Pseudo-label with Keras

This is an implementation of the semi-supervised aproach called pseudo-label using Keras. This algorithm is based on article "Pseudo-Label : The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks" by Dong-Hyun Lee.

warning: This project is under development! To report a problem, create an Issue

author: Gabriel Kirsten Menenezes (https://github.com/gabrielkirsten/)

Pseudo-label algorithm

graph LR
A[Your dataset] --> B(Unlabel dataset)
B -- predict --> G
A --> C(Labeled dataset) 
C --> E{Supervised trainning}
E --> G[Your trained model]
G --> F(Pseudo labels)
F --> D{Semi-supervised trainning}
C --> D
D --> H[Your final trained model]

Requirements:

You must Install:

1. Python 2.7;
2. Nvidia cuda libraries;
3. Nvidia cuDCNN libraries;
4. Tensorflow or Theano*;
5. Keras;
6. Sklearn;
7. h5py.

note:
* never tested on Theano.

How to use:

usage: main.py [-h] [-a ARCHITECTURE [ARCHITECTURE ...]] [-f FINETUNINGRATE] -d DATASETPATH [-n NOLABELPERCENT [NOLABELPERCENT ...]]

optional arguments:

  -h, --help            show this help message and exit
  -a ARCHITECTURE [ARCHITECTURE ...], --architecture ARCHITECTURE [ARCHITECTURE ...]
                        Select architecture(Xception, VGG16, VGG19, ResNet50,
                        InceptionV3, MobileNet)
  -f FINETUNINGRATE, --fineTuningRate FINETUNINGRATE
                        Fine tuning rate
  -d DATASETPATH, --datasetPath DATASETPATH
                        Dataset location
  -n NOLABELPERCENT [NOLABELPERCENT ...], --noLabelPercent NOLABELPERCENT [NOLABELPERCENT ...]

Parameters:

Some parameters are defined at the main.py as:

• Image width (IMG_WIDTH), Image height (IMG_HEIGHT) and Image channels (IMG_CHANNELS);
• Supervised batch size (BATCH_SIZE);
• Pseudo-label batch size (PSEUDO_LABEL_BATCH_SIZE);
• Epochs (EPOCHS);
• Class names (CLASS_NAMES).