 4 years ago

Hi @maplefallfrost , can you share the code of feature extraction When you are convenient ?

Marcovaldong commented 4 years ago

Thankyou very much for share! I had train your model by using features, which extracted by my code. The result is lower than using your features data, so could you share the code of feature extraction When you are convenient.

Hi, @sheirving , can you share your code of feature extraction?

HugoRomat commented 1 year ago

Any update on this? @sheirving @maplefallfrost Can you share the feature extraction module?