Add Stochastic Multiplicative Updates

[bantin]

No description provided.

[ahwillia]

Nice work! Does this enable us to fit to the full Maze dataset?

Small notes:

• I wonder if it makes sense to define a separate StochasticMU class.
• What is the intended API? Something like model = CMF(alg_name="stochastic_mu")?

Thanks again - excited to try this out!

[bantin]

I haven't tried with the full maze dataset but from my short experiments it can work well with some tuning of the window size parameter. I thought about defining a separate class but I thought it might get out of hand if we have stochastic variants of many algorithms. Right now the API is to pass "stochastic" = True, "window_size" = N. I can make a separate class if we think that would be better.

…

On Wed, Nov 21, 2018 at 12:23 PM Alex Williams ***@***.***> wrote: ***@***.**** commented on this pull request. Nice work! Does this enable us to fit to the full Maze dataset? Small notes: - I wonder if it makes sense to define a separate StochasticMU class. - What is the intended API? Something like model = CMF(alg_name="stochastic_mu")? Thanks again - excited to try this out! — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIrPZp7is_G3kL10bi2AWvvssCphtBrSks5uxbZCgaJpZM4Ytrkb> .

[bantin]

Also -- I just realized that the API isn't so great currently. If you pass "stochastic" = True with an algorithm that doesn't support that option it won't give an error or a warning. I think I should fix that before we merge.

…

On Wed, Nov 21, 2018 at 1:06 PM Benjamin Antin ***@***.***> wrote: I haven't tried with the full maze dataset but from my short experiments it can work well with some tuning of the window size parameter. I thought about defining a separate class but I thought it might get out of hand if we have stochastic variants of many algorithms. Right now the API is to pass "stochastic" = True, "window_size" = N. I can make a separate class if we think that would be better. On Wed, Nov 21, 2018 at 12:23 PM Alex Williams ***@***.***> wrote: > ***@***.**** commented on this pull request. > > Nice work! Does this enable us to fit to the full Maze dataset? > > Small notes: > > - I wonder if it makes sense to define a separate StochasticMU class. > - What is the intended API? Something like model = > CMF(alg_name="stochastic_mu")? > > Thanks again - excited to try this out! > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#23 (review)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AIrPZp7is_G3kL10bi2AWvvssCphtBrSks5uxbZCgaJpZM4Ytrkb> > . >

[ahwillia]

see some suggestions in #24

[bantin]

@ahwillia Updated the cache_resids function as you suggested. Thoughts?

[ahwillia]

Before merging, can you add a small notebook testing this out on the Maze dataset?

[bantin]

Will do!

…

On Thu, Nov 22, 2018 at 11:43 AM Alex Williams ***@***.***> wrote: Before merging, can you add a small notebook testing this out on the Maze dataset? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIrPZjutw-sdIYimxeb-Oph-YFXPvy2Tks5uxv5lgaJpZM4Ytrkb> .

[degleris1]

Merge issues (just in __init__.py) resolved in the commit above. The stochastic update rule looks good to me!

[bantin]

@ahwillia Just added a simple notebook that shows fitting a matrix with 20k timebins in 8.5 mins. Admittedly, the fit isn't great, but if you think it's good enough to show the stochastic functionality, we can probably merge this to master.

[degleris1]

Looks good!