dfalbel

R torch vs Python PyTorch: Benchmark Summary

Benchmark based on py_r_torch_benchmarking, related to mlverse/torch#1435.

Setup

• Machine: Apple Silicon (arm64), macOS Darwin 24.5.0
• R: 4.5.2
• Python: 3.13 (via uv)
• PyTorch: 2.8.0

dfalbel

# Benchmark: overhead of async+await vs the sync C++ path
#
# Both paths end up calling the same PJRT C API, but the sync path
# awaits in C++ while async+await does it through R. This measures the
# R-side overhead difference to decide if the dedicated sync path is worth keeping.
#
# Usage: Rscript bench/bench-async-vs-sync.R

library(pjrt)

dfalbel

This is a Guild AI runs repository. To access runs, install Guild AI and run guild pull gist:dfalbel/denoising. For more information about Guild AI Gist based repositories, see Guild AI - Gists.

dfalbel

This is a Guild AI runs repository. To access runs, install Guild AI and run guild pull gist:dfalbel/denoising-diffusion-runs. For more information about Guild AI Gist based repositories, see Guild AI - Gists.

dfalbel

library(torch)

dat <- dataset(
  "mydataset",
  initialize = function(time, size, len = 100 * 32) {
    self$time <- time
    self$len <- len
    self$size <- size
  },
  .getitem = function(i) {

dfalbel

library(keras)
library(tensorflow)

input <- layer_input(shape = list(365, 10))
representation <- input %>%
  layer_lstm(units = 32, input_shape = list(365, 10)) %>%
  layer_dropout(rate = 0.2)

output1 <- representation %>%
  layer_dense(units = 2, name = "out1")

dfalbel

library(torch)
library(ggplot2)

# we want to find the minimum of this function
# using the gradient descent.

f <- function(x) {
  x^2 - x
}

dfalbel

todoist_token <- config::get("TODOIST_TOKEN")

get_tasks_week <- function(week = 0, offset = 0) {
  
  res <- httr::POST(
    url = "https://api.todoist.com/sync/v8/activity/get",
    body = list(
      token = todoist_token,
      limit = 100,
      page = week,

dfalbel

dataset

Daniel Falbel 4/13/2019

Context

In tf.data in python the api for iterating over the elements of a dataset is the following:

dfalbel

---
title: "Quora Question Pairs"
output: 
  flexdashboard::flex_dashboard:
    orientation: columns
runtime: shiny    
---

```{r global, include=FALSE}
library(keras)