torch
torch copied to clipboard
mlverse
R Interface to Torch
``` # This returns only weight1 net$named_parameters(TRUE) #> $weight1 #> torch_tensor #> -2.2368 #> [ CPUFloatType{1,1} ][ requires_grad = TRUE ] # This returns weight1 and weight2 net$named_parameters(FALSE) ``` This...
``` # This does not work, because `$register_parameter()` is not called. # I would expect it to work though or at least throw a warning net$weight1 = torch_clone(net$weight1) + 1000...
After installing R torch loading the library (`library(torch)`) results in the following message requesting to install additional software even though all the dependencies (listed in `DESCRIPTION`) were already present ```...
Hi @dfalbel, While benchmarking some R packages for DL (torch is the fastest) we made a strange observation. The runtime of torch suddenly doubles at a certain network size (the...
This can speed up some operation for newer generations of GPUs: https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
The shape of e.g. the convolution operators depends on whether there is a batch dimension. Compare https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html with https://torch.mlverse.org/docs/reference/nn_conv1d.html?q=nn_conv1d#null
I just stumbled upon the the following surprising behaviour and I just wanted to report it. As far as I understand, `torch_tensor()` simply converts the NAs into their underlying machine...
It would be nice to have support for `vmap` on the R side. It seems that this is mainly implemented in LibTorch so we should be able to have that...
Hi, I'm using a fairly up-to-date Manjaro Linux. I've noticed that R's torch does not seem utilize my CPU (AMD Ryzen 7, series 5000) to its full extent. Further, using...
