LSTM inputs.

[affromero]

Could you please share the features for the LSTM inputs? I am trying to reproduce them but I cannot reach 10% of improvement only by using LSTM as you report in the paper. For instance, I cannot improve my results using LSTM at all.

I am asking this because I am having a bad time trying to install Theano under Cuda 9.1 and all your coding dependencies.

[machanic]

@affromero I am also working on this area. I think the paper he report is not correct. I am using end-to-end LSTM which cannot achieve 66% F1-score, that is too high! You can contact me in sharpstill@163.com

[machanic]

@affromero Which file is the train file and are you runing?

[affromero]

@sharpstill these lines set the train/test files: https://github.com/wiibrew/EAC-Net/blob/master/BP4D_EAC_dlib_train.py#L400-L401. Here you can find those files: https://www.dropbox.com/sh/nshrqeuh4gu95mn/AACgyL67MGUwWMQAebcigXbla?dl=0

[YXWu-bjtu]

@sharpstill I am trying to reproduce the codes 'ENet' and 'EAC'. However, my results about f1-score are lower than the results recorded in his paper. Moreover, compared to ENet, the performance of EAC-Net is not significantly improved. I am wondering if your results matched with his results?
Please contact me!
393148129@qq.com

[YXWu-bjtu]

@affromero I am trying to reproduce the codes 'ENet' and 'EAC'. However, my results about f1-score are lower than the results recorded in his paper. Moreover, compared to ENet, the performance of EAC-Net is not significantly improved. I am wondering if your results matched with his results?
Please contact me!
393148129@qq.com

[affromero]

@YXWu-bjtu Not at all.

I exactly ran the scripts here distributed:
I trained on one of the three folds (BP4D_faceim_tr_ag_shp.txt), first on ENet, and then using ENets weights to train EAC-Net. After 20000 iterations I can only get an F1 score of 31% (using 'BP4D_faceim_ts_shp.txt' for testing), which highly differs from the 56% reported in the paper.

I have tried to contact the author (for a long time) inquiring for some clarification about this. Unfortunately, no success so far.

[machanic]

@affromero @YXWu-bjtu I think the F1 score reported in that paper has problem. My new research will be coming soon, it achieved F1 score of 63.1% on BP4D 3-fold test.
You can watch my new project(does not release yet, but will be soon)

[machanic]

@affromero @YXWu-bjtu Although the author present a great paper, but unfortunately, the result of LSTM is incorrect of PAPER, it will not surpass 63% in BP4D of F1 score.
I have open source all my code in https://github.com/sharpstill/AU_R-CNN

My paper is published in Neurocomputing: https://www.sciencedirect.com/science/article/pii/S0925231219305338

google drive download link: pdf download

I believe I have completed all the experiments , LSTM cannot help too much.