One of the projects I have done at Comcast Applied AI Lab. This training scheme would guarantee low losses for all patterns for synthetic data In Feedback mini-batch training, the distribution of patterns within a minibatch is dynamically changed to ensure that all the patterns are optimally trained. If a specific pattern is difficult to learn, it is more likely to be found within a minibatch.​ This method increases the performance of the model. More importantly it guarantees that no patterns are undertrained, which provides a common ground for further improvements/interpretation.

You need to install corti-dnlp and corti-data-manager in your virtual enviroment.

• tf_adversarial_minibatching.py is the main file to run, remember to specify parameters and directories.

• generate_data.py can generate data based on given distribution. This is for testing, and traing a normal model.

• test_minibatching.py test the performance of minibatch trained model on a test dataset.

• view_pattern.py,test.py,test1.py tests various properties of the model.

• plotting.py,plot_dist.py,plotting_scatter.py plots loss distributions or just distributions.

• train_normal.py trains a normal model for comparision.