Hyperparameter Optimization via Randomization

classic Classic list List threaded Threaded
9 messages Options
Reply | Threaded
Open this post in threaded view
|

Hyperparameter Optimization via Randomization

Phillip Henry
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Sean Owen-2
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Phillip Henry
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Sean Owen-2
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Phillip Henry
Hi, Sean.

Perhaps I don't understand. As I see it, ParamGridBuilder builds an Array[ParamMap]. What I am proposing is a new class that also builds an Array[ParamMap] via its build() method, so there would be no "change in the APIs". This new class would, of course, have methods that defined the search space (log, linear, etc) over which random values were chosen.

Now, if this is too trivial to warrant the work and people prefer Hyperopt, then so be it. It might be useful for people not using Python but they can just roll-their-own, I guess.

Anyway, looking forward to hearing what you think.

Regards,

Phillip



On Fri, Jan 29, 2021 at 4:18 PM Sean Owen <[hidden email]> wrote:
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Sean Owen-2
I was thinking ParamGridBuilder would have to change to accommodate a continuous range of values, and that's not hard, though other code wouldn't understand that type of value, like the existing simple grid builder.
It's all possible just wondering if simply randomly sampling the grid is enough. That would be a simpler change, just a new method or argument.

Yes part of it is that if you really want to search continuous spaces, hyperopt is probably even better, so how much do you want to put into Pyspark - something really simple sure. 
Not out of the question to do something more complex if it turns out to also be pretty simple.

On Sat, Jan 30, 2021 at 4:42 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

Perhaps I don't understand. As I see it, ParamGridBuilder builds an Array[ParamMap]. What I am proposing is a new class that also builds an Array[ParamMap] via its build() method, so there would be no "change in the APIs". This new class would, of course, have methods that defined the search space (log, linear, etc) over which random values were chosen.

Now, if this is too trivial to warrant the work and people prefer Hyperopt, then so be it. It might be useful for people not using Python but they can just roll-their-own, I guess.

Anyway, looking forward to hearing what you think.

Regards,

Phillip



On Fri, Jan 29, 2021 at 4:18 PM Sean Owen <[hidden email]> wrote:
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Phillip Henry
Hi, Sean.

I don't think sampling from a grid is a good idea as the min/max may lie between grid points. Unconstrained random sampling avoids this problem. To this end, I have an implementation at:


It is unit tested and does not change any already existing code.

Totally get what you mean about Hyperopt but this is a pure JVM solution that's fairly straightforward.

Is it worth contributing?

Thanks,

Phillip





On Sat, Jan 30, 2021 at 2:00 PM Sean Owen <[hidden email]> wrote:
I was thinking ParamGridBuilder would have to change to accommodate a continuous range of values, and that's not hard, though other code wouldn't understand that type of value, like the existing simple grid builder.
It's all possible just wondering if simply randomly sampling the grid is enough. That would be a simpler change, just a new method or argument.

Yes part of it is that if you really want to search continuous spaces, hyperopt is probably even better, so how much do you want to put into Pyspark - something really simple sure. 
Not out of the question to do something more complex if it turns out to also be pretty simple.

On Sat, Jan 30, 2021 at 4:42 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

Perhaps I don't understand. As I see it, ParamGridBuilder builds an Array[ParamMap]. What I am proposing is a new class that also builds an Array[ParamMap] via its build() method, so there would be no "change in the APIs". This new class would, of course, have methods that defined the search space (log, linear, etc) over which random values were chosen.

Now, if this is too trivial to warrant the work and people prefer Hyperopt, then so be it. It might be useful for people not using Python but they can just roll-their-own, I guess.

Anyway, looking forward to hearing what you think.

Regards,

Phillip



On Fri, Jan 29, 2021 at 4:18 PM Sean Owen <[hidden email]> wrote:
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Sean Owen-2
It seems pretty reasonable to me. If it's a pull request we can code review it.
My only question is just, would it be better to tell people to use hyperopt, and how much better is this than implementing randomization on the grid.
But the API change isn't significant so maybe just fine.

On Mon, Feb 8, 2021 at 3:49 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

I don't think sampling from a grid is a good idea as the min/max may lie between grid points. Unconstrained random sampling avoids this problem. To this end, I have an implementation at:


It is unit tested and does not change any already existing code.

Totally get what you mean about Hyperopt but this is a pure JVM solution that's fairly straightforward.

Is it worth contributing?

Thanks,

Phillip





On Sat, Jan 30, 2021 at 2:00 PM Sean Owen <[hidden email]> wrote:
I was thinking ParamGridBuilder would have to change to accommodate a continuous range of values, and that's not hard, though other code wouldn't understand that type of value, like the existing simple grid builder.
It's all possible just wondering if simply randomly sampling the grid is enough. That would be a simpler change, just a new method or argument.

Yes part of it is that if you really want to search continuous spaces, hyperopt is probably even better, so how much do you want to put into Pyspark - something really simple sure. 
Not out of the question to do something more complex if it turns out to also be pretty simple.

