chainer-fast-neuralstyle
chainer-fast-neuralstyle copied to clipboard
yusuketomoto
Any plans to update with Instance normalization?
Hi, do you have any plans to update your solution with Instance normalization as described here "Instance Normalization: The Missing Ingredient for Fast Stylization" https://arxiv.org/abs/1607.08022
@igorbarinov what's the difference between instance nomalization and batchnorm with batchsize = 1 ?
@igorbarinov got it. instance normalization only normalized in spatial not cross feature map and batch
@austingg Have you tried it? I think it does the same thing with batch normalization when you set batchsize = 1
@bucktoothsir I haven't tried. but batchsize = 1 still normalize cross the feature maps. instance normalization only do normalization within the feature map plane. As @yusuketomoto say, batchsize >=2 have some problem, i do all my experiments with batchsize = 1
@bucktoothsir sorry . After i checked the paper, i found batchsize = 1 is instance normalization.
Are you sure ? According to the paper that would mean that using batchsize=1 yields better results than batchsize >= 2
@ttoinou batch normalization is across instances, however, when batch size = 1, there is only one instance. You can also implement multi batch training with normaliztion within instance.
From the paper we can see the effect is greatly improved. Anybody have the implement in python, I can test it.
@logic1988 : Yes they updated their project yesterday with instance normalization with torch ! https://github.com/DmitryUlyanov/texture_nets @austingg : Great so there's still some hope to get the training faster by fixing the batchsize issue (I guess that loading images, resizing them and giving them to the GPU would be faster in batch, no ? there would be less CPU <-> GPU communication)
There are a great Torch Implementation released by jcjohnson a few days ago. https://github.com/jcjohnson/fast-neural-style
In addition to the original method, the new method inspired by the the paper [Instance Normalization: The Missing Ingredient for Fast Stylization] has been implemented.
The latter seems to have a better result and faster speed than the former.
It said that:
These models use the same architecture as those used in our paper, except with half the number of filters per layer and with instance normalization instead of batch normalization. Using narrower layers makes the models smaller and faster without sacrificing model quality.
How to modify the corresponding code in this project? @yusuketomoto
On the same system, is the Torch implementation or this implementation faster? better? Anyone have comparisons?
