dlab-berkeley / r-deep-learning Goto Github PK

Add batch-norm to content and slides

Myself and other participants in the workshop have issues when fitting models using the image_data_generator() function in notebook 2 & 3.

The issues I and other have gotten so far include:

1. "Error in py_call_impl(callable, dots$args, dots$keywords) : ImportError: Could not import PIL.Image. The use of load_img requires PIL."
2. "Error in py_call_impl(callable, dots$args, dots$keywords) : ImportError: Could not import PIL.Image. The use of load_img requires PIL."

# For R-Deep-Learning package missing issues when using datagen

# Step 1: Figure out what package is missing, for example: 
#"Error in py_call_impl(callable, dots$args, dots$keywords) : ImportError: Could not import PIL.Image. The use of `load_img` requires PIL."  
# Issue is no python library PIL

# Step 2: Install the library using conda_install() from the reticulate package
library(reticulate)

# install SciPy
conda_install("r-reticulate", "scipy")
# install PIL
conda_install("r-reticulate", "PIL") # replace PIL with whatever python package is missing


# If conda_install() fails try:
py_install("PACKAGE_NAME_HERE")

# Step 3: Restart your R-session and try the code that broke again!

This worked well for most participants.

Add auto-image downloader section to learn how to apply deep learning to image classes downloaded from Google.

A markdown file will be a lot easier to read than txt on github, will allow lists, etc.

Workshop title should be "R-Introduction-to-Deep-Learning:-Parts-1-2"

Where to set seed in fit_generator?

Just ran into this today (via https://www.coursera.org/learn/introduction-tensorflow/home/welcome), but if the loss function is sparse categorical crossentropy then the y_label doesn't need to be one-hot encoded:

https://keras.rstudio.com/articles/tutorial_basic_classification.html#compile-the-model

So this would provide a minor streamlining benefit to the part 1 preprocessing.

Add Auto-Keras analysis

Datasets look good except in notebook 2, where the people in the human image training and validation datasets look maybe ~90%+ white (a classic problem with a lot of this research: https://www.aaihs.org/race-after-technology/).

I didn't feel quite comfortable enough with the subject matter to revise the code to incorporate new datasets, but this could be a less white option:
https://paperswithcode.com/dataset/fairface

Since those are images of faces instead of full people, though, we could then compare it (possibly? not sure if the model would pick up white space differences more than faces) with this dog face dataset:
https://github.com/GuillaumeMougeot/DogFaceNet

Might also be great to incorporate into the slide deck some of the recent literature on how machine learning algorithm output can reproduce or magnify racist trends in the feeder datasets.