Indexing model parameters doesn't agree with PyTorch semantics

[alexpeys]

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()<0] <- 0

and get

Error in model$lhs_emb$weight$data()[model$lhs_emb$weight$data() < 0] <- 0 : invalid (NULL) left side of assignment

In PyTorch this is totally legal.

Is this an R limitation or fixable?

[dfalbel]

This works in theory:

library(torch)
x <- torch_tensor(c(1,2,3))
x[x < 2] <- 0
x
#> torch_tensor
#>  0
#>  2
#>  3
#> [ CPUFloatType{3} ]

^{Created on 2020-10-19 by the reprex package (v0.3.0)}

I wonder if model$lhs_emb$weight is really a tensor in your module. Can you provide a reproducible example?

[dfalbel]

Actually it seems to be a bug with using the $data() method. Here's an equivalent implementation:

library(torch)
m <- nn_embedding(num_embeddings = 100, embedding_dim = 50)
with_no_grad({
  m$weight[m$weight < 0] <- 0  
})
m$weight$min()
#> torch_tensor
#> 0
#> [ CPUFloatType{} ]

^{Created on 2020-10-19 by the reprex package (v0.3.0)}

Minimal reprex for the error:

library(torch)
m <- nn_embedding(num_embeddings = 100, embedding_dim = 50)
m$weight$data()[m$weight$data() < 0] <- 0  
#> Error in m$weight$data()[m$weight$data() < 0] <- 0: invalid (NULL) left side of assignment

^{Created on 2020-10-19 by the reprex package (v0.3.0)}

[alexpeys]

Here is the class:

require(torch)
embedder <- nn_module(
  clasname = "embedder",
  initialize = function(n_lhs, n_rhs, dim=10) {
    self$lhs_emb <- nn_embedding(n_lhs, dim)
    self$lhs_bias <- nn_embedding(n_lhs, 1)
    
    self$rhs_emb <- nn_embedding(n_rhs, dim)
    self$rhs_bias <- nn_embedding(n_rhs, 1)
    
  },
  forward = function(lhs_indexes, rhs_indexes) {
    scores <- torch_sum(self$lhs_emb(lhs_indexes) * self$rhs_emb(rhs_indexes), 2)
    scores <- scores + torch_reshape(self$lhs_bias(lhs_indexes), list(-1))
    scores <- scores + torch_reshape(self$rhs_bias(rhs_indexes), list(-1))
    return(scores)
  }
)

model <- embedder(n_lhs=3, n_rhs=3, dim=2)

Now running

model$lhs_emb$weight$data()

Yields

torch_tensor -0.9773 -0.5023 0.6282 -1.4819 -1.0336 1.5777 [ CPUFloatType{3,2} ]

Running

model$lhs_emb$weight$data()[model$lhs_emb$weight$data()<0] <- 0

Yields

Error in model$lhs_emb$weight$data()[model$lhs_emb$weight$data() < 0] <- 0 : invalid (NULL) left side of assignment

Running

[alexpeys]

with_no_grad({
  m$weight[m$weight < 0] <- 0  
})

Works except bugs out if you try to use an empty slice, which is another thing PyTorch allows (precisely for ease of this kind of projection).

[dfalbel]

Ok! This should be an easy fix, I'll do it as soon as possible.

[alexpeys]

Thanks!

[dfalbel]

Wait, that doesn't seem to error for me:

library(torch)
m <- nn_embedding(num_embeddings = 100, embedding_dim = 50)
m$weight$min()
#> torch_tensor
#> -3.66959
#> [ CPUFloatType{} ]
with_no_grad({
  m$weight[m$weight < -100] <- 0    
})
m$weight$min()
#> torch_tensor
#> -3.66959
#> [ CPUFloatType{} ]

^{Created on 2020-10-19 by the reprex package (v0.3.0)}

Are you sure this is the cause? What exactly is an empty slice?

[alexpeys]

Oops, had a typo, was a non-existent rather than empty slice.

[dfalbel]

Ok, so IIUC the with_no_grad solution is working as expected?

[alexpeys]

Ok, so IIUC the with_no_grad solution is working as expected?

Yes, though would be good to also have the data() method correct since this is copy-pasted PyTorch code and it would be great for that to just work after swapping . to $.