Features Requests

Be able to customize the loss function so that more non-standard losses can be used. This would be implemented so that each Target component would have its own loss function that can be changed and customized using custom code. A summary of which loss function(s) are used would be seen in the Training settings so it's not easy to miss.

So that any customized component that's created can be saved and re-used in other models.

So that it's easier to handle cases with multiple numerical inputs, rather than having each become its own input component.

This would add a component which can be placed on the workspace and then used for loading an already existing TensorFlow model. This is so that you can easily get any models you already have built into PerceptiLabs.

https://www.figma.com/file/tJktsFg85QstCARNnGBJpp/%5BTool%5D-Wireframes-2?node-id=4265%3A22779

These are two very important training settings that are currently missing. They can help a lot with making sure that the model does not overfit.

This is to support older models inside the tool, and to make sure that there is always a conversion path from the older versions to the newer versions. This can be done by maintaining the tool version inside the model.json file and then build compatibility functions for every version in which we make changes to the model.json files. Could also be good to maintain a repository of model.json example versions.

This would add a searchbar to the Workspace which lets you search for any Keras operation and turn it into a PerceptiLabs component. This provides a huge amount of extra flexibility to the custom modelling part of the tool (without having to create custom components).

So that it's easy to go between model iterations/versions and find the one that performed the best. A version would be saved every time the model was ran, and the different versions together with their performance would be viewable from a modal in the ModelHub. Setting a version as "Active" would change your your model into that version.

This would show notifications during training that explains how the training is going. As an example, it might alert you that the gradients are 0 in a component, or that it's starting to overfit, etc.