torch

torch copied to clipboard

``` r library(torch) blank_canvas [ CPUFloatType{2,3} ] blank_canvas[3:4, col_idx] # torch_tensor #> 0 0 0 #> 0 0 0 #> [ CPUFloatType{2,3} ] blank_canvas[3:4, 1:3] # torch_tensor #> 0 0...

Athospd

Building lantern with CUDA and existing libtorch

6

I'll split this in three sections: - allowing for existing libtorch to be used when building lantern - bug in pytorch cmake config - allowing for existing libtorch to be...

skeydan

Check hashes of downloaded libraries

2

One option to increase securiy of the package is to hardcode the hashes of the files downloaded in [`R/install.R`](https://github.com/mlverse/torch/blob/master/R/install.R) and check the hashes after downloading. This protects against corrupted files...

dirkschumacher

Test coverage

4

dfalbel

missing tensor methods

1

- [x] bool - [x] element_size - [ ] get_device - [ ] imag - [ ] is_shared - [ ] map_ - [ ] new_ones - [ ] new_tensor...

dfalbel

Indexing model parameters doesn't agree with PyTorch semantics

10

I'm trying to constrain my solution space to the positive orthant so I'm projecting negative parameters (from nn_embedding()) to 0 at each gradient step. I use: model$lhs_emb$weight$data()[model$lhs_emb$weight$data()

alexpeys

Add some CI infrastructure to test the examples. Shouldn't be at every commit, but perhaps dailies or weekly tests... Examples in vignettes/examples

dfalbel

add an `str` method for tensors

1

dfalbel

This allows converting between torch tensors and numpy arrays with no copy: ``` library(torch) x

dfalbel

Continuous benchmarking

1

Explore https://github.com/r-lib/bench#continuous-benchmarking to avoid performance regressions

dfalbel