Comments on The Beginner Programmer: Image recognition tutorial in R using deep convolutional neural networks (MXNet package)

Thank you for your insights and information. Prepa...

2025-11-20T05:43:04.711+00:00

Thank you for your insights and information. Prepare for Data Science Jobs in Chennai
with 360DigiTMG. Gain hands-on experience in Python, R, ML, AI, and analytics through live projects and mentorship. Placement support equips learners with industry-ready skills to secure competitive roles in top Chennai-based companies.

Prepare for Data Science Jobs in Chennai with 360D...

2025-11-20T05:08:46.721+00:00

This comment has been removed by the author.

Good information Thanks for sharing 360DigiTMG hel...

2025-09-24T12:05:21.986+01:00

Good information Thanks for sharing 360DigiTMG helps learners prepare for Data Science Fresher Jobs in Bangalore. Gain hands-on experience in Python, R, ML, AI, and analytics through projects, mentorship, and placement guidance to become industry-ready.data science fresher jobs bangalore

Learn from 360DigiTMG’s Data Science and Artificia...

2025-09-16T09:38:30.519+01:00

Learn from 360DigiTMG’s Data Science and Artificial Intelligence Online Courses, designed for flexible learning. Covering Python, R, AI, ML, and analytics, these courses offer projects, mentorship, and placement support. Ideal for professionals and students, they provide practical exposure to data and AI tools from anywhere, anytime.

data science and artificial intelligence online courses

Really awesome blog. Very interesting and knowledg...

2021-11-27T18:49:19.043+00:00

Really awesome blog. Very interesting and knowledgeable content. Keep sharing more blogs with us.
Data Science Training in Hyderabad

Extremely overall quite fascinating post. I was se...

2021-11-16T06:10:46.863+00:00

Extremely overall quite fascinating post. I was searching for this sort of data and delighted in perusing this one. Continue posting. A debt of gratitude is in order for sharing. iot course in lucknow

I have read your excellent post. This is a great j...

2021-07-12T06:59:23.914+01:00

I have read your excellent post. This is a great job. I have enjoyed reading your post first time. I want to say thanks for this post. Thank you...
digital marketing courses in hyderabad with placement

We are really grateful for your blog post. You wil...

2021-05-17T03:58:59.549+01:00

We are really grateful for your blog post. You will find a lot of approaches after visiting your post. Great work

Best Data Science courses in Hyderabad

Amazon.com mytv - you merely unboxed a fresh Amaz...

2021-03-26T11:09:14.794+00:00

Amazon.com mytv - you merely unboxed a fresh Amazon device and are excited to research -- just how can you begin? Discover how to connect your devices to Prime so that you can easily see and listen to exclusive Prime manhood content from anywhere. You simply have to make amazon accounts and activate it using amazon my television activation code.

This blog resolved all my queries I had in my mind...

2021-03-15T16:58:19.802+00:00

This blog resolved all my queries I had in my mind. Really helpful and supportive subject matter written in all the points. Hard to find such kind of blogs as descriptive and accountable to your doubts.This blog resolved all my queries I had in my mind. Really helpful and supportive subject matter written in all the points. Hard to find such kind of blogs as descriptive and accountable to your doubts.How to get rid of Discord Javascript Error [100% SOLVE ISSUE]

Thanks for sharing this super helpful guide. I rea...

2020-12-22T03:00:16.474+00:00

Thanks for sharing this super helpful guide. I really appreciate your effort.
Learn How to Resize an Image in Pixlr

Great Article Artificial Intelligence Projects ...

2020-08-11T06:28:56.367+01:00

Great Article
Artificial Intelligence Projects

Project Center in Chennai

JavaScript Training in Chennai

JavaScript Training in Chennai

Your blog is very informative and interactive. Tha...

2019-11-25T06:22:26.624+00:00

Your blog is very informative and interactive. Thanks for sharing useful information.R Programming training in bangalore. Keep updating valuable content.

Thanks for sharing this valuable information and w...

2019-09-28T10:07:51.015+01:00

Thanks for sharing this valuable information and we collected some information from this blog.

