mcfly
mcfly copied to clipboard
NLeSC
A deep learning tool for time series classification and regression
 from azure blog: https://learn.microsoft.com/en-us/azure/machine-learning/how-to-auto-train-forecast?view=azureml-api-2 paper: https://arxiv.org/abs/1803.01271
Before training any model, we should shuffle the data. This is best practice and failing to do so can totally impair training.
According to https://mcfly.readthedocs.io/en/latest/reference.html?highlight=early#mcfly.train_models_on_samples, ```Unless ‘None’ early Stopping is used for the model training. Set to integer to define how many epochs without improvement to wait for before stopping. Default is...
See for a list of possible architectures on the bottom of p3 in Ismael et al - Benchmarking Deep Learning Interpretability in Time Series Predictions: https://proceedings.neurips.cc/paper/2020/file/47a3893cc405396a5c30d91320572d6d-Paper.pdf
This is (multi-dimensional) regression for each time step, where the target is the value of the next time step
The documentation hasn't built correctly on readthedocs, so it still shows an old version of the documentation.
We should drop 3.6 and include up to 3.10 (and maybe already 3.11?)
In find_best_architecture using this combination throws a cryptic error a user is probably not going to understand.
maybe check mexca or dianna-ai/dianna for possible examples
When giving a tf.keras.metrics object as metric in `find_best_architecture()`, e.g., `find_best_architecture(..., metric=tf.keras.metrics.Precision())`, an error is raised: ``` models = modelgen.generate_models(X_train.shape, y_train.shape[1], number_of_models=number_of_models, task=task, metrics=[metric], **kwargs) _, val_performance, _ = train_models_on_samples(X_train,...
