search

aimagelab / VATr

MIT License
72 stars 4 forks source link

About net.W Issue. #10

Closed Fyzjym closed 1 year ago

Fyzjym commented 1 year ago

hi，thanks your great wok! I noticed code： model.py @397 lines ''' self.netW = WDiscriminator( resolution=self.args.resolution, n_classes=self.args.vocab_size, output_dim=self.args.num_writers ).to(self.args.device) ''' Why n_classes is self.args.vocab_size, does net.W classify the author wid? net.W is still responsible for classifying characters in the vocabulary table?

vittoriopippi commented 1 year ago

Hi @Fyzjym! The WDiscriminator class is responsible to the Writer Identification. The fact that class WDiscriminator takes as input the n_classes is a refuse of old code from the Handwriting-Transformers GitHub repo.

The n_classes parameter is stored into the class variables (Line 646) and it is used to create the variable embed (Line 710) that is never used into the class.

Fyzjym commented 1 year ago

I completely agree with your point of view. Thank you for your prompt and effective responses! :D