Machine learning in-house Corporate training in Nigeria

why in your predicted_labels, you have to add -1 t...

2019-02-27T05:28:29.999+00:00

why in your predicted_labels, you have to add -1 to the label number. Is it possible that somewhere you messed up with the labels?

Hi there, Thanks for your post. It's a really...

2019-01-06T20:28:49.919+00:00

Hi there,

Thanks for your post. It's a really nice easy to follow tutorial for users like myself to get into this. Just a few things that I wanted to discuss that others might struggle with:

1. The installation of Mxnet varies drastically depending on where you look across the internet. Your link on how to build unfortunately doesn't exist anymore. The simplest option I found was on their page (see below) where you input your platform and it gives you the required code to install Mxnet.

https://mxnet.incubator.apache.org/install/index.html

2. As what was mentioned in the comments, the way that the testing sample is chosen highly influences the performance of the model as favourable testing samples can make the model appear better than what it is.

I was not able to reproduce your results unfortunately. The way random number generation is done from seeding appears to be dependent on the operating system (see https://stackoverflow.com/questions/48626086/same-seed-different-os-different-random-numbers-in-r)

For those who are looking for someone to compare against. I am running a Windows 10 machine (update 1803), Python version 3.7 and R version 3.5.0 (2018-04-23) -- "Joy in Playing" and using the seeding in this article I achieved a 35% test accuracy. Without setting the seeding, my test accuracy varied (depending on the seed chosen by the operating system) between 17.5% and 100%.

3. Another thought that I thought should be mentioned is that Neural nets are notorious for needing huge amounts of training data to acquire good results. I know that this isn't the aim of the tutorial, but perhaps it might be misleading for those starting out to expect a 97.5% accuracy every time when given such a small training set.

Thanks again for the tutorial!
Chris

Here's my custom callback function I used abov...

2018-04-19T10:47:11.361+01:00

Here's my custom callback function I used above to track validation error and to use early stopping.

mx.callback.custom_early.stop_log.train.metric <-
function (period,
logger = NULL,
train.metric = NULL,
eval.metric = NULL,
bad.steps = NULL,
maximize = FALSE,
verbose = FALSE)
{
function(iteration, nbatch, env, verbose = verbose) {
if (!is.null(env$metric)) {
if (!is.null(train.metric)) {
result <- env$metric$get(env$train.metric)
if ((maximize == F & result$value < train.metric) |
(maximize == TRUE & result$value > train.metric)) {
return(FALSE)
}
}
if (!is.null(eval.metric)) {
if (!is.null(env$eval.metric)) {
result <- env$metric$get(env$eval.metric)
if ((maximize == F & result$value < eval.metric) |
(maximize == TRUE & result$value > eval.metric)) {
return(FALSE)
}
}
}
}

## @TG log train + eval error
## start insertion
if (nbatch %% period == 0 && !is.null(env$metric)) {
result <- env$metric$get(env$train.metric)
if (nbatch != 0 & verbose)
message(paste0(
"Batch [",
nbatch,
"] Train-",
result$name,
"=",
result$value
))
if (!is.null(logger)) {
if (class(logger) != "mx.metric.logger") {
stop("Invalid mx.metric.logger.")
}
logger$train <- c(logger$train, result$value)
if (!is.null(env$eval.metric)) {
result <- env$metric$get(env$eval.metric)
if (nbatch != 0 & verbose)
message(paste0(
"Batch [",
nbatch,
"] Validation-",
result$name,
"=",
result$value
))
logger$eval <- c(logger$eval, result$value)
}
}
}
## end insertion

if (!is.null(bad.steps)) {
if (iteration == 1) {
mx.best.iter <<- 1
if (maximize) {
mx.best.score <<- 0
}
else {
mx.best.score <<- Inf
}
}
if (!is.null(env$eval.metric)) {
result <- env$metric$get(env$eval.metric)

# @TG stop if validation error is NaN
if(is.na(result$value)) {
return(FALSE)
}
#

if ((maximize == F & result$value > mx.best.score) |
(maximize == TRUE & result$value <= mx.best.score)) {
# (maximize == TRUE & result$value < mx.best.score)) { # TG, 19.04.2018
if (mx.best.iter == bad.steps) {
if (verbose) {
message(
paste0(
"Best score=",
mx.best.score,
", iteration [",
iteration - bad.steps,
"]"
)
)
}
return(FALSE)
}
else {
mx.best.iter <<- mx.best.iter + 1
}
}
else {
mx.best.score <<- result$value
mx.best.iter <<- 1
}
}
}
return(TRUE)
}
}

"Building the model" with batch normaliz...

