Added option for custom dataloader and the possibility of using metrics from torchmetrics

Fixed a series of breaking changes related to the metrics returning arrays. Adapted docs to the new functionalities

I kept almost all previous code as it was, I just added functionality for custom dataloaders and possibility to use torchmetrics

Tabnet

Specified python 3.7 version to avoid a test fail when using travis that I cannot reproduce locally neither on the mac or on a ubuntu machine

Specified python 3.7 version to avoid a test fail when using travis that I cannot reproduce locally neither on the mac or on a ubuntu machine

Refined documentation. Added some test for saving model. Added an example notebook. Ready to test installations and publish v1 to pypi

Added test for new trainer methods. Back to documentation style more in line with some popular packages like pytorch

added get and setstate methods in the EarlyStopping and ModelCheckpoint callbacks. Added the possibility of using GRUs in the deeptext component and also predict using the hidden state or the output. Fixed a small bug in the text processor. Improved the save method in the Trainer

ModelCheckpoint saves best epoch as well. Added dropout option for tabnet. Adjusted RAdam for new signatures. Adjusted the training so it can take ReduceLROnPlateau. Also so that it automatically restores the best weights after training

Modified the callbacks to accomodate the neccessity of the ReduceLROnPlateau scheduler. Replace weight by pos_weight in BCEWithLogitsLoss. Added on_eval_begin method to reset metrics

Fixed a documentation error. For the tabtransformer the input_embed is a list with tuples of 2 elements, not 3

Change tab_resnet to tab_resnet_blks. Added types to tab_net.py. Cleaned up trainer by moving some methods to tab_net_utils. Adjusted tests

Added a few types, refined the docs and added tests for tabnet and the corresponding functionalities

added an option to automatically set embed size via fastai's rule of thumb. Re-structure the preprocessing module, breaking it up in a smaller modules for better debugging and 'tractability'

added explain and compute feature importance methods to the trainer. Need to add error handling and messaging since these funcionalities are intended only for tabnet. Move the general_utils module to the training module and rename it as trainer_utils. Adapted the create_explain_matrix function to WideDeep

finished implementing TabNet encoder. Next I need to implement the interpretability methods. Also adjusted WideDeep class to be able to work with TabNet. Renamed some example scripts to be consistent with each other

fixed a naming inconsistency. The Transformer encoder blocks were named 'blks'. A more consistent name relative to the other two tab models would be 'tab_transformer_blks', and so I have changed it. This leads to a small inconsistency between pypi version and that in github, which is almost irrelevant

for reasons beyond me both lates and stable doc versions as built by Read the Docs look pretty different than the html built locally, which is really frustrating. I have just built a new version pointing to master and changed the link to the documents

Tabtransformer