Skip to content

Latest commit

 

History

History
36 lines (23 loc) · 1.05 KB

File metadata and controls

36 lines (23 loc) · 1.05 KB

tensorflow-finetune-flickr-style

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 flickr style dataset

# Download dataset
$ python assemble_data.py images train.txt test.txt 500

Download the pre-trained model

Download link: here

Or follow the tutorial and extract bvlc_alexnet.npy from https://github.com/ethereon/caffe-tensorflow

Lauch the finetune process

$ python finetune.py train.txt test.txt bvlc_alexnet.npy

Finetune result

// Fine-tuning result
Iter 1280: Testing Accuracy = 0.3250
// From scratch result
Iter 1280: Testing Accuracy = 0.1655