Multi-Label-Image-Classification

This project uses a pre-trained network for ImageNet, adding a new layer that will be learned for new labels, and displays a resume in TensorBoard.

DESCRIPTION:

This model take a pre-trained Inception v3 shape graph (trained for the ImageNet images) and replaces the top layer of the model with new fully-connected layer that can recognize other classes of images (multi-label classification). The former top layer (node by name - 'softmax') receives as input a 2048-dimensional vector. And that model create a new layer instead, which receives as the input this vector and has the sigmoid activation function that enables multi-label classification. This model base on 'Retrain an Image Classifier for New Categories' TensorFlow tutorial , but there are significant changes. First of all - single-label classification (softmax layer) was changed to multi-label classification (sigmoid layer), as well as all system of caching and loading images for learning was rebuild for better optimisation.

PREPARATIONS:

Before training the model, you should to prepare some file structure and some separate files. The absence of these preliminary actions will entail a program error - access-error or faulty output model. You can download dataset wich I used for this project. (There is no verification of the correctness of the input data, it is left for lightness of the source code) The LABELS_DIR/labels.csv file describing all the available labels in the certain order in which they will be described during the training. This is described in the csv file format, for example:

label_1,label_2, ... ,label_N

The training images folder (IMAGE_DIR) should have a structure like this:

./images/train/image_name_1.jpg
./images/train/image_name_2.jpg
...
./images/train/image_name_N.jpg

Also there should be a LABELS_DIR/train.csv file describing (giving the correct labels for each image) images from the IMAGE_DIR directory in the csv format. For instance, the file may look like this:

image_name_1.jpg,0,1,0,0, ... ,0,1,1
image_name_2.jpg,0,0,1,0, ... ,1,0,0
...
image_name_N.jpg,1,0,0,0, ... ,0,0,0

OUTPUT:

This script produces a new model file (tensorflow graph) that can be loaded and run by any TensorFlow-compatibility program, for example the label.py or OpenCV.

TENSORBOARD:

By default, this script will log summaries in the format for TensorBoard to SUMMARIES_DIR (./tmp/logs by default) directory. Visualize the summaries with this command:

$ tensorboard --logdir ./tmp/logs

REFERENCE:

This network was used for the image multi-labeling at the competition by Intel. The initial data set is a 36000 images described by 17 categories. It is required to mark 4000 more images. This architecture allowed me to achieve the score at 0.962 without significant time expenditure for learning the whole model.