4 ms·

Very interesting. For my curiosity I did some searching on VGG vs ResNet. * ResNets were better at feature extraction for image clustering [0] * One of the tr

by icc97 8y ago

Very interesting. For my curiosity I did some searching on VGG vs ResNet. * ResNets were better at feature extraction for image clustering [0] * One of the trade-offs of ResNets seems to be their relative complexity to VGG [1] * It surprises me that ResNets aren't significantly faster to train based on the large reduction of FLOPs (from the ResNet paper VGG-19 had 19.6 billion FLOPs vs ResNet-34 with 3.6 billion FLOPs) - I think people just train deeper ResNets e.g. ResNet-50 [0]: https://medium.com/@franky07724_57962/using-keras-pre-trained-models-for-feature-extraction-in-image-clustering-a142c6cdf5b1 https://medium.com/@franky07724_57962/using-keras-pre-traine... [1]: https://www.reddit.com/r/MachineLearning/comments/6e6mlf/d_is_vgg_common_in_newer_research_or_is_resnet/ https://www.reddit.com/r/MachineLearning/comments/6e6mlf/d_i...

hardmaru 8y ago
You might enjoy this recent article on Distill: https://distill.pub/2018/differentiable-parameterizations/#section-styletransfer https://distill.pub/2018/differentiable-parameterizations/#s... They discuss VGG vs non-VGG architectures in the context of style transfer in Section 2, which was interesting to me.
- icc97 8y ago
  Excellent, digging down into that leads to this Reddit post [0] > One thing they noticed was that using features from a pretrained ImageNet VGG-16/19 CNN from 2014 (4 years ago), like the original Gatys paper did, worked much better than anything else; indeed, almost any set of 4-5 layers in VGG would provide great features for the style transfer optimization to target (as long as they were spread out and weren't exclusively bottom or top layers), while using more modern resnets (resnet-50) or GoogLeNet Inception v1 didn't work - it was hard to find sets of layers that would work at all and when they did, the quality of the style transfer was not as good. Interestingly, this appeared to be true of VGG CNNs trained on the MIT Places scene recognition database too, suggesting there's something architectural going on which is not database specific or peculiar to those two trained models. And their attempt at an upscaling CNN modeled on Johnson et al 2016's VGG-16 for CIFAR-100 worked well too. [0]: https://www.reddit.com/r/MachineLearning/comments/7rrrk3/d_eat_your_vggtables_or_why_does_neural_style/ https://www.reddit.com/r/MachineLearning/comments/7rrrk3/d_e...
- icc97 8y ago
  I now realise that that article comes back to your very own twitter post [0], thanks for your patience! [0]: https://twitter.com/hardmaru/status/954173051330904065 https://twitter.com/hardmaru/status/954173051330904065

by icc97 8y ago

hardmaru 8y ago

You might enjoy this recent article on Distill: https://distill.pub/2018/differentiable-parameterizations/#section-styletransfer https://distill.pub/2018/differentiable-parameterizations/#s... They discuss VGG vs non-VGG architectures in the context of style transfer in Section 2, which was interesting to me.

icc97 8y ago
Excellent, digging down into that leads to this Reddit post [0] > One thing they noticed was that using features from a pretrained ImageNet VGG-16/19 CNN from 2014 (4 years ago), like the original Gatys paper did, worked much better than anything else; indeed, almost any set of 4-5 layers in VGG would provide great features for the style transfer optimization to target (as long as they were spread out and weren't exclusively bottom or top layers), while using more modern resnets (resnet-50) or GoogLeNet Inception v1 didn't work - it was hard to find sets of layers that would work at all and when they did, the quality of the style transfer was not as good. Interestingly, this appeared to be true of VGG CNNs trained on the MIT Places scene recognition database too, suggesting there's something architectural going on which is not database specific or peculiar to those two trained models. And their attempt at an upscaling CNN modeled on Johnson et al 2016's VGG-16 for CIFAR-100 worked well too. [0]: https://www.reddit.com/r/MachineLearning/comments/7rrrk3/d_eat_your_vggtables_or_why_does_neural_style/ https://www.reddit.com/r/MachineLearning/comments/7rrrk3/d_e...
icc97 8y ago
I now realise that that article comes back to your very own twitter post [0], thanks for your patience! [0]: https://twitter.com/hardmaru/status/954173051330904065 https://twitter.com/hardmaru/status/954173051330904065