In javierluraschi/kartmodels: kartmodels: Models for the 'kartsim' package.

This project makes use of kartsim to explore multiple training models.

Capturing Data

We will use the kartsim package and kartsim_capture() to capture training and test data to train the models in this repo.

1. Capture training data using:

library(kartsim)
kartsim_capture("capture/train", circuit = 3)

1. Followed by test data with:

kartsim_capture("capture/test", circuit = 3)

Training using the CIFAR model

We will first try modeling this as an image classification problem using the cifar10_cnn Keras example.

tfruns::training_run("models/tf-cifar.R")

or in cloudml runnning:

cloudml::cloudml_train("models/tf-cifar.R")

Next we will preload the model and predict over a graph object as follows:

sess <- tensorflow::tf$Session()
graph <- tfdeploy::load_savedmodel(sess, "savedmodel")

tfdeploy::predict_savedmodel(list(array(0, c(32,32,3))), graph, type = "graph", sess = sess)

   user  system elapsed 
  0.027   0.001   0.025

Which we can use to control the kart based on this model:

kartsim::kartsim_control(function(image, direction) {
  labels <- c("left", "forward", "right")
  input <- array(png::readPNG(image), c(32,32,3))
  result <- tfdeploy::predict_savedmodel(input, graph, type = "graph", sess = sess)
  scores <- result$predictions[[1]]$activation

  labels[which(scores == max(scores))][[1]]
}, circuit = 3)

517/517 [==============================] - 54s 105ms/step - loss: 6.4093 - acc: 0.6013 - val_loss: 5.4995 - val_acc: 0.6588
Epoch 2/2
517/517 [==============================] - 55s 107ms/step - loss: 6.3691 - acc: 0.6048 - val_loss: 5.5059 - val_acc: 0.6577

Improving the CIFAR model with context

In order to improve accuracy, we can consider making the model "remember" the state of the previous direction to help give continuity while steering.

tfruns::training_run("models/tf-cifar-prev.R")

library(tensorflow)
sess <- tensorflow::tf$Session()
graph <- tfdeploy::load_savedmodel(sess, "savedmodel/")

tfdeploy::predict_savedmodel(
  list(
    list(
      conv2d_1_input = array(0, c(32,32,3)),
      previous = 0
    )
  ),
  graph,
  type = "graph",
  sess = sess)

$predictions
            activation_6
1 0.0462, 0.9226, 0.0312

We can apply this control policy as follows:

previous <- 0
control_cifar_prev <- function(image, direction) {
  labels <- c("left", "forward", "right")
  input <- array(png::readPNG(image), c(32,32,3))

  result <- tfdeploy::predict_savedmodel(
    list(
      list(
        conv2d_1_input = input,
        previous = previous
      )
    ),
    graph, type = "graph", sess = sess)

  message(result)

  angle <- result$predictions$activation[[1]]
  previous <<- angle

  angle
}

kartsim::kartsim_control(control_cifar_prev)

javierluraschi/kartmodels documentation built on May 18, 2019, 8:07 p.m.