Test accuracy not crossing 0.50

[vinayakarannil]

HI, I have tried training for my own data. Intially i trained with 150k data. Then 300k and finally i trained with half a million data. But in all cases my test accuracy never crossed 0.50. The model converges in around 5-6 epochs itself with accuracy values between 0.40 and 0.50. I have tried using learning rates 0.01, 0.001,0.0001. No improvements. However all the validation results are showing correct labels. [13/25][6000/8146] Loss: 0.232784 Start val P---------e------r---**--- => Per* , gt: Per* -1------.--3----- -x------ => 1.3 x , gt: 1.3 x r---e----p----e----a---t-- => repeat , gt: repeat --------0------.--2------- => 0.2 , gt: 0.2 S--C-H--U--Y--L-KK-I-L-L-- => SCHUYLKILL , gt: SCHUYLKILL
w--------h-----e----n----- => when , gt: when RR------------------------ => R , gt: R A------c---t--i--v---ee--- => Active , gt: Active A------c----t-ii-v---e---- => Active , gt: Active R-----AA---MM------Z---I-- => RAMZI , gt: RAMZI Test loss: 1.171150, accuray: 0.445625

[avhirupc]

How long did it take to train? What dataset did you use?

[vinayakarannil]

it doesnt take much time for convergence... I used custom dataset

[Fighting-JJ]

Have you solved it? I trained on dataset synth90k ， and test accuracy never crossed 0.86. I dont know is there any promotion can be done.

[Aurora11111]

I think it beacause of your trianing image size and it must match with the training model

[zhengshangdong]

@Fighting-JJ Hi,bro.I trained on dataset synth90k too.But the loss is always 12.Even I divide the trainsets into 7,000,000 or 800,000 or 5000 pictures,the loss was not change.And the accuray is 0,It could not recognize any number pictures.Did you meet this problem.I use pytorch 1.0 and python 3.6..

[hiendoan95]

I have the same problem when training the model with my own Japanese dataset. The max length is 50 characters (based on data requirement). The accuracy never crossed 0.4. Can you give me some advice?