Simple caffe Classifier

GUI tool to test caffe neural networks.

Description

Provides a simple GUI tool to use CNN. To classify images, you only need to drag and drop.

Requirements

• Python 2.7
• PyQt4
• pycaffe (python interface of caffe)
• numpy
• matplotlib
• cv2 (python interface of OpenCV)

Preparation

Prepare files to load CNN

• net.caffemodel - binary proto file for weights of CNN. Obtain by downloading or training.
• deploy.prototxt - define the network structure. Set input layer dim = 1x3x227x227, and set the top layer name as "prob".
• labels.txt - simple text file to list the classes to be classified.
• Create a folder with network name and put these 3 files into that folder.
• Edit classifier_simpleGUI.py line 152. Set the argument of UI() to the created folder name.

Folder alexnet is an example.

alexnet/deploy.prototxt is a modified file of BVLC alexnet (github). Corresponding caffemodel file can be downloaded from there.

alexnet/labels.txt is a list of classes. This is extracted from synset_words.txt, which is downloaded from http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz.

Usage

1. Run classifier_simpleGUI.py. Running on terminal is better because some messages will be flushed on stdout/err.
2. After load the caffe model, an empty window whose title is the same as the network name will launch. Please wait for a while.
3. Select image files you want to classify. Drag them to the window and drop.
4. After forward calculations, a table which lists top 5 classes for each image will appear. Please wait for a while.
5. The actual probability values for each class can be confirmed by pushing "Details" button on the bottom of the window.
6. During detail view, you can return to the top 5 list table.
7. Dragging and drop other images, the previous results will be discarded an then calculate for new images and prepare a new table.
8. To terminate the program, simply close the window.

Note

This program is written to learn how to use PyQt4.

Author

Seiji Ueno

ブログ(日本語)