didw / imagenet-transfer.torch

Imagenet Transfer Learning on torch using caffe model

This repository is torch code to transfer learning using caffe model(used loadcaffe). You can make good performance using small amount of data if you fine tuning or transfer learning using pretrained model.

This code is based on imagenet-multiGPU.torch.

Dependencies

• loadcaffe(https://github.com/szagoruyko/loadcaffe)
• cudnn(https://github.com/soumith/cudnn.torch)

Data Preprocessing

Traning and test data needs to store in root/train and root/val respectively. First subdirectory name is going to be class name. You don't need to have label files in data.

Training Type

• Transfer Learning: Use pretrain model as feature extractor and train last fully connected layer only.
• Fine Tunning: Use pretrain model as initializer and train whole modules including convolution layer.

Running

Setting up the parameters in opt.lua. And run it.

th main.lua

Thanks to

• Soumith Chintala
• Sergey Zagoruyko