search

yusuketomoto / chainer-fast-neuralstyle

Chainer implementation of "Perceptual Losses for Real-Time Style Transfer and Super-Resolution".
MIT License
803 stars 229 forks source link

is there a tiny train dataset? #77

Open albertyou2 opened 7 years ago

albertyou2 commented 7 years ago

HI @yusuketomoto Nice work. The generate.py is fast and the result is good!

But the training speed is too slow! I have download the train2014 dataset and followed the instructions to train my own model. But I found that training on this dataset for 1 epoch needs nearly 24 hours with GFORCE GTX 1070(8G memory).

Then I set the batch_size to 2 , I got an out of memory error. Is there an small alternative dataset?Or is there a method for speed up?

yusuketomoto commented 7 years ago

@albertyou2 Please try resize-conv branch which converges much more first and quality might be better. Please also note that this branch is not compatible with master branch so to be careful when you use models trained by the other branch. I hope this answers your question.

albertyou2 commented 7 years ago

@yusuketomoto Thank you for reply ! I 'll try as you said.

albertyou2 commented 7 years ago

@yusuketomoto Thank you for reply ! I 'll try as you said.

keefeWu commented 4 years ago

Also you don't need use the whole data, you can split data, just choose what you want