Finetune AlexNet with Tensorflow

Update 15.06.2016

I revised the entire code base to work with the new input pipeline coming with TensorFlow >= version 1.2rc0. You can find an explanation of the new input pipeline in a new blog post You can use this code as before for finetuning AlexNet on your own dataset, only the dependency of OpenCV isn't necessary anymore. The old code can be found in this past commit.

This repository contains all the code needed to finetune AlexNet on any arbitrary dataset. Beside the comments in the code itself, I also wrote an article which you can find here with further explanation.

All you need are the pretrained weights, which you can find here or convert yourself from the caffe library using caffe-to-tensorflow. If you convert them on your own, take a look on the structure of the .npy weights file (dict of dicts or dict of lists).

Note: I won't write to much of an explanation here, as I already wrote a long article about the entire code on my blog.

Requirements

• Python 3
• TensorFlow >= 1.2rc0
• Numpy

TensorBoard support

The code has TensorFlows summaries implemented so that you can follow the training progress in TensorBoard. (--logdir in the config section of finetune.py)

Content

• alexnet.py: Class with the graph definition of the AlexNet.
• finetune.py: Script to run the finetuning process.
• datagenerator.py: Contains a wrapper class for the new input pipeline.
• caffe_classes.py: List of the 1000 class names of ImageNet (copied from here).
• validate_alexnet_on_imagenet.ipynb: Notebook to test the correct implementation of AlexNet and the pretrained weights on some images from the ImageNet database.
• images/*: contains three example images, needed for the notebook.

Usage

All you need to touch is the finetune.py, although I strongly recommend to take a look at the entire code of this repository. In the finetune.py script you will find a section of configuration settings you have to adapt on your problem. If you do not want to touch the code any further than necessary you have to provide two .txt files to the script (train.txt and val.txt). Each of them list the complete path to your train/val images together with the class number in the following structure.

Example train.txt:
/path/to/train/image1.png 0
/path/to/train/image2.png 1
/path/to/train/image3.png 2
/path/to/train/image4.png 0
.
.

were the first column is the path and the second the class label.

The other option is that you bring your own method of loading images and providing batches of images and labels, but then you have to adapt the code on a few lines.