In this project, we use flickr style dataset to demonstrate finetune in TensorFlow. The details please refer to the example from the Caffe website
Thank @ethereon and @sergeyk for their code. We modify the network.py from caffe-tensorflow and flickr.py from vislab for our use.
# Download dataset
$ python assemble_data.py images train.txt test.txt 500
Download link: here
Or follow the tutorial and extract bvlc_alexnet.npy from https://github.com/ethereon/caffe-tensorflow
$ python finetune.py train.txt test.txt bvlc_alexnet.npy
// Fine-tuning result
Iter 1280: Testing Accuracy = 0.3250
// From scratch result
Iter 1280: Testing Accuracy = 0.1655