huntler / timeseriesdl Goto Github PK

Add missing dataset.scale_back() unit test.

Add a model similar to the ConvLSTM model, but replace the LSTM with a GRU. This should speed up the training and inference, while lowering the capability to learn complex behaviours. As complex structures may not be the case, a GRU can be beneficial.

If a key in the config gets the value 'auto' assigned, the config manager should look for equally named keys. Then, the value of the other key should replace 'auto'. Make sure, that at least one key needs a real value instead of 'auto'

Add a unit test for transcoding a dataset, which includes testing a (trained) AE.

Test the collate_fn which utilizes an autoencoder. Transfer the DummyAutoEncoder to its desired package and make it more usable for the future.

Add a dynamic Conv2d layer, which takes the arguments

• from_size: describing the input size
• to_size: describing the desired output size
• hidden_layers: describing the number of Conv2d layers generated

Use this custom Conv2d layer in every model implemented so far and adapt the config.yaml files accordingly. This should simplify the testing with model/parameter sizes.

Use pytorch lighting for training the model as this is less likely to intorduce bugs and offers more features e.g. distributed training

Replace TensorBoard with MLFlow to gain more insights into training

Create a pipeline executing trainings of multiple models (e.g. ConvAE -> LSTM) with a single log folder containing:

• TensorBoard logs
• saved models
• configuration YAML files

Auto-Tune hyperparameters, use existing frameworks with methods e.g. Bayesian Optimization