Adrian Sieradzki

Sounds like something I need especially since when introducing changes I can break learning for some use-cases and unit tests wont catch it - only benchmarks.

> ### Discussed in #1126 > > Originally posted by **pkese** October 28, 2023 If anyone is interested... > > I made a small language model inspired by https://github.com/karpathy/nanoGPT in...

Hi Kerod. Thanks for trying RL Matrix. I am in the process of making it a lot more polished and I'll be ditching binary serialiser for buffer saving. I never...

> Thanks a lot for the prompt response! I'm currently using **PPO**. Thats a shame. I'm a bit behind on this one. Let's see what I can do.

Just an update that I have commited an updated version. Its not on nuget yet For you this means massive performance improvements for both PPO & DQN. Also RNN now...

You're not supposed to put torchlib on user devices, its too massive. You pretty much have two options: 1. Export your models as ONNX and use that for inference. 2....

> > You're not supposed to put torchlib on user devices, its too massive. You pretty much have two options: > > > > 1. Export your models as ONNX...

> Great issue. What are the some of the situations where you need a module to have tensor members, but they're not buffers that become part of the state_dict? RNN...

Seconding this, I would like to keep scheduler and swap out optimizer without having to instantiate a new one. This sucks for model loading.

> As mentioned here [#1314 (comment)](https://github.com/dotnet/TorchSharp/issues/1314#issuecomment-2145274102), although the fields would be exposed, they may not be modifiable. Actually if I were the one to decide, I'd prefer to make them...