Freeze Localization

[yash-bhat]

Hello @Bartzi

Great work!!

I am training a text recognition model based off your 'text_recogniton_model' as such:

python3 chainer/train_text_recognition.py curriculum.json log -b 16 --char-map map.json -r text_recog_model/model_190000.npz -g 0 -si 100

Can you please give a brief as to what --load-localization, --load-recognition and --freeze-localization actually does.

My goal is to use your model as is for text localization but fine tune text (new set of characters) recognition for my cause.

Thank you!

• Yashu

[Bartzi]

Hi,

thanks for using our work!

Here is the explanation of the command line options:

• --load-localization with this option you can indicate that you only want to load the weights of the pre-trained localization network that is contained in the snapshot, you are referring to with the -r option
• --load-recognition is the same, but here you will only load the weights of the recognition network
• --freeze-localization indicates that the localization network shall not be updated during training, i.e. the weights remain constant through the training.

For your case, you should use --load-localization and --freeze-localization.

[yash-bhat]

Thanks for the explanation. I was able to train through till 400 iter but now it throws me the issue #18

Any ideas? I'm feeding single word images and the maximum length of the words are 10 which I have specified as below in the input data csv.

10 1

Error :

chainer.utils.type_check.InvalidType: 
Invalid operation is performed in: SoftmaxCrossEntropy (Forward)

Expect: in_types[0].shape[0] == in_types[1].shape[0]
Actual: 1600 != 2080

• Yashu

[Bartzi]

Do you have a full traceback?

[yash-bhat]

Hello @Bartzi

I figured the issue. The predicted and supplied labels were of different lengths. The problem is with python parsing I think. Whenever it encounters double quotes (") it wraps around it. For example: 10 foot 2 inches represented as string - 10' 2" when read from CSV in python it becomes --- "10' 2""" !!!

I have eliminated the double quotes altogether but ideally I need the model to learn the - " Thanks though. Will update here if I figure it out.

• Yashu

[yash-bhat]

Just an update before I close this issue. For python string parsing issue with quotes - " ... i just went ahead with a placeholder character which I clean up during inference.

• Yashu