⌛ Time-Series / Sequence Multiclass Classification - TensorFlow

This repo contains a write-up for the first stage of a paid project I completed for a client on Upwork. At the time of writing, we are waiting to start stage 2. The client was delighted with the work and left a 5-star review:

🤔 The Problem

The client gave us ~200 rows of data to work with. The data was the output from a machine which analysed seawater. An electric current was passed through each seawater sample and output ~1000 different values. The values vary based on the metals the sample contains. The goal was to classify each of these samples into one of four possible classes: cadmium, copper, lead, and seawater (i.e., no metal found) corresponding to the metal which appears most in the sample. This was complicated somewhat by the concentration of each metal (stage 2 of this project aims to predict the concentration of each metal in the sample).

We built a range of LSTM models and eventually found that an attention-based LSTM worked best and it achieved 97% accuracy. Moreover, to deal with the tiny amount of data we had, we performed extensive data augmentation.

📕 Noteable Notebooks / Files

• data_plots.ipynb - plots of all samples, plus plots coloured by un/successful model prediction
• electro_augmenter.py - data augmentor class that lead to a huge gain in performance
• train_attention_with_augmentation.ipynb - training of final attention-based model with data augmentation. This model was saved and submitted to the client.
• utils.py - main file containing all functions used for model training

🏗 This Repo is a Work-in-Progress

This portfolio is a work in progress. It probably won't be in perfect condition when you read it. But I hope it gives you an idea of the quality of my work and what I can do.

If you are interested in working together, please reach out via my Upwork profile or email me at: adamdmurphy4 [at] gmail [dot] com

📝 Notes

I completed this project alongside a Senior Machine Learning Engineer Waylon Flinn. Waylon wrote a custom attention-based LSTM model which provided a significant gain in performance over vanialla LSTMs. All code files he wrote have 'waylon_' appended to the start.