search

Dexterp37 / fontina

Data generation, model training and inference for Visual Font Recognition using PyTorch
MIT License
10 stars 1 forks source link

Cu subnetwork #10

Closed alliegu closed 1 month ago

alliegu commented 6 months ago

Hi! Testing this implementation on real images and trying to figure out if I can improve performance. In the paper they mention that Layers 1 and 2 are imported to the CNN in Fig. 5, as the Cu sub-network and fixed. But in the implementation you seem to call the full autoencoder upon training the full network. I wonder if that was intentional? Thanks! Edit: nevermind, it's just the encoder

alliegu commented 6 months ago

Oh nevermind, it's just the encoder

Dexterp37 commented 6 months ago

Oh nevermind, it's just the encoder

No worries! Please let me know if you need further explanation. If you manage to find out what the problem is, I'm willing to spend GPU time to retrain this and share the checkpoints.

alliegu commented 6 months ago

Can I ask what python / pytorch / torchvision versions you're using? I did the installation steps but ran into versioning issues; downgraded and new issues popped up. Thank you so much!

Dexterp37 commented 4 months ago

Can I ask what python / pytorch / torchvision versions you're using? I did the installation steps but ran into versioning issues; downgraded and new issues popped up. Thank you so much!

Whops, sorry I missed this reply. @alliegu I'm using Python 3.10, torch 2.0.1 and the related torchvision version. See here.

anruigu commented 1 month ago

Just wanted to comment that with your pretrained full checkpoint, even though I couldn't reproduce correct predictions using test samples in the paper, this is a sufficiently good feature extractor for my use case on unlabeled fonts. Really appreciate it!

Dexterp37 commented 1 month ago

Just wanted to comment that with your pretrained full checkpoint, even though I couldn't reproduce correct predictions using test samples in the paper, this is a sufficiently good feature extractor for my use case on unlabeled fonts. Really appreciate it!

@anruigu I'm glad you found this helpful! Keep up the great work!