Home

Awesome

multilabel-MXNet

This is the implement of the multilabel image classificaton in MXNet. Multilabel means each image belong to 1 or more than 1 labels and it is different from multi task.

This implement doesn't need recompile MXNet and is very convenient for you to use. Firstly, I assume that you can use MXNet normally. Then, do as follows:

Data

If you are doing a single label image classification, your .lst file may like this(take 4 classes as example):

IDlabelimage_name
52470.000000image1.jpg
339861.000000image2.jpg
398292.000000image3.jpg
156473.000000image4.jpg
103691.000000image5.jpg
224083.000000image6.jpg
25982.000000image7.jpg

For multilabel image classification, you should create .lst file like multilabel-MXNet/data_example/train_data.lst (take 4 classes as example)

In this implement, we only use .lst and raw image as the input instead of .rec file.

Train

run_train.sh is the train script for you to start fine-tune quickly. You should open this script and change the path of train_multilabel.py, .lst file, imagefile and model-prefix after you clone the project.

Then run:

sh run_train.sh

More details