search

machanic / AU_R-CNN

The official implementation code of paper: "AU R-CNN：Encoding Expert Prior Knowledge into R-CNN for Action Unit Detection".
https://arxiv.org/abs/1812.05788
87 stars 24 forks source link

save and load all of the parameter #5

Open AminRahmatii opened 4 years ago

AminRahmatii commented 4 years ago

Dear Mr I ran your code(AURCNN) step by step you told in github but after training,i see loss function doesn't descend and have random behavior.i just changed batch size from 20 to 7 because of out of memory error in runtime. and accuracy was zero all the time of training . can you help me why this happening? and another thing this model is very big and the time of training is about 20 days in system with gpu 1080 ti.i sometimes stop and resume training but all of the parameters doesn't save like epoch and when i stop in epoch 6,and then run again,project run from epoch 1.can you please change your code so that this problem solved? thanks a lot

machanic commented 4 years ago

Please use SGD optimizer , the loss function problem will be fixed.