reshapor function now supports general number of inputs ...

Using 90% as train data
New helper function correct_array_steps which is used in scaling added
min_max scaler and std_scaler now much more slim and stable
standard minibatch size changed to 100
scaler function now supports external scalers loaded with full_load
RNNPlacePrediction class has loss_validation_list
Early stopping now based of off loss on validation set (Checkpoints as well)
Full_save function updated to support new characteristic of RNNPlacePrediction class (loss_validation_list)
Full load function updated and bug fix so it always loads the scaler correctly as well

New trained model added (best performance yet)

Some trained models added and their .pkl data added ...

Minor change in the rnn test function

1_to_1_multi_layer: Some minor changes to full_save and full_load to optionally include the train and test data and load it accordingly ...

1_to_1_multi_layer: Added a function to test the rnn

1_to_1_multi_compact: A shortened version which loads all the modules, my functions and helperfunctions and the RNNPlacePrediction class with requirements.py and just does the training and testing of the rnn
requriements.py: Loads all modules, does the preprocessing and defines all functions and helperfunctions (to be loaded into 1_to_1_multi_compact)

2 trained models (5 layer GRU and 5 layer lstm) added and their respective pkl data to allow a full_load()

Test on second laptop

Smarter checkpoints ...

Going back to earlier last checkpoint if persormance was better than at end of training
Scalers now work on on 2d and 3d arrays
full_load and full_save added which save and restore the rnn with all parameters
get_rnn_folder added which is a helper function for full_load and provides the folder name for an architecture with the correct notation
Also dataset added

Standard scaler (and inverse) added for 2d and 3d arrays of all shapes ...

Min Max scaler (and inverse) added for 2d and 3d arrays of all shapes
Scalers now also work on test set
Huge performance increase
Finetuning for patience trigger for scalers missing
Trained model for newest version added and old version discarded

Normalization with skicit learn added (Doesnt work on test_set yet) ...

Early stopping with patience refined (How to find optimal patience and trigger value?)
Elu activation added
Last activation i always tanh
Checkpoints added

Loss normalized ...

Early stopping with patience improved and fine tuned
Improved checkpoints
Bugfixes
Plot of loss added
Trained save of rnn_lstm_5l_[50,40,30,20,10] added

Early stopping with patience added ...

Checkpoints during training added to go back to best state if perfomance decreases

Removed n_layer as with the new ncells its implicitly stated and added Errors if ncells is of the wrong type (has to be int or a list of ints)

Delete copy

Revert "Revert "This is the more advanced version of the old program"" ...

This reverts commit 963add1.

This is the more advanced version of the old program Added early stopping Improved performance It supports now many layers and an arbitrary amount of cells per layer. It also saves in a folder named after the models architecture

This is the more advanced version of the old program ...

Added early stopping
Improved performance
It supports now many layers and an arbitrary amount of cells per layer.
It also saves in a folder named after the models architecture

Revert "This is the more advanced version of the old program" ...

This reverts commit ca028c5.

This is the more advanced version of the old program ...

Added early stopping
Improved performance
It supports now many layers and an arbitrary amount of cells per layer.
It also saves in a folder named after the models architecture

Revert "Test for rnn state" ...

This reverts commit 82de3ca.

Test for rnn state

first test

Add files via upload

Initial commit