search

Yelp / MOE

A global, black box optimization engine for real world metric optimization.
Other
1.31k stars 139 forks source link

Jialei gh368 cuda python interface #394

Closed jialeiwang closed 9 years ago

jialeiwang commented 10 years ago

***** PEOPLE ***** Primary reviewer: @suntzu86

Reviewers: @sc932

***** DESCRIPTION ** Branch Name: jialei_gh368_cuda_python_interface Ticket(s)/Issue(s): #368

***** TESTING DONE ***** make test

suntzu86 commented 10 years ago

couple of questions. if they don't require any changes then you have a shipit from me. This really makes me regret how I designed the high level calls for optimization! So many massive 50-argument functions :(

Please make a ticket for making the GPU components invokeable from the REST interface. That should be a pretty simple matter of adding a few more parameters.

congrats on getting all the GPU stuff put together! :D Really awesome work Jialei!

suntzu86 commented 10 years ago

left a few more brief comments; see email I sent you

jialeiwang commented 10 years ago

Why my comment does not show corresponding line of code as you did before :(

suntzu86 commented 9 years ago

ok, posted changes finally.

  1. your comments aren't showing up b/c you commented on an individual commit vs at the "global" level (the 'commits' tab at the top, which shows the combo of all commits). When you do that, github doesn't try to map your comment onto the global view. I'm guessing this is lines of an individual commit aren't guaranteed to be in the global (e.g., if they were deleted earlier)
  2. the in the screenshot is due to attempting to access the move ctor of CudaExpectedImprovementState(), which was defaulted. That ctor calls the move ctor of CudaDevicePointer which is implicitly deleted. I'm fixing this as well as changing CudaDevicePointer to be type-blind and having it internally use a unique_ptr (safer).
suntzu86 commented 9 years ago

compiles & runs now, but the GPU tests don't pass :(

jialeiwang commented 9 years ago

Before shipping, we may need to address the problem that GPU becomes uncallable after certain number of consecutive set/ release the GPU.

suntzu86 commented 9 years ago

thanks for looking over the changes.

the set/reset issue we've been talking about offline should be the topic of a new issue/pull request. this one is busy enough already :)

jialeiwang commented 9 years ago

Shipit! Sorry for the late reply!