search

sergeywong / cp-vton

Reimplemented code for "Toward Characteristic-Preserving Image-based Virtual Try-On Network"
MIT License
474 stars 182 forks source link

training on larger images #9

Closed solitarysandman closed 5 years ago

solitarysandman commented 5 years ago

Hi @sergeywong,

I'm trying to train GMM on images of size 512, 384 - double of what has been used in cp-vton training. What changes would I need to make in networks.py?

I've tried a couple of things but they haven't shown promise:

Can you please let me know how can I train on images of size (512,384)?

Thanks.

mr-johnlocke8 commented 5 years ago

Would love to know if you solved it, I am trying as well to adjust the model to x2 the size.

solitarysandman commented 5 years ago

@mr-johnlocke8

changed the number of layers in feature extraction from 3 to 4 (lines 408, 409 in networks.py)

^this worked for me.

rishabhdhakarwal commented 5 years ago

@solitarysandman is there any significant difference between the outputs if GMM is trained on larger images as you mentioned.

solitarysandman commented 5 years ago

@rishabhdhakarwal not really, no. But increasing the grid size has a beneficial change. Finer warping can be noticed. But increasing grid size comes at a computational cost, so keep that in mind!

rishabhdhakarwal commented 5 years ago

@solitarysandman I'm trying to train the model on images of size 4x i.e. 1024 X 768 What should I keep the grid size for satisfactory outputs?

Thanks

solitarysandman commented 5 years ago

@rishabhdhakarwal , You'll have to run multiple experiments, gradually increasing the grid size. Unfortunately, I cannot exactly tune that hyperparameter for you.