How to train dataset with different number of classes?

[AsadRehman-319836]

Hi,

I have trained your model with parameters mentioned in your paper. Now I want to train it with more number of classes. In paper you mentioned that you train it with 113 classes. Now I want to train it with around 140 classes. Can you please tell me what parameters are required to change in your model to start training?

Thanks in advance.

[XH-B]

Change 'dictionary.txt' using your label and change 'K'=113' to 'K=140'

AsadRehman-319836 @.***> 于2022年10月16日周日 12:29写道：

Hi,

I have trained your model with parameters mentioned in your paper. Now I want to train it with more number of classes. In paper you mentioned that you train it with 113 classes. Now I want to train it with around 140 classes. Can you please tell me what parameters are required to change in your model to start training?

Thanks in advance.

— Reply to this email directly, view it on GitHub https://github.com/XH-B/ABM/issues/8, or unsubscribe https://github.com/notifications/unsubscribe-auth/AL4PQCDUBTWRTS6LIG23LFLWDOACNANCNFSM6AAAAAARGF6KCU . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[AsadRehman-319836]

@XH-B Thanks for the reply. I did that but accuracy dropped drastically. Now WER is 90% on validation data. Any suggestion on how to increase the accuracy.

Also I changed batch_Imagesize to 80000 maxImagesize to 80000 and valid_batch_Imagesize to 80000 since my GPU ran out off memory if I select these parameters to 320000.

Waiting for your reply.

[SuperHHzy]

@XH-B Thanks for the reply. I did that but accuracy dropped drastically. Now WER is 90% on validation data. Any suggestion on how to increase the accuracy.

Also I changed batch_Imagesize to 80000 maxImagesize to 80000 and valid_batch_Imagesize to 80000 since my GPU ran out off memory if I select these parameters to 320000.

Waiting for your reply.

Hello, did you solve your problem? What caused your accuracy rate to drop?

[hygxy]

Hi,

I have trained your model with parameters mentioned in your paper. Now I want to train it with more number of classes. In paper you mentioned that you train it with 113 classes. Now I want to train it with around 140 classes. Can you please tell me what parameters are required to change in your model to start training?

Thanks in advance.

Hi, the CROHME2014_data folder contains the following sub-folders:

isolatedTest2014.zip
MatricesTest.zip
MatricesTestLG.zip
MatricesTrain.zip
MatricesTrainLG.zip
TestEM2014.zip
TestEM2014GT.zip
TestEM2014LG.zip

which one do you use to train the author's model?

[AsadRehman-319836]

Hi,

I'm extremely sorry for the late reply. for your first question, the answer is I didn't change the number of classes in test.py code that's why my accuracy was not correct. In CROHME 2014, the dataset given on CROHME website does not contain a training dataset. Try CROHME 2012 dataset. Maybe that dataset contains training data. Moreover, use such datasets that have mathematical equations because CROHME datasets have Matrices, symbols, and Mathematical expressions in separate folders.

Thanks. Regards

On Mon, Dec 5, 2022 at 8:39 AM hygxy @.***> wrote:

Hi,

I have trained your model with parameters mentioned in your paper. Now I want to train it with more number of classes. In paper you mentioned that you train it with 113 classes. Now I want to train it with around 140 classes. Can you please tell me what parameters are required to change in your model to start training?

Thanks in advance.

Hi, the CROHME2014_data folder contains the following sub-folders:

isolatedTest2014.zip MatricesTest.zip MatricesTestLG.zip MatricesTrain.zip MatricesTrainLG.zip TestEM2014.zip TestEM2014GT.zip TestEM2014LG.zip

which one do you use to train the author's model?

— Reply to this email directly, view it on GitHub https://github.com/XH-B/ABM/issues/8#issuecomment-1336702081, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANXKAMIRB2ST7O6MK2FB5ATWLVPXBANCNFSM6AAAAAARGF6KCU . You are receiving this because you authored the thread.Message ID: @.***>

-- Regards, Engr. Asad Rehman NUST-PNEC Electrical Engineer Contact No: +923135131880