ssu_geosciences

Sonoma State University Geosciences - Koret Scholarship, Sigmaclast Classifier

Setup

Creating an environment from an environment.yml file

Use the Terminal or an Anaconda Prompt for the following steps.

Create the environment from the environment.yml file:

conda env create -f environment.yml

Activate the new environment:

Windows: activate Keras
macOS and Linux: source activate Keras

Verify that the new environment was installed correctly:

conda list

Config

There are a few options that can be adjusted. To change settings open config.py in any text editor.

emacs config.py

or

vim config.py

Options that can be edited are as follows:

model_name: the model names are listed in the config.py file, the only model that is custom is the SSUGeosciences model

batch_size: this is the number of images processed per iteration in an epoch.

num_epochs: number of times to train

learning_rate: the smaller the number the longer it takes to learn. Too small of a number or too large of a number 
               can cause our learning to be unsuccessful
               
ratio_train: since we do not break our images into train/dev/test sets we must do that dynamically. This is the % of
             our images that we want to use in our training set. ratio_dev is derived from ratio_train and ratio_test

ratio_test: --CURRENTLY UNUSED AS OUR DATASET IS TOO SMALL-- this is the percent of our images that we will use to test upon. 

image_directory: the directory that our images to binary classify our stored in. Our images must be stored in their own 
                 subdirectories. I.e. images/with and images/without. Only 2 directorie will be considered when the program 
                 is ran. 
output_directory: This directory will store the output of our model. Every time the model is ran the results will be 
                   stored in output_directory/model_name/%.txt where % is the accuracy percent
                   
optimizer: An optimizer is one of the two arguments required for compiling a Keras model. There are a few keras optimizer  
           options that are all listed in the config.py just above the optimizer definition.

Use

Activate the Keras environment

source activate Keras

Run the binary classifier

python main.py

Output is produced in multiple ways: 1. The output is displayed to the user 2. The config file used is copied and it's results are appended as a comment to the bottom of the file. Additionally the file is copied to the results folder defined in the config.py file with it's name based on the accuracy of the script run.

GPU Time Issues

When utilizing multiple GPU's and K-Fold cross validation, it is worth noting that the time to setup multiple GPU models can rapidly increase the time it takes to test our data. It seems to takes ~3s per K in K-Folds to initialize our models.

Additionally: If your dataset is small enough, there is a speed up to be had by keeping our model on one gpu.

Visualize GPU Utilization

watch -n 0.5 nvidia-smi

Note on K-Fold cross Validation

We use K-Fold cross val merely to test the validity of a model. After we choose a model, based on our cross validation we should then train our model on all of our data.

https://stats.stackexchange.com/questions/52274/how-to-choose-a-predictive-model-after-k-fold-cross-validation

Say we have two models, say a linear regression model and a neural network. How can we say which model is 
better? We can do K-fold cross-validation and see which one proves better at predicting the test set 
points. But once we have used cross-validation to select the better performing model, we train that
model (whether it be the linear regression or the neural network) on all the data. We don't use the 
actual model instances we trained during cross-validation for our final predictive model.