2018-04-19T10:43:34.724+01:00

"Building the model" with batch normalization: After 6-8 iterations, test accuracy is same as above (97.5%). Uses an own implementation of early stopping (see comment below!)

#'
#' with batch normalization
#' with early stopping
#'

# Clean workspace
rm(list=ls())

# Load MXNet
require(mxnet)

# load custom loss function (early stopping)
setwd("/gpudev/projects/E610046/sandbox/cnn_example1/fun/early_stopping.R")

# Loading data and set up
#-------------------------------------------------------------------------------

# Load train and test datasets
load("data/test_64.RData")
load("data/train_64.RData")
train <- train_64; rm(train_64)
test <- test_64; rm(test_64)

# Set up train and test datasets
train <- data.matrix(train)
train_x <- t(train[, -1])
train_y <- train[, 1]
train_array <- train_x
dim(train_array) <- c(64, 64, 1, ncol(train_x))

test_x <- t(test[, -1])
test_y <- test[, 1]
test_array <- test_x
dim(test_array) <- c(64, 64, 1, ncol(test_x))

# Set up the symbolic model
#-------------------------------------------------------------------------------

data <- mx.symbol.Variable('data')
# 1st convolutional layer
conv_1 <- mx.symbol.Convolution(data = data, kernel = c(5, 5), num_filter = 1)
tanh_1 <- mx.symbol.Activation(data = conv_1, act_type = "tanh")
pool_1 <- mx.symbol.Pooling(data = tanh_1, pool_type = "max", kernel = c(2, 2), stride = c(2, 2))
# 2nd convolutional layer
bn_1 <- mx.symbol.BatchNorm(pool_1)
conv_2 <- mx.symbol.Convolution(data = bn_1, kernel = c(5, 5), num_filter = 50)
tanh_2 <- mx.symbol.Activation(data = conv_2, act_type = "tanh")
pool_2 <- mx.symbol.Pooling(data=tanh_2, pool_type = "max", kernel = c(2, 2), stride = c(2, 2))
# 1st fully connected layer
bn_2 <- mx.symbol.BatchNorm(pool_2)
flatten <- mx.symbol.Flatten(data = bn_2)
fc_1 <- mx.symbol.FullyConnected(data = flatten, num_hidden = 500)
tanh_3 <- mx.symbol.Activation(data = fc_1, act_type = "tanh")
# 2nd fully connected layer
bn_3 <- mx.symbol.Flatten(data = tanh_3)
fc_2 <- mx.symbol.FullyConnected(data = bn_3, num_hidden = 40)
# Output. Softmax output since we'd like to get some probabilities.
NN_model <- mx.symbol.SoftmaxOutput(data = fc_2)

# Pre-training set up
#-------------------------------------------------------------------------------

# Set seed for reproducibility
mx.set.seed(100)

# Device used. CPU in my case.
devices <- mx.cpu()

# Training
#-------------------------------------------------------------------------------
logger <- mx.metric.logger$new()
# iterator for eval error
valIter <-
mx.io.arrayiter(
data = test_array,
label = test_y,
batch.size = 40,
shuffle = F
)

# Train the model
model <- mx.model.FeedForward.create(
NN_model,
X = train_array,
eval.data = valIter,
y = train_y,
ctx = devices,
num.round = 130,
array.batch.size = 10,
learning.rate = 0.01,
momentum = 0.9,
eval.metric = mx.metric.accuracy,
epoch.end.callback = mx.callback.custom_early.stop_log.train.metric(
period = 40,
bad.steps = 3,
logger,
verbose = T,
maximize = T
)
)

# Testing
#-------------------------------------------------------------------------------

# Predict labels
predicted <- predict(model, test_array)
# Assign labels
predicted_labels <- max.col(t(predicted)) - 1
# Get accuracy
sum(diag(table(test_y, predicted_labels)))/40

################################################################################
# OUTPUT
################################################################################
#
# 0.975
#

# Author: TG

I had issues with the EBImage library. Hence I did...

