In botchkoAI/vertaReticulateClient: Reticulated-based client to the Verta AI Platform

This is closely following this jupyter notebook. The point is to go trough several of the functionalities of VERTA modeldb.

Setup

library(reticulate)
library(data.table)
library(magrittr)
library(stringr)
library(yardstick)
# use_condaenv("verta_reticulate",required=T)
# library(vertaReticulateClient)

reg_random_string <- rpois(1,1000)

HOST = "dev.verta.ai"
PROJECT_NAME <-  str_glue("MTCARS Regression{reg_random_string}")
EXPERIMENT_NAME <-  "Linear Regressions"
WORKSPACE <-  "Stefan_Petrov"
# Sys.setenv(VERTA_EMAIL= "...",
#            VERTA_DEV_KEY = "...")

Sys.setenv(VERTA_EMAIL= "...",
          VERTA_DEV_KEY = "...")

Initialize Client

library(vertaReticulateClient)
vertaReticulateClient::init_verta(HOST)
proj <- vertaReticulateClient::set_project(name="Reticulate MTCars")
experiment <- vertaReticulateClient::set_experiment(name="Verta R Experiment")

Use mtcars for the experiment

N <- nrow(mtcars)
trainSet <- mtcars[1:floor(N/2),]
testSet <- mtcars[(floor(N/2)+1):N,]

Create 2 registered models to compate

experiment_run_name <- str_glue("MTCars Run {rpois(1,1000)}")
run <- set_experiment_run(name=experiment_run_name,desc = experiment_run_name)
log_run_hyperparameters(run,hyperparams = list(hyperparam1=1,hyperparam2=2),overwrite = T)
log_run_hyperparameters(run,hyperparams = list(formula="mpg~."),overwrite = T)

mod_mtcars <- lm(mpg~.,data = trainSet)

summary_lm <- summary(mod_mtcars)
rsq <- summary_lm$adj.r.squared
# run_log_metric(run,key = "rsquared",value = rsq)

run2 <- set_experiment_run(name=str_glue("MTCars Experiment 3 {rpois(1,10000)}"),
                           desc = "MTCars Experiment 3")
log_run_hyperparameters(run2,hyperparams = list( formula=
                                                 "mpg~cyl+disp+hp+drat+wt+qsec+vs"
                                               ),overwrite = T)
mod_mtcars_2 <- lm(mpg~cyl+disp+hp+drat+wt+qsec+vs,data = trainSet)
summary_lm_2 <- summary(mod_mtcars_2)
rsq_2 <- summary_lm$adj.r.squared
run_log_metric(run2,key = "rsquared",value = rsq_2,overwrite = T)

So we created some 2 experiment runs and added some metrics for them...

Log Model and Upload Artifacts

first_model_object <- list(
  model_file = mod_mtcars,
  formula_obj = "mpg~.",
  original_df = trainSet,
  required_packages="",
  additionalObjects=list()
)

saveRDS(first_model_object,"mpg_experiment_object.RData")
run$log_model(model = "mpg_experiment_object.RData",overwrite = T)# TODO- add
# run_log

The downloaded model is almost what's needed. From the web UI it works fine.

# there is some issue with this
# run$download_model("downloaded_model.RData")
run$log_model(model = "mpg_experiment_object.RData",overwrite = T)# TODO- add
run$download_model("./dl_model/download_model.RData")
# run$download_model("./dl_model/")

downloaded_model <- readRDS("./dl_model/download_model.RData")# this
downloaded_model # this is almost what it should be, but not quite...
downloaded_model_lm <- downloaded_model[["model_file"]]
downloaded_model_lm

We'll use downloaded_model to build a docker service

lm_model_from_download <- downloaded_model$model_file

Find the best run and it's hyperparameters

best_run <- get_best_run_by_metric(experiment,
                                   "metrics.rsquared",descending = T)
best_run_hyperparameters <- best_run$get_hyperparameters()

Register the Best Model as in the jupyter

registered_model_best_mtcars <- get_or_create_registered_model(name="mtcars_best",
                                                               workspace = WORKSPACE)
registered_model_best_mtcars$create_version_from_run(best_run$id,name=str_glue("best_model_{rpois(1,10000)}"))

Model Deployment

Here we simply use the downloaded_model in order to deploy a service at localhost:8000

runner_template_location <- file.path(path.package("vertaReticulateClient"),
"to_run.R"
)

plumber_template_location <- file.path(path.package("vertaReticulateClient"),
                                         "plumber_sample.R"
)


# createDockerfile(
#   model_file_location = "../mpg_model.RData",
#   # model_file_location = downloaded_model, 
#   runner_template_location = runner_template_location,
#   plumber_template_location = plumber_template_location,
#   required_packages = "forecast"
#   )


# log_model_data(best_run,file_to_save_at = "../mpg_model.RData",
#                 downloaded_model,
#                 required_packages = c()
#                )

log_model_data(best_run,modelObject = downloaded_model,required_packages = c())

best_run$download_model("mpg_experiment_best_model.RData")

createDockerContextZip("mpg_experiment_best_model.RData")

The above docker context could be build

buildDocker(location = "docker_folder_YCEQR4149N/",tag = "verta-mtcars:latest")

Run the client with: ```{bash,eval=FALSE} docker kill $(docker container ls -q) docker run -t 8000:8000 verta-mtcars:latest

```

now test w/ the following reques request:

curl -X POST --location "http://localhost:8000/score"\ -H "Accept: application/json"\ -d "{"cyl":8,"disp":440,"hp":230,"drat":3.83,"wt":5.345,"qsec":17.42,"vs":0,"am":0,"gear":3,"carb":4,"_row":"Chrysler Imperial"}"

---

Or

POST http://localhost:8000/score Accept: application/json

{"cyl":8,"disp":440,"hp":230,"drat":3.83,"wt":5.345,"qsec":17.42,"vs":0,"am":0,"gear":3,"carb":4,"_row":"Chrysler Imperial"}

botchkoAI/vertaReticulateClient documentation built on Dec. 19, 2021, 10:50 a.m.