Fix/nn transformer block #367
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
cxzhang4
commented
Mar 21, 2025
| ) | ||
| mlr3pipelines::mlr_pipeops$add("transformer_layer", PipeOpTorchTransformerLayer) | ||
|
|
||
| # TODO: remove deofault values from here |
There was a problem hiding this comment.
Since defaults should be handled by the PipeOp wrapper class
cxzhang4
commented
Mar 21, 2025
|
|
||
| # TODO: remove layer_idx and ask about how we want to handle this condition | ||
| # layer_idx = -1 | ||
| if (!is_first_layer || !prenormalization || first_prenormalization) { |
There was a problem hiding this comment.
Should match the official implementation now, but it is still confusing that !prenormalization is here...
cxzhang4
commented
Mar 21, 2025
| d_embedding = 32 | ||
|
|
||
| # # TODO: access x[, -1] first | ||
| # # TODO: sometimes there is no normalization, i.e. nn_identity instead of nn_layer_norm, figure out how to handle this |
There was a problem hiding this comment.
Could parameterize like the pre-implemented nn_ft_head, i.e. with an activation parameter
cxzhang4
commented
Mar 21, 2025
| self$last_layer_query_idx = param_vals$last_layer_query_idx | ||
| } | ||
| ), | ||
| private = list( |
There was a problem hiding this comment.
Confirm the ordering of the dimensions of the tensors. It may be the case that here, the first dimension is the sequence dimension (NOT the batch dimension)
#' @title Custom Function
#' @inherit torch::nnf_linear description
#' @section nn_module:
#' Calls [`torch::nn_linear()`] when trained where the parameter `in_features` is inferred as the second
#' to last dimension of the input tensor.
#' @section Parameters:
#' * `out_features` :: `integer(1)`\cr
#' The output features of the linear layer.
#' * `bias` :: `logical(1)`\cr
#' Whether to use a bias.
#' Default is `TRUE`.
#'
#' @templateVar id nn_linear
#' @template pipeop_torch_channels_default
#' @templateVar param_vals out_features = 10
#' @template pipeop_torch
#' @template pipeop_torch_example
#'
#'
#' @export
PipeOpTorchFn = R6Class("PipeOpTorchFn",
inherit = PipeOpTorch,
public = list(
#' @description Creates a new instance of this [R6][R6::R6Class] class.
#' @template params_pipelines
initialize = function(id = "nn_fn", param_vals = list()) {
param_set = ps(fn = p_uty(...))
super$initialize(
id = id,
param_set = param_set,
param_vals = param_vals,
module_generator = nn_linear
)
}
),
private = list(
.shapes_out = function(shapes_in, param_vals, task) {
# Implement this.
# 1. Generate a tensor of shape shapes_in (fill NA with something)
# 2. Apply function private$.f
# 3. Meausre shapes and fill dimensions with NA again
# Should also be possible to implement shapes_out properly
# Also take inspiration from pipeop_preproc_torch
},
.make_module = function(shapes_in, param_vals, task) {
self$param_set$values$fn
},
.fn = NULL
)
)
#' @include aaa.R
register_po("nn_fn", PipeOpTorchFn) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
A sketch of the FT-Transformer graph.