All examples are complete runnable scripts and are self-contained to ease understanding
(the only exception is dataset fetching that has been moved to common.py).
- Hello World: a very simple example to run your existing model in HyperEngine environment.
- Getting started with tuning:
tuning a single parameter (
learning_rate) to find the best value for a simple CNN. - Saving best models: tuning several hyper-parameters and saving the best CNN models on the disk.
- Fine-tuning the saved model: training the selected model further to squeeze the highest possible accuracy out of it.
- Learning curve prediction: optimizing the process with the learning curve prediction.
- Regression problem: HyperEngine is designed for classification problems, but with some tricks can be applied in regression settings as well.
- Exploring CNN architectures to beat MNIST record: exploring and tuning all possible variations of the CNN design to get to the top accuracy for the MNIST dataset.
- Exploring CNN architectures for CIFAR-10 dataset: exploring different variations of the CNN design to get good accuracy for the CIFAR-10 dataset (state-of-the art accuracy would require a bit more efforts - in the next examples).
- Exploring architectures of All Convolutional Nets for CIFAR-10 dataset: exploring different variations of the fully convolutional nets (FCNN), which achieve state-of-the-art accuracy with few parameters and computational costs. See "Striving for Simplicity: The All Convolutional Net" paper for details.
- LSTM to classify MNIST digits: recurrent neural networks can process images too. Let's see if it can get to 99% with right hyper-parameters.
- RNN to detect SMS spam: solving a simple NLP problem with different types of RNN. This example shows how hyper-parameters can be used in data processing.
- Word2Vec training: an example of custom data provider to train Word2Vec embedding vectors. In addition, it shows how to use the solver without accuracy metric and validation.