prediction of handwritten ✅ or ❌ boxes.
It is written using the golem framework for production-grade and robust shiny apps. The shiny app is embedded inside the structure of an R package which allows not only a concise and well structured workflow but also the integration of documentation (roxygen2) and tests (testthat). The prediction model is a convolutional neural network (CNN) implemented via the Keras API of Tensorflow.
The main objectives for me writing this code were to dive into the following packages/ frameworks:
- golem: framework for building production-grade shiny applications
- shinymodules: modularize shiny apps
- tensorflow/ keras: modeling deep neural networks
The following packages were also new to me:
- fullPage: fullPage.js framework for shiny
- waiter: waiting screens for shiny
- emo: emojis 😜
- googlesheets4: use googlesheets (as database)
Please note that the reposatory is no longer maintained and thus the code might be deprecated due to newer non backward compatible versions of the used dependencies. - I promise future projects will make use of
{renv}💡
You can install the current version of YesNoDetect with the following line of code:
# install.packages("devtools")
devtools::install_github("EmanuelSommer/YesNoDetect")You can run the app locally by calling the following:
YesNoDetect::run_app()Note: Of course for the local usage the functions in fct_db_handling.R
must be changed in such a way that they use your permissions file
for googlesheets which also has to be initiated.
- Database: Explore and amend the database which is the foundation of the modeling process.
- Model: The current model architecture, stats and valuation metrics.
- Predict: Challenge the model with your handwritten boxes!
To model the handdrawn boxes I obviously needed training data. To get it I did set up this section. As database I chose a googlesheets file and implemented functions to write and read it and supplied the permissions needed. The boxes are saved as 28x28 black and white pictures (as a flattened 28x28 matrix).
The first of the two slides of this section does give a short overview of the database:
- absolute number of boxes (yes, no, total)
- plot displaying the proportion of yes vs no boxes
- plots of the average yes and no box
The second of the two slides of this section allows the user to amend the database:
- draw the box
- label it
- save it
In this section one can find the most important information about the
current model used. After reading several papers to get a good grasp of
some of the state of the art architectures of CNNs for image recognition
I implemented 3 CNNs in the Rmarkdown document in the folder modeling.
These were trained and tested on independent test sets and then the best
model is chosen and saved for usage in the app. For a closer look at the
modeling process just read through the Rmarkdown document. It also
contains code to visually inspect wrong predictions and more.
The first slide of this section contains a visualization of the archictecture of the current model.
The second slide of this section shows essential core and valuation stats of the current model.
In this last section one can challenge the current model to predict ones handwritten ✅ or ❌ boxes and recieves a probability for each class.
The drawing is not done by clicking, holding and releasing the mouse but by click, hover and click again. This is not the intuitive way. As the drawing was implemented by an interactive plot using the hover and click parameters I could not find a way to change to the intuitive way of drawing, at least not in a reasonable searching period. Moreover I did not find any other (no interactive plots) solutions for handdrawing in shiny apps and efficiently saving the drawing numerically. I guess some Javascript via the shinyjs package could do the trick.
The app is not touch friendly this does not only apply to the drawing of the boxes but the whole fullPage UI seems not to work on mobile well at all. To deal with this one could switch to shinyMobile but I will try this out another time.
If you have any suggestions to solve the known issues or you have new ones I would really like to hear from you!
👋 Have fun! 👋