nrunx --network host --env USER=$USER --env HOME=$HOME --env DISPLAY=$DISPLAY nvcr.io/idmod/keras_tf_mnist:latest
cd DataScience/MNIST
python3 mnist_tutorial.py --train
python3 mnist_tutorial.py --test
`
What just happened?
This is the famous MNIST 'hello world' of neural networks. At first we train a simple convolutional neural net on the MNIST
digits dataset. We just use 5 epochs for speed but you can see error gets small fast. The code shows how to do this in
Keras. The trained network is saved and when you run --test you reload that network and do inferences. Rather than just
measure the accuracy we go digit by digit and find errors and display those images to the user, along with what the NN
thought it was vs what it actually is. You can see that those failures are actually pretty hard for humans to classify also.
nrunx --network host --env USER=$USER --env HOME=$HOME --env DISPLAY=$DISPLAY nvcr.io/idmod/keras_emulator_fromfile
cd <...>/DataScience/RANDOM_FORESTS/
./generate.py > generated.csv
python3 random_forest.py generated.csv
What just happened? This demo shows Random Forests in Python. This does not require Keras, just SciKit Learn. This first example isn't an actual dataset but a numerical generated 'dataset' where the output column is literally a mathematical function of the inputs. This is my version of an RF hello world. Next will be the bostonhousing dataset and then the PIMA indian one.
python3 random_forest.py ../DATA/boston_housing_csv.csv
python3 random_forest.py ../pima-indians-diabetes.data
python3 random_forest.py ../DATA/surveillance_data.csv
The final one there is a dataset from Brittany related to costs reported from different countries for doing Polio program surveillance.
In each case the final result is a sorted barchart plot of the correlates to the output. Note that Random Forests on this size of data really don't need any special hardware.
- Emulator
- Neural Net Regressor
- Spotify
- Sentiment Analysis