martinjm97 / ennui Goto Github PK

Select multiple objects by dragging a selection box across the canvas for deletion and multi dragging

To allow users to send each other urls with their architectures.

Right now, every time train is pressed you have to download the mnist dataset. We should cache it.

What should the user be able to change in the menu on the left?

When clicked a first time, start training. If clicked during training check if the user wants to cancel the previous job and start a new job. This should probably be implemented with observables

k should be like 5 or 10. Check that performance improves and plots look more smooth

Add in the option to downsample data so slower computers are more able to run app.

Add a template with non-sequential model (Maybe add a second template of a model that performs very badly on MNIST)

If you drag a shape (layer) to the edge of the canvas, it can be partially obscured.

This implies making the layer parameters an actual member of Layer and adding hooks with the parameter boxes.

I tried the architecture of input straight to output and that got 85% validation accuracy. That seems fishy to me...

When you're on the Progress tab or the Visualization tab both are displayed. This damages performance.

If someone clicks train on an invalid architecture raise the tf error to a little red message box in the lower left corner. These messages should be friendly and understandable.

This only happens for the very first layer you select.

View Images that pass and fail classification.

https://stackoverflow.com/questions/13821994/d3-js-transition-bug-in-firefox

Multiple clicks on the train button launches multiple runs

They are not very informative e.g. if input and output are not connected it gives errors.

This means removing all appearances of 'px' everywhere and fixing with floating, relative positions, and percentages.

Generate the Keras code in Python for the given network architecture.

Batch size, learning rate, optimizers (Adam, Momentum, etc), training and validation accuracy on graph.
We can add even more for V2

Currently overwrites

When you click on a layer the parameters that show up on the right have to be parenthesized properly.