About training details about im2vec

[preddy5]

Hey @DYF-AI
i) Yes there are only a few images in the emojis dataset.
ii) Yes the training process is end-to-end. Could try training by setting https://github.com/preddy5/Im2Vec/blob/master/configs/emoji.yaml#L13 to 0
iii) Also it looks like you only have 2 curves can you recheck if https://github.com/preddy5/Im2Vec/blob/master/models/vector_vae_nlayers.py#L44 has 4 colors and 4 colors correspond to the ones in the dataset. No the results should look better. Refer to https://github.com/preddy5/Im2Vec/tree/master/logs/VectorVAEnLayers/version_110 for an example training run.

[DYF-AI]

Thank you very much for your reply, your suggestion is very effective：
ii) Yes the training process is end-to-end. Could try training by setting https://github.com/preddy5/Im2Vec/blob/master/configs/emoji.yaml#L13 to 0

This is the result of 185 epochs, which is much better than before，thank...

[DYF-AI]

Hey, Dear Author
Epoch 401: 100%|██████████████| 13/13 [00:07<00:00, 1.83it/s, loss=0.170, v_num=110, Reconstruction_Loss=0.0777, KLD=0, aux_loss=0, other losses=0]beta: 0.0
learning rate: 0.00029936846961918924

I have another question. I saw KLD=0, aux_loss=0, other losses=0 during training. Do these losses need to be turned on during training? In other words, will the result be better after turning it on?

[preddy5]

Hey @DYF-AI
KLD is the KL divergence loss like in a VAE
aux_loss is for the auxiliary network we discussed in the main paper.
other losses is the loss term to encourage minimal surface area for the curves.

using other_losses term after 500 epoch will improve the results a tiny bit, KLD and aux should not affect your results.

[DYF-AI]

Okay, thank you very much for your reply, this is a very interesting and meaningful project, thanks. I want to use path-to-path to generate fonts based on this project...

[preddy5]

@DYF-AI you are welcome. All the best with the project :)