Fix GPU's gradient!
#74
Conversation
|
Works on GPU too. |
|
I will test the last commit on a GPU tomorrow. edit: It works. |
| training_minibatch_iterator = create_minibatch(xtrn, ytrn, size_minibatch) | ||
| test_minibatch_iterator = create_minibatch(xtst, ytst, size_minibatch) | ||
| current_training_minibatch = rand(training_minibatch_iterator) | ||
| current_test_minibatch = rand(test_minibatch_iterator) | ||
|
|
||
| nested_array = build_nested_array_from_vec(chain_ANN, x0) | ||
| layers_g = similar(params(chain_ANN)) # create a Vector of layer variables | ||
| layers_g = similar.(params(chain_ANN)) # create a Vector of layer variables |
There was a problem hiding this comment.
why is param here lower case and in the other src/KnetNLPModels_methods.jl it is Param
There was a problem hiding this comment.
One thing I am not sure is how to test on GPUin unit test as of now, so I wanted to double check if it is possible
There was a problem hiding this comment.
why is param here lower case and in the other src/KnetNLPModels_methods.jl it is Param
params is a method and not the type.
The method return a type Param if needed.
One thing I am not sure is how to test on GPUin unit test as of now, so I wanted to double check if it is possible
I pull on the GPU the branch I want to test, and I run manually the tests on the GPU.
Which is nice with KnetNLPModels.jl is that the code works for both CPU and GPU, so i just run ] test once julia in loaded on Atlas.
There was a problem hiding this comment.
so the test run both GPU and CPU then
If you want to test on GPU, you must have access to GPU, this is achieved by the job buildkite/knetnlpmodels-dot-jl. |
|
Can we open a bug here and put it conditional on knet.jl update? |
|
And I think you mentioned this buildkite/knetnlpmodels-dot-jl fails, |
I don't understand what you mean. I prefer keeping buildkite/knetnlpmodels-dot-jl even if it fails, because Dominique had to ask for us to get access to GPU from buildkite's community. |
|
Sorry for confusion Sure if we already expect it to fail and have it in the github issue then I am okay with it I will pass approve it then, just don''t forget the issue |
No description provided.