C++ frontend Tutorial #408
Conversation
|
Deploy preview for pytorch-tutorials-preview ready! Built with commit 0e665a6 https://deploy-preview-408--pytorch-tutorials-preview.netlify.com |
| b = register_parameter("b", torch::randn(M)); | ||
| } | ||
| torch::Tensor forward(torch::Tensor input) { | ||
| return torch::admm(b, input, W); |
| <https://arxiv.org/abs/1511.06434>`_ -- one of the first and simplest of its | ||
| kind, but entirely sufficient for this task. | ||
|
|
||
| What was a GAN aGAN? |
There was a problem hiding this comment.
Is it a typo or an intentional pun?
|
Here's a direct link to the preview for easier reading. https://deploy-preview-408--pytorch-tutorials-preview.netlify.com/advanced/cpp_frontend.html |
|
It looks reasonable. Do you want editing or to ship as is? |
|
looks pretty good and well-structured to me. thanks a lot Peter! :) |
| another_bias = register_parameter("b", torch::randn(M)); | ||
| } | ||
| torch::Tensor forward(torch::Tensor input) { | ||
| return linear(input) + another_bias; |
There was a problem hiding this comment.
A compile error occurred at this line when building with vs2017. The error disappeared after I changed linear(input) to linear->forward(input)
There was a problem hiding this comment.
there seems to be an issue here regarding the general syntax of the forward pass.
I suppose an overload of the call operator was intended here but as of now is not supported by neither the stable version nor the nightly build of libtorch. I find the overload highly desirable as it lends itself towards the python API and seems way more natural than having to call a forward function.
There was a problem hiding this comment.
I appreciate everyone trying out the examples, but there are some PRs that I am landing before this code works, which is why this tutorial is not yet published. The call operator is coming with pytorch/pytorch#15831 which I'll land right now, and then the example repository, and then this PR
apaszke
left a comment
apaszke
left a comment
There was a problem hiding this comment.
Looks great, but I have one question
| torch::load(generator_optimizer, "generator-optimizer-checkpoint.pt"); | ||
| torch::load(discriminator, "discriminator-checkpoint.pt"); | ||
| torch::load( | ||
| discriminator_optimizer, "discriminator-optimizer-checkpoint.pt"); |
There was a problem hiding this comment.
Just curious: how does C++ API deal with optimizers holding references to module parameters in this case? In Python one has to be careful to pickle the optimizer together with a module to keep the sharing intact, but in here you're not doing that.
There was a problem hiding this comment.
Tensor deserialization is in-place in the C++ serialization API, so the tensors to which the optimizers hold references should stay valid (as I see it)
* Wrote C++ frontend tutorial * Adding image and index entry * Update links to examples repo
8 participants
An end-to-end tutorial for the C++ frontend.
@soumith @ezyang maybe want to give it a quick glance?