UI suggested change: "Network" -> "Model Configuration"

[Make42]

In https://clear.ml/docs/latest/docs/webapp/webapp_model_viewing#model-configuration you talk about networks. I am guessing you mean neural networks. I am not using any though ANNs - maybe this can/should be adjusted to "Model Configuration"?

[ainoam]

Thanks for suggesting @Make42.

Indeed the tab content already discusses "Model Configuration". You'll notice the ClearML currently uses the terms Model, Network and Configuration somewhat interchangeably. Do you find this especially confusing?

[Make42]

@ainoam: I am saying the following not to bash, but as in "tough love", because I want to see ClearML improve and succeed! Just to clarify my motivation :-).

It is confusing, because it is factually wrong.

A model is not its own configuration. When you just say "configuration", I have to ask "Configuration of what?" If you answer "the configuration of the model" then, why don't you just write "model configuration". This says it is the model's configuration.

A network is a thing which has nodes and edges between those nodes. That is it. A network can be all kinds of things (these examples are all relevant in data science):

• The internet is a network.
• If I have a computer cluster, they create a network, e.g., in a big data context.
• If I work work with graph data, I might do "network analysis".
• And a artificial neural network is a network, but only one (more specific) type among many types of networks.

A model - in general - is NOT a network.

A "graph", by the way, is a graphical representation of a network, not the network itself. Alternatively a "graph" is a graphical representation of a function, but this is off-topic here ;-).

[ainoam]

Much appreciate the love @Make42.

You are, of course, correct. The only point I was trying to make was that "network" was somewhat (if ill-) fitting as in some of the use cases the values therein indeed outline the architecture of a neural network. Obviously, since this tab is in the context of a model page, we're not discussing any abstract network, but the one relating to the model.

In any case, I think perhaps "Configuration" would be the most befitting choice for a rename in our case - WDYT?

[Make42]

@ainoam: I understand that "network" here is in the context of "artificial neural network". The problem is that the tab is not only for neural network models, but for any model. (And in my case, I do not use a NN, but still the "network" tab.) Anyways, I think you already understood what I mean :-D.

I think "configuration" or "model configuration" is fine. However, this directly opens up the question, how this "configuration" information differs from the information displayed in the tab "metadata". I started to discuss this in https://github.com/allegroai/clearml-docs/issues/477. A similar discussion is raised by me in https://github.com/allegroai/clearml-docs/issues/478. It could be that a simplification of ClearML might be reasonable; however, it might be that the current way it is makes totally sense, but it need to be better communicated (at least to me, but possibly others) with other naming and clearer documentation.

PS: I am not sure if the issue should be there at "clearml-docs" or here.

[Make42]

@ainoam: There is another, related issue, so I will just add it here: the Model API has the method update_weights. This is so related to ANNs. I am using it to upload my entire model, disregarding any "weights". It is pretty weird as a user. Wouldn't "update model" would have been a good option? Or maybe you have something very specific in mind in the background where you take advantage on online learning and gradual updating of weights? Maybe then you need two different methods with different names?

I am not deep enough into ClearML yet, to say exactly how this should be solved, but me uploading a zip file containing a random forest with "update_weights" and an argument "weights_filename" is definitely weird.

[idantene]

@Make42 Jumping in here, just out of common pain points (I 100% agree with all your suggestions and comments).

This may or may not be known to you, depending on when ClearML first popped on your radar, but ClearML definitely started with CNNs as the driving force. Hence the terms relating specifically to ANNs, the Datasets (and Hyper Datasets) that are mostly aimed at images, the forced iteration argument in logging, the placement of images as "debug samples" (though they can easily be more correctly listed under Plots or Artifacts), etc.

[Make42]

@idantene: Thanks for this helpful clarification! This ties in with the issues I have described in https://github.com/allegroai/clearml-docs/issues/478