Show the parameters and the results

[Rob174]

Potential solutions

Solution	:+1:	:-1:
Tensorboard	Already ready	Curves cannot be used directly in papers (too small axes, curves, values)
Dash+plotly	Graph easy to do	Interface to make. Potentially no regex filtering
Vuejs+node	Entirely personnalized	Time consuming

[Rob174]

Other possibility:

Tensorboard to log curves but also a script to analyse event files of the tensorboard to allow the user to compare the curves in a notebook with plotly.

Confusion matrix as html to avoid formatting problems

[Rob174]

Finally: all jupyter notebook with widgets solution

[Rob174]

Visualize the network

Goals:

• see the structure of the network, its building blocks
• see the transformations applied to the image : print or graphviz processing
• how to compare and filter architectures

--> 2 levels of detail

• see the parameters

Graphviz/d3js interactive graph allowing to choose the structure detail level and if we want to display the hyperparameters.

Eventually add visualization of the kernel on hover event

Problems to solve:

• retrieve parameter name when the user directly supply the value without the parameter name (eventually with eval and using already imported classes)
• how to reduce the area taken by the graph (manually ? But if so how to save the changes)
• how to identify the repeating patterns in the architecture ? (Automatic ? Thanks to additional grouping layers)
• how to integrate view user choices ? (ipywidgets ??)
• last functionality ?? (Highlighting common patterns, showing number of parameters, clustering networks with neighbour method )

[Rob174]

Not the main priority at the moment