Fine tuning a Convolutional Recurrent Neural Network (CRNN) with CTC layer

1

Hello,

l have downloaded a pre-trained CRNN (convolutional recurrent neural network) for image based sequence recognition. Here is the pre-trained model and the related code for the model code

Now, l want to fine tune the model. So, l’m wondering how can l fine tune this model. The idea is to remove the last classify layer and and add a new one. How can l do that for this CRNN ?

The pre-trained CRNN has alphabet="0123456789abcdefghijklmnopqrstuvwxyz"
and l want to fine tune the model on the following alphabet
alphabet=“0123456789abcdefghijklmnopqrstuvwxyz,;:!%@)(/.°”

Thank you.

1 Like
2

search on this forum for “finetuning”, there are a few threads that explore that.
I dont think giving you particular solutions for your specific code is feasible (unless someone here has a lot of free time :slight_smile: )

3

Looks like you already have an issue on the implementor’s GitHub page: https://github.com/meijieru/crnn.pytorch/issues/23

I responded there with a solution to your problem. (Or at least, a solution to one of your problems, can’t guarantee the rest of your code will run).

Probably best to keep any further discussion to implementor’s GitHub since this is pretty specific to that code base.

Best,
Stephen