Hey there! I hope it's ok for me to answer only by now.

Am I correct in assuming the algorithm stores all the data points it sees in memory (i.e. the X attribute)?

Hey there! I hope it's ok for me to answer only by now.

Am I correct in assuming the algorithm stores all the data points it sees in memory (i.e. the X attribute)?

Hello, sorry for replying that late.
I use the window_size attribute and when there are more data than allowed, it deletes the oldest data point.
If window_size < 1, then it stores all the data points

If window_size < 1, then it stores all the data points

Ok I see, fair. But I don't think we'll ever want that behavior. Could you remove it?

If window_size < 1, then it stores all the data points

Ok I see, fair. But I don't think we'll ever want that behavior. Could you remove it?

Yes I can. So I add an error if the value of window_size is not an integer > 0 ?

Nope, no need to check for an error. An exception will raise itself at some point. In general, we don't do input validation. Instead, we document well.

Okay ! I will delete it

@MaxHalford I think I'm quite happy with the current state of the code! Only one small concern is that the unit tests keep failing because there is an attribute error, saying that HierarchicalClustering does not have the attribute distance_func, although I believe atm there is no distance_func in my code anywhere anymore. Do you think of any potential reasons behind this?

@MaxHalford @kchardon If possible, I would really hope that both of you would be able to take a final look at the current state of the implementation and see if there are any changes you want to make. In case there is no opposition from both of you, when applicable, I would want to merge this into the main branch of River.

I will review next week :)

I (finally) took a look at this. There's a lot of nits to fix. But the main issue I see is that the code relies on numpy. It converts input dictionaries to numpy arrays. Would it be possible to use dictionaries only instead? I don't see any good reason to rely on numpy. Indeed, it's not in the spirit of River to rely on numpy when it can be avoided.