[WIP] Transfer Learning using ResNet and tinyImageNet

Open shashikdm opened this issue 6 years ago • 4 comments

Todo

[Data Augmentation (cropping, flipping, light and color variation)]
[Resample data to 224*224 using bicubic interpolation]
[Bicubic interpolation of first Conv layer]

Feb 20 '19 02:02 shashikdm

@shashikdm The ResNet model in the Metalhead master is broken. That is the primary cause for the poor results. Also the exact preprocessing steps are given here.

The updated model is in the newmodels branch if you want to test using that. Your current code will work, only change the Resnet() line to trained(ResNet50).

Feb 20 '19 07:02 avik-pal

Using the manifest you can depend on a branch of Metalhead, if you want to try that.

Feb 26 '19 13:02 MikeInnes

https://github.com/avik-pal/Metalhead.jl/tree/newmodels is the branch. :)

Feb 26 '19 18:02 avik-pal

yes I'm currently training the last layer using ResNet50 from newmodels branch

Mar 01 '19 13:03 shashikdm