Trained model weights

The trained model weights, model architectures, training histories and performance visualizations are available for download in release v0.1.

Each experiment is named according to the following scheme:

datasetModel-properties-target-normalization-window.pickle

Where:

• datasetModel denotes the certain model used in the experiment dataset and subsequently the neural model. "matrix" is used in LSTM experiments, while "stacked" is used with CNN.
• properties is a list of property names, separated by a comma. These are the properties included in the dataset.
• target is the name of the target property
• normalization is the name of the normalization strategy used. In the paper, pixel is defined as image and property is defined as prop.
• window is the window size of the dataset.

The following files are available for each experiment:

• (experiment name).h5: The trained model weights, including the neural architecture and hyperparameters. This file can be loaded by the neural_trainer module and evaluated or additionally trained.
• (experiment name).pickle: A metadata file, containing information about the experiment, such as the used dataset, training parameters and detailed model history for each of the defined error measures in the paper. This file can be loaded by the result_visualizer tool and the histories can be visualized. Furthermore, the tool can read multiple metadata files and compare the experiments based on a chosen measure.
• (experiment name).svg: A rendered image of the trained model prediction graphs on the train dataset (top) and test dataset (middle). The visualized performance histories during the epochs for every error measure are found on the bottom.