UI \ UX Interface

This is the topic on UI\UX issues, comments and proposals.

  1. [feature request] It is extremely necessary to have Submodule block, where part of the model can be stored. Block should have multiple inputs and multiple outputs, just regular layers connections, that are converted to TF Tensors when compiling model and verifying shapes, not Input and Output layers of the network.
  2. [issue] please, fix copy paste and delete options when selected group of connected blocks, under Chrome browser on Win 8.1 option seems working randomly, so best choice for now is to operate one-by-one component
  3. [issue] limited canvas, that depends on browser scale, model scale and monitor resolution. Why do we need such? We will be building big models having numbers of similar sub-blocks ending up with hundreds elements in the whole model.
  4. [issue] operations with block becomes slower with number of blocks increase, please use more optimal algo and structure for storing, accessing\changing, deleting elements. It seems algo iterates over whole model with every UI operation, or reallocates memory or copying all stuff for some reason…
  5. [feature request] Add possibility to change the color of the block, so blocks performing as functional unit in the network can be colored in one color, that makes model better readable
  • on the results tab:
    – make the checkbox color same as the charts color.

  • make it possible to rename a Training Run after started or at least after finished (on the results tab)

  • a confusion matrix would be great

  • on the inference tab a treshold setting (like in customvision.ai) and a filter to search for misclassified samples would be great

  • if user is dragged a pretrained model onto the canvas (like inception) and not set the proper preprocessing on the tab - popup a warning message.

[usability] turn off “deselect all” option when using scroll, its frustrating to select things again and again when scrolling canvas with horizontal and vertical scroll controls. (mouse scroll up\down with wheel works fine)

Thanks for sharing feedback. This helps us understand user point of view and make our software further user friendly.

Our team will go through this feedback and file enhancement requests.

One of these (renaming a run) has already been requested and we have implemented it. It will be available in next version soon.

[bug] No recompute of dimensions at the output of pre-trained model block when changing from ‘default’ or ‘th’ to ‘tf’, causing wrong or even negative numbers