On Sat, Jan 30, 2021 at 4:42 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

Perhaps I don't understand. As I see it, ParamGridBuilder builds an Array[ParamMap]. What I am proposing is a new class that also builds an Array[ParamMap] via its build() method, so there would be no "change in the APIs". This new class would, of course, have methods that defined the search space (log, linear, etc) over which random values were chosen.

Now, if this is too trivial to warrant the work and people prefer Hyperopt, then so be it. It might be useful for people not using Python but they can just roll-their-own, I guess.

Anyway, looking forward to hearing what you think.

Regards,

Phillip



On Fri, Jan 29, 2021 at 4:18 PM Sean Owen <[hidden email]> wrote:
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip

Reply | Threaded
Open this post in threaded view
|

Re: Hyperparameter Optimization via Randomization

Phillip Henry
Hi, Sean.

I've added a comment in the new class to suggest a look at Hyperopt etc if the user is using Python.

Anyway I've created a pull request:


and all tests, style checks etc pass. Wish me luck :)

And thanks for the support :)

Phillip



On Mon, Feb 8, 2021 at 4:12 PM Sean Owen <[hidden email]> wrote:
It seems pretty reasonable to me. If it's a pull request we can code review it.
My only question is just, would it be better to tell people to use hyperopt, and how much better is this than implementing randomization on the grid.
But the API change isn't significant so maybe just fine.

On Mon, Feb 8, 2021 at 3:49 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

I don't think sampling from a grid is a good idea as the min/max may lie between grid points. Unconstrained random sampling avoids this problem. To this end, I have an implementation at:


It is unit tested and does not change any already existing code.

Totally get what you mean about Hyperopt but this is a pure JVM solution that's fairly straightforward.

Is it worth contributing?

Thanks,

Phillip





On Sat, Jan 30, 2021 at 2:00 PM Sean Owen <[hidden email]> wrote:
I was thinking ParamGridBuilder would have to change to accommodate a continuous range of values, and that's not hard, though other code wouldn't understand that type of value, like the existing simple grid builder.
It's all possible just wondering if simply randomly sampling the grid is enough. That would be a simpler change, just a new method or argument.

Yes part of it is that if you really want to search continuous spaces, hyperopt is probably even better, so how much do you want to put into Pyspark - something really simple sure. 
Not out of the question to do something more complex if it turns out to also be pretty simple.

On Sat, Jan 30, 2021 at 4:42 AM Phillip Henry <[hidden email]> wrote:
Hi, Sean.

Perhaps I don't understand. As I see it, ParamGridBuilder builds an Array[ParamMap]. What I am proposing is a new class that also builds an Array[ParamMap] via its build() method, so there would be no "change in the APIs". This new class would, of course, have methods that defined the search space (log, linear, etc) over which random values were chosen.

Now, if this is too trivial to warrant the work and people prefer Hyperopt, then so be it. It might be useful for people not using Python but they can just roll-their-own, I guess.

Anyway, looking forward to hearing what you think.

Regards,

Phillip



On Fri, Jan 29, 2021 at 4:18 PM Sean Owen <[hidden email]> wrote:
I think that's a bit orthogonal - right now you can't specify continuous spaces. The straightforward thing is to allow random sampling from a big grid. You can create a geometric series of values to try, of course - 0.001, 0.01, 0.1, etc.
Yes I get that if you're randomly choosing, you can randomly choose from a continuous space of many kinds. I don't know if it helps a lot vs the change in APIs (and continuous spaces don't make as much sense for grid search)
Of course it helps a lot if you're doing a smarter search over the space, like what hyperopt does. For that, I mean, one can just use hyperopt + Spark ML already if desired.

On Fri, Jan 29, 2021 at 9:01 AM Phillip Henry <[hidden email]> wrote:
Thanks, Sean! I hope to offer a PR next week.

Not sure about a dependency on the grid search, though - but happy to hear your thoughts. I mean, you might want to explore logarithmic space evenly. For example,  something like "please search 1e-7 to 1e-4" leads to a reasonably random sample being {3e-7, 2e-6, 9e-5}. These are (roughly) evenly spaced in logarithmic space but not in linear space. So, saying what fraction of a grid search to sample wouldn't make sense (unless the grid was warped, of course).

Does that make sense? It might be better for me to just write the code as I don't think it would be very complicated.

Happy to hear your thoughts.

Phillip



On Fri, Jan 29, 2021 at 1:47 PM Sean Owen <[hidden email]> wrote:
I don't know of anyone working on that. Yes I think it could be useful. I think it might be easiest to implement by simply having some parameter to the grid search process that says what fraction of all possible combinations you want to randomly test.

On Fri, Jan 29, 2021 at 5:52 AM Phillip Henry <[hidden email]> wrote:
Hi,

I have no work at the moment so I was wondering if anybody would be interested in me contributing code that generates an Array[ParamMap] for random hyperparameters?

Apparently, this technique can find a hyperparameter in the top 5% of parameter space in fewer than 60 iterations with 95% confidence [1].

I notice that the Spark code base has only the brute force ParamGridBuilder unless I am missing something.

Hyperparameter optimization is an area of interest to me but I don't want to re-invent the wheel. So, if this work is already underway or there are libraries out there to do it please let me know and I'll shut up :)

Regards,

Phillip