Tensor tch functions

[Athospd]

tch functions implemented:

• tch_log
• tch_log2
• tch_log10
• ttch_log1p
• tch_tan
• tch_tanh
• tch_sin
• tch_sinh
• tch_std
• tch_var
• tch_min
• tch_max
• tch_prod
• tch_logsumexp
• tch_mode <- new
• tch_median <- new
• tch_tril <- new
• tch_triu <- new
• tch_rsqrt <- new
• tch_sigmoid <- new
• tch_sign <- new
• tch_sqrt <- new
• tch_trunc <- new
• tch_lerp <- new

Methods:

• rsqrt_
• sigmoid_
• sin_
• tan_
• sqrt_
• tanh_
• trunc_
• abs_
• acos_
• asin_

tch functions that I failed to implement:

• ~tch_mode~
• ~tch_median~

Notes:

• tch_logsumexp() seems to have some divergence between pytorch doc and c++ implementation. I'll discuss this in a issue.
• I kept the same structure of tch_mean() function for the major of these implementations, so they suffer of the same lack of other arguments code.

  // TODO handle other sum arguments.
  Rcpp::stop("Not yet implemented");

[dfalbel]

@Athospd I think it didn't work median and mode, because when you have a dim argument median will return 2 tensors: the medians for each row and the index of the median for each row.

In this case I think we will need 2 functions: 1 to calculate the median (without any additional args) and another to calculate the median with and return both tensors to R. tensor_qr_ has a reference implementation for this. Then we need to create some glue code on the R side.

[dfalbel]

@Athospd I created a reference implementation for tch_mean here: https://github.com/dfalbel/torch/blob/4bd94ae7db65f0dbf0308e1cca892eec359f13a0/src/tensor.cpp#L1037-L1049

We need to adapt other functions that have this kind of behavior to use this.

[Athospd]

@dfalbel I found out that the following functions has the tuple (dim, keepdim) args:

argmax argmin logsumexp median mode norm prod std sum var kthvalue max min

We have to adapt these using mean as reference, I got it right?

[dfalbel]

Exactly!

[dfalbel]

I will close this PR and merge #31 .