Ethan Meitz
Ethan Meitz
@Jutho could you please push this through. Thanks
I believe cuTensor restricts me to GPUArrays 9 which has a memory double free issue when using multiple threads. I was hoping to update to 10 but I believe this...
Apparently this issue can cause crashes. To be clear this only happens when using TensorOperations (and more specifically GPUArrays.jl) inside of multiple separate threads. In my case I have one...
I'll test it out, thanks for making some changes! For the record here is the issue on GPUArrays: https://github.com/JuliaGPU/GPUArrays.jl/issues/503
Are there docs anywhere on incorporating a function made by `build_function` into a package? There's no version of `build_function` that targets a Julia output that have the `fname` parameter so...
Right now I call `build_function` with expr false and I save that to a file. I get something like : `function (ˍ₋arg1,) ... end` printed to my file. I just...
Yes, I believe we discussed this on Slack yesterday haha. That is what I plan on using for my implementation as it is only for a single model. Just thought...
I'll let you know if what I come up with is general enough. Flux should have all the same layer types & hyperparameters as PyTorch correct (with different names)?
Another thought for default optimization methods is the default optimizer in scipy least_squares. It is extremely robust in my experience. There's more knobs exposed to tune the optimizer than Optimziation.jl...
The optimizer is not though? The loss function is just LSQ by default.