Training with lower and upper case alphabets, numbers

[MounikaMaddula]

Hi

I am trying to train the model with lower & upper case alphabets plus numbers. Model seems to converge well with lower case alphabets and numbers but when I added upper case alphabets, model is not converging.

To give more context, I used opencv to generate dataset. Length of the words varied from 1 character to 15 characters. Learning rate - 0.001. Shuffled the dataset and I observed that without shuffling the dataset based on word length helped in quicker convergence compared to sorted dataset.

[meijieru]

What the difference between no shuffling and sorted dataset?

[MounikaMaddula]

Sry shuffled dataset gave quicker convergence compared to unshuffled dataset.

[meijieru]

First, you may have to acquiring more data as the problem is more difficult.

For the second question, it is natural as the assumption of IID data. Following paradigm may help

1. train with shuffled dataset, all images are resized into same size
2. then train without shuffling where images' ratio are kept

[MounikaMaddula]

Before loading to the model, I resize all the images to 100*32. With same parameters, model is working well with only lower case and digits. You think complexity of the model has to improved to make it work for upper case alphabets ?

Will try with 2nd point.

[MounikaMaddula]

I increased the training data significantly but still model is performing poor on separate classes for small and capital letters. I also tried changing the model parameters such as no. of hidden units of RNN, adding a CNN layer. But these didn't help. Can you suggest me what can be done to improve the accuracy?

[ooooverflow]

@MounikaMaddula I am trying to train with alphabets, numbers too,but the loss seems wrong. would you mind posting your code, thank you

[haneSier]

@ooooverflow you can refer to https://github.com/Sierkinhane/crnn_chinese_characters_rec