missing batch_size error inconsistent with torch docs, batch_sampler arg #421
Description
?torch::dataloader says
batch_sampler: (Sampler, optional): like sampler, but returns a batch
of indices at a time. Mutually exclusive with ‘batch_size’,
‘shuffle’, sampler, and ‘drop_last’. Custom samplers can be
created with sampler().
So batch_size should not be provided at the same time at batch_sampler.
However that contradicts the error message I get from mlr3torch when I run the code below:
batch_sampler_class <- torch::sampler(
"BatchSampler",
initialize = function(data_source) {
self$data_source <- data_source
},
.iter = function() {
batch_size <- 2
indices <- 1:self$.length()
batch_vec <- (indices-1) %/% batch_size
batch_list <- rev(split(indices, batch_vec))
count <- 0L
function() {
if (count < length(batch_list)) {
count <<- count + 1L
return(batch_list[[count]])
}
coro::exhausted()
}
},
.length = function() {
length(self$data_source)
}
)
sonar_task <- mlr3::tsk("sonar")
sonar_ingress <- list(feat=mlr3torch::TorchIngressToken(
features=sonar_task$col_roles$feature,
batchgetter=mlr3torch::batchgetter_num))
target_batchgetter <- function(data){
browser()
torch::torch_tensor()
}
sonar_dataset <- mlr3torch::task_dataset(sonar_task, sonar_ingress, target_batchgetter)
batch_sampler_instance <- batch_sampler_class(sonar_dataset)
inst_learner <- mlr3torch::LearnerTorchMLP$new(task_type="classif")
inst_learner$param_set$set_values(
epochs=10,
batch_sampler=batch_sampler_class)
inst_learner$train(sonar_task)In the code above I specified batch_sampler but I get the error message below:
> inst_learner$train(sonar_task)
Erreur dans .__ParamSet__get_values(self = self, private = private, super = super, :
Missing required parameters: batch_sizeThe error message says I need to add batch_size parameter, but that contradicts the documentation in torch. @sebffischer Can this error message be removed please? It would be much less confusing for users if error messages in mlr3torch were consistent with the documented behavior in torch.