New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
autodiff gaussian width parameter #4782
Conversation
src/shogun/kernel/GaussianKernel.cpp
Outdated
float64_t element=distance(j, k); | ||
derivative(j, k) = std::exp(-element) * element * 2.0; | ||
auto f = exp(-CShiftInvariantKernel::distance(j, k) / constant_part); | ||
f.grad(); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Wow it is quite simple to do this for the bandwidth ... I guess I tried to diff wrt the features last time I tried, which raised problems regarding passing the matrices to stan...but this is cool!
I assume the same results as there are tests covering this?
what are the computational differences compared to the previous line? Maybe a benchmark would be good?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yes, the results are correct and are test in unittests, mostly by GP stuff.
The computation is very slow... Basically it creates a new variable in the heap in the inner loop which is in the heap, and then it redoes the gradient calculation.
What we need is a way to calculate the "formula" of the gradient once outside the loop and then reuse it.. Not sure how to do this in stan or if it is even possible
Rethinking the kernels might be a good idea to parallelise things anyways ... The singe kernel value approach we have is more suited to very expensive kernel functions (string kernels), including the kernel cache (useless for things like the gaussian kernel, here vectorization is much more useful) Any ideas how to go about with this concretely? |
i noticed that for example for euclidean distance this is really innefficient, as we compute one value at the time and do some checks before the computation. I think shogun does the calculation fast, but could be better with some refactoring: |
I think in addition to computing a single kernel value, we would need to offer an API that gets you the matrix. But not via simply looping over single values but rather directly in the lazy way you suggest ... |
@karlnapf turns out eigen does what we want! This would fit nicely with using more eigen and lazily evaluate the distance in the kernel! The latest commit is as fast as the manual derivative! |
64540d1
to
aa229fa
Compare
compile time differentiation is much faster Revert "autodiff gaussian width parameter" This reverts commit e9bced7.
aa229fa
to
a35cf20
Compare
@karlnapf I am thinking we could write our own AutoDiffScalar, but instead AutoDiffArray, that does the same compile time optimisation as the current autodiff but with arrays, so should handle parallelism better? and then we can go through stan and add more ops than there in eigen right now! |
Benchmark calculating the Gaussian kernel gradient with different training example sizes and feature dimensions
|
src/shogun/kernel/GaussianKernel.cpp
Outdated
derivative(j, k) = std::exp(-element) * element * 2.0; | ||
eigen_log_width.derivatives() = EigenScalar::Unit(1,0); | ||
auto el = CShiftInvariantKernel::distance(j, k); | ||
Eigen::AutoDiffScalar<EigenScalar> kernel = exp(-el / (exp(eigen_log_width * 2.0) * 2.0)); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It kinda sucks that the code for the kernel needs to be in here as well as in the kernel itself dont you think?
require(lhs, "Left hand side features must be set!"); | ||
require(rhs, "Rightt hand side features must be set!"); | ||
require(rhs, "Right hand side features must be set!"); | ||
|
||
if (!strcmp(param->m_name, "log_width")) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
so I guess a next step could be to start thinking about getting rid of this explicit code, and rather automatically offer this derivative through registering something in the ctors ....
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think for that we should maybe have a base class for classes that have parameters that we can take the derivative wrt. This class registers the gradient parameters in some vector and then we can get the index from there.
Basically when we do watch_param(...) this would add the variable in such a vector if it has the flag GRADIENT
using EigenScalar = Eigen::Matrix<float64_t, 1, 1>; | ||
Eigen::AutoDiffScalar<EigenScalar> kernel = kernel_function(j, k); | ||
// 0 is the index of the width parameter | ||
derivative(j, k) = kernel.derivatives()(0); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this seems pretty automatable to me (at least say for scalar valued kernels with scalar parameters)
src/shogun/kernel/GaussianKernel.cpp
Outdated
auto CGaussianKernel::kernel_function(int32_t idx_a, int32_t idx_b) | ||
{ | ||
// this could be written as Eigen::Matrix<float64_t, n_differentiable_params, 1>; | ||
m_eigen_log_width = Eigen::AutoDiffScalar<Eigen::VectorXd>(m_log_width); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@karlnapf this should fix the temporary being deleted. Ideally we would replace m_log_width with this... Because this isn't thread safe
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
and for the distance matrix we woudn't have this issue because features holds a reference to the data!
6ed74f7
to
fe0a4aa
Compare
fe0a4aa
to
bb85e1b
Compare
float64_t result=distance(idx_a, idx_b); | ||
return std::exp(-result); | ||
float64_t result; | ||
#pragma omp critical |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@karlnapf the issue now is that this is not thread safe anymore
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
because i assign the value of m_eigen_log_width in kernel_function. Basically would have to do the assignment outside the function, but if a user does ->put("log_width", ..) there is no way of tracking this.. hence the gradient dependent parameters would have to be added to the parameter framework somehow. Or AutoDiffScalar would have a reference instead of a value..
@@ -174,6 +185,7 @@ class CGaussianKernel: public CShiftInvariantKernel | |||
protected: | |||
/** width */ | |||
float64_t m_log_width; | |||
Eigen::AutoDiffScalar<Eigen::VectorXd> m_eigen_log_width; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this idea is actually not too bad: just make the parameters one can diff wrt class members. They just wrap the other member if I understand correctly? These two getting out of sync is he problem you mean with the user putting a parameter? Callback stuff might help? Should be possible to automate something for that or?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
a more consequent version of this would be to make all model parameters live in eigen, and all the "model" function simply returning composed versions of those... but i guess that goes too far ;)
@@ -123,6 +125,15 @@ class CGaussianKernel: public CShiftInvariantKernel | |||
return std::exp(m_log_width * 2.0) * 2.0; | |||
} | |||
|
|||
#ifndef SWIG | |||
/** | |||
* Returns a lazily evaluated Eigen expression template |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If it is already evaluated, then it won't be lazily evaluated (at least not from the point of view of deferring evaluation after this method has been executed).
What about without the lazily evaluated part, or refactoring to "to be lazily evaluated" if you really want to keep that part.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yes, good point, i'll change that
// this could be written as | ||
// eigen_log_width.derivatives() = EigenScalar::Unit(n_differentiable_params, i); | ||
// where i is the idx of the adjoint | ||
m_eigen_log_width.derivatives() = Eigen::VectorXd::Unit(1,0); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is the derivative of the log width wrt the log width itself?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this the derivative of the kernel function wrt to the log with, df/dw. It gives the same result as the manual derivative that was in the loop before in get_parameter_gradient
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Then kernel function should be equal to the log width (for the derivative to be one), no?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
what do you mean? The derivative is set to one, so that when you apply the chain rule, you get it wrt to this parameter, i.e. you initialise the dependent variable. Like shown here https://eigen.tuxfamily.org/bz_attachmentbase/attachment.cgi?id=395. The same happens in stan gradient function: https://github.com/stan-dev/math/blob/b42294a57318a27d2b6723b96a0b620485ba83e0/stan/math/rev/core/grad.hpp#L39
Closing this in favour of work being done graphs |
@karlnapf I was testing out autodiff with stan for the Gaussian kernel and it seems to work fine.
The issue is that the forward pass isn't thread safe, so we might have to rethink a bit how kernels compute values.
Then we can have a virtual method that defines the kernel using arrays rather than scalars, which is then called from forward and backward functions. The linalg parallelism would be the completely handled by eigen and we don't worry about cache, etc...