2018-04-19T10:39:12.688+01:00

I had issues with the EBImage library. Hence I did the step "Some more preprocessing with R" using this. Attention to Changes: (1) No downscaling to 28x28. I keep the full resolution of 64 x 64 and at the end. (2) At the end a RData file instead of .CSV is created.

#
# This code is from https://firsttimeprogrammer.blogspot.de/2016/08/image-recognition-tutorial-in-r-using.html
#

# This script is used to resize images from 64x64 to 28x28 pixels

# Clear workspace
rm(list=ls())
setwd("/gpudev/projects/E610046/sandbox/cnn_example1")

# Load data
X <- read.csv("data/olivetti_X.csv", header = F)
labels <- read.csv("data/olivetti_y.csv", header = F)

# Dataframe of resized images
rs_df <- data.frame()

# Main loop: for each image, resize and set it to greyscale
rs_df <- data.frame(labels,X)
rs_df[1:10,1:10]
colnames(rs_df) <- c("labels",paste0("pixel",1:ncol(X)))

# Train-test split
#-------------------------------------------------------------------------------
# Simple train-test split. No crossvalidation is done in this tutorial.

# Set seed for reproducibility purposes
set.seed(100)

# Shuffled df
shuffled <- rs_df[sample(1:nrow(rs_df)),]

# Train-test split
train_64 <- shuffled[1:360, ]
test_64 <- shuffled[361:400, ]

# Save train-test datasets
save(test_64, file="data/test_64.RData")
save(train_64, file="data/train_64.RData")

# Done!
print("Done!")

Hi, the labels are stored into the tran_y and test...

2017-12-02T21:56:31.717+00:00

Hi, the labels are stored into the tran_y and test_y variables.

I cannot see there the dependent variable is store...

2017-12-02T10:52:42.193+00:00

I cannot see there the dependent variable is stored? Somehow the information needs to be transmitted which picture blongs to which person. Where is that done?

Hi Dimri, thanks for reading the post! Sorry for m...

2017-06-30T21:39:26.830+01:00

Hi Dimri, thanks for reading the post! Sorry for my late reply. I had a bit of a struggle too when installing Mxnet.

Yes changing the seed should (generally) give you roughly similar accuracy results. But it may also not: it can also happen that a particular combination of test sample happened to contain particularly critical tests for the model, for instance you trained on a training set of 5 different pictures and the classifier always struggles with picture 1 (let's say it always gets 20% accuracy on it). By chance and by splitting randomly the testing set you got a testing set of 90% picture 1 samples. Now it may easily look that your classifier has an accuracy of around 20% while instead that result is simply driven by a biased testing set.

In order to partly account for this problem and to get a better estimate of how good your model is performing, a cross validation may help testing that scenario.

Hi, I was able to fix "Mxnet" issue ment...

2017-06-30T21:03:26.904+01:00

Hi, I was able to fix "Mxnet" issue mentioned in my previous comment.

It ran perfectly and gave similar results. However, when I changed the seed from 100 to some other number (Just before shuffling itno training and testing), I get an accuracy of 0.325.

Isn't it supposed to be somewhat similar of model is working fine ?
Am I missing on something ?

Hi, First of all thanks for this post. Very infor...

2017-06-28T16:32:59.962+01:00

Hi,

First of all thanks for this post. Very informative.
I am struggling to install "mxnet" package in my R.

Could you please guide what needs to be done in order to use it.

Hi Lenchik! Thanks for reading the post! I didn...

2017-03-07T18:49:34.271+00:00

Hi Lenchik! Thanks for reading the post! I didn't know it was already available in R, thanks for letting me know.