Dividing images to train,test and val

[SaadAliHafiz]

greetings Lukas,
!gdown -O dataset/data/crohme.zip --id 13vjxGYrFCuYnwgDIUqkxsNGKk__D_sOM !gdown -O dataset/data/pdf.zip --id 176PKaCUDWmTJdQwc-OfkO0y8t4gLsIvQ !gdown -O dataset/data/pdfmath.txt --id 1QUjX6PFWPa-HBWdcY-7bA5TRVUnbyS1D os.chdir('dataset/data') !unzip -q crohme.zip !unzip -q pdf.zip os.chdir('images') !mkdir ../valimages !ls | shuf -n 1000 | xargs -i mv {} ../valimages os.chdir('../../..')
by running this command, test, train and val directories are made, but its not of handwritten images.

[lukas-blecher]

Because the handwritten images are located in images/ and valimages/, but you're looking at trian/ which are the regular images.
I clearly stated in the notebook that the model will be trained on the mixed domain.

[SaadAliHafiz]

Got it thanks,
Lukas, I tried training only on handwritten images only I didn't include pdf equations but It's still giving me correct predictions on pdf equations while not recognizing handwritten equations why?
is it coz of we downloaded pretrained model before training ? by this command
!curl -L -o checkpoints/weights.pth https://github.com/lukas-blecher/LaTeX-OCR/releases/download/v0.0.1/weights.pth

[SaadAliHafiz]

I did, in my case, it was mixed_e14.pth so I use this

!python pix2tex.py --checkpoint checkpoints/mixed/mixed_e14.pth

[lukas-blecher]

ok perfect. Well did it work?

[SaadAliHafiz]

no :(

[lukas-blecher]

but BLEU is still good? What if you compute it for the training set as well? These should be exactly the same images.

[SaadAliHafiz]

Yep BLEU is still good, Okay I'll compute it on train.pkl

[SaadAliHafiz]

both are same

[lukas-blecher]

That's very good. Don't know why it won't work for pix2tex though. Try adding the config file:

!python pix2tex.py --checkpoint checkpoints/mixed/mixed_e14.pth --config settings/colab.yaml

[SaadAliHafiz]

it finally worked thanks, Lukas!
You're a life saver

[lukas-blecher]

Nice :D Make sure to check the valimages too

[SaadAliHafiz]

I did and it worked on that as well

[SaadAliHafiz]

I've even tried to check on other handwritten images from google and it worked on those as well.