README:

---

Deep Learning and Chemometrics: Quantitative and Qualitative Spectroscopy Interpretation of Aqueous Solutions

Colin Murphy; CS 615 Deep Learning - Spring 2020

---

Download copy of the final paper for analysis details

Packages Needed:

• NumPy, Matplotlib, Pandas, Keras, Sklern, TensorFlow (backend)

Code for cleaning of the original .txt files (provided by the International Diffuse Reflectance Conference) is performed in "NIR_DATA_CLEAN.py"; however, you do not need to execute the script, as the cleaned csv datasets have been included in this distribution (within data folder).

All further data processing (standardization, data augmenting, and one-hot encoding) can be found in "DATA_PREP_MODELING.py"

The modeling structure (built using Keras and TensorFlow) can also be found within "DATA_PREP_MODELING.py"

CODE NOTES:

"GlobalStand" function performs standardization over the entire dataset that is given (zero-mean and unit standard deviation).

"load_dataset" function loads the csv data file into a pandas dataframe - indexed by the sample ID and applies "GlobalStand" to spectroscopy dataset.

"plot_spectrum" function plots a given spectrum (requires dataframe format) for visualization.

"one_hot" converts analyte names to unique ID keys.

"data_augment" function create a random augmented spectrum from a given sample.

"model_in_data" function performs the final data prep before training - indicating "one_hot_y" to be applied if data is being used for classification. Furthermore, it performs data augmentation on the training dataset, creating 10 augmented samples, per original sample and concatenating them.

"classification_model" defines the structure of the CNN network for classification, with the KerasClassifier wrapper being used to feed the model into the sklearn cross_validation function to perform a stratified 5-fold cross-validation and return the mean and standard deviation of the accuracy.

"regression_model" defines the structure of the CNN network for regression, with the training being performed when calling the function. Function returns a tuple of the trained model and the testing dataset error.

"coeff_determination" function calculates the coefficient of determination (R^2) between true and predicted.

"huber" function calculates the huber error between true and predicted.

"rise" function calculates the rise error between true and predicted.

"f_map_visual" function graphs the feature maps of the trained model's first convolution layer for visualization.