GSoC 2020 Amogh Jahagirdar
Name: Amogh Jahagirdar
University: National Institute of Technology, Tiruchirappalli, India
Course: Bachelor of Technology, Instrumentation and Control Engg. (Major) and Computer Sci.(Minor)
Email: jramogh@gmail.com
Github: https://www.github.com/amogh-jrules
Kaggle: https://www.kaggle.com/amoghjrules
LinkedIn: https://www.linkedin.com/in/amoghjahagirdar/
I am a 3rd year undergraduate student and OSS enthusiast and love contributing to projects that impact human lives. Apart from programming, I enjoy reading and going on runs.
I have not participated in GSoC before and will not be applying to any other projects. I am eligible to receive payments from Google.
I already have been working with the idea and have been sharing my progress with the mentors regularly. During the program I will be comfortably able to allocate 5-6 hours a day and will put in extra hours if the schedule demands it.
I will be having a round of examinations in the Summer because my University has shifted semester dates because of the COVID-19 outbreak. The dates for the same have not been fixed yet, I shall begin working from the community period itself if need be. Apart from that, I will not take any vacations over the duration of the program.
My first hobby project was a sprite based game back in 10th grade. It received 70,000+ downloads on the Play Store. I have been professionally coding for the past 3 years. In the past I have worked with the Hyperledger Blockchain framework and with GroceriStar, which is an OSS driven startup in Ukraine.
I am ranked in the top 0.5% (Rank 495 of 100000+) on Google's Competitive Machine Learning and Data Science platform, kaggle.com. Having worked in the field of Data Science from the past 2 years, I have acquired decent experience with respect to Linear ML Models, Python and Neural Network based models. (Some of my work related to modelling methods, feature engineering, etc.)
Recently, I co-authored a research paper for Beebots, which is a blockchain-based decentralised swarm robotic system. I implemented and tested a density-based clustering algorithm in Python which helped optimize the bots' behaviour. (Github repo)
Last winter, I mentored at the Tensorflow organisation at Google Code-In. I meticulously reviewed 70+ machine-learning tasks and provided constructive criticism and assisted students in debugging their code. (Certificate)
In September 2019, I worked on Ytality, a product ecosystem to transmit vitals of an ailing patient via the cloud, pass them through disease-diagnosis models and display them to doctors. I trained models using TensorFlow and Keras on extensive datasets of EEG, ECG and heart-sound classifications. (Details and Code)
All my above projects have git-based version control and are hosted on github.com. I have a working understanding of git, which I have gained from practical use. My OSS experience can be tracked via my Github contributions.
Contributions to SunPy (5 PRs, 3 merged, 4 issues, 3 closed)
Notebook containing my exploratory analyses and implementation of baseline models
Colab notebook with Sunspotter EDA with interacting with SunPy modules
Most people do not realise how significantly the Sun and solar weather affect their daily life. Although there have been multiple attempts to forecast solar weather, I feel there remains a lot left to be done in the field. I want to achieve the state of the art forecasting ability for Solar weather.
What excites me about this project is the long lasting impact my work will have on human lives and the fact that I will be able to contribute a tiny bit to the overall progress of mankind.
I chose this project because it aligns with my technical skills and interests. Moreover, the SunPy community is very responsive and receptive, which is a sign that they are passionate about the project.
As mentioned earlier, I have a sound background in Machine Learning, Neural networks and Data Science. Being fluent in Python will enable me to write clean, readable code and maintain the high coding standards of the library. Self motivated and enthusiastic, I will not have any problem working independently, asking the mentors for help only when it is really needed.
On plotting out the target variable (the classification variable) a clear bias towards the negative class is visible. This unbalanced target distribution can lead to poorly trained models. Having discussed this with the mentors before, I decided to implement Stratified K-Fold validation. Stratified K-Fold takes K-Fold validation a step further in that it ensures that the distribution of classes in every fold is identical to that of the overall dataset. I have already implemented a flexible Stratified K-Fold generating pipeline as a means of generating Out of Fold training data.
For simplicity, I have broken down the project into a number of subtasks, as shown below:
It involves feature cleaning (e.g., converting timestamps into an ordinal variable), handling missing values, normalizing data.
- Metric for normalization - Z-Score
Exploratory Data Analysis or EDA involves plotting graphs inorder to uncover underlying trends in the data and relations between the independent features. It consists of :
- Visualising the distribution of variables
- Plotting Pearson Correlation Coefficients and much more
The inferences obtained on analyzing these plots will provide the basis on which the modeling technique(s) are chosen.
The inference obtained will not only be used to choose the appropriate models, but will also be used to analyze the possibility of generating new more information rich features by combining 2 or more features with a low correlation score, or dropping redundant features which will optimise the model.
Here is where most of the heavy-lifting will be done.
- Implementing a Deep Flare Net (DeFN) for Solar Flare Prediction. It will involve:
- Multilayer Image-based DNN to predict the probability of flares occurring within the next X hr of time in an active region
- Further used to determine the most likely maximum classes of flares
- The cost function this NN will be the minimization of cross entropy
- A major advantage of this implementation are that the features are human readable unlike CNNs which extract imaging features that we cannot understand.
- It is also possible to rank these features in order of importance
- Will be implemented with Pytorch or TF Keras
- Using linear and non-linear models from sklearn, xgboost and lightgbm. They will be:
- Logistic Regression
- Support Vector Machine
- Random Forest Regressor
- Elastic Net
- Lasso
- XGBoost Regressor
- Light Gradient Boosting Machine Regressor
- Ensemble learning pipeline. This will consist of:
- Averaged Voting - predictions of all models are averaged
- Weighted Voting - the predictions of models are given varied weights based on their importances/contribution to accuracy
- Out of Fold Meta-Modeling - the predictions of all models are taken in as training data for a meta-model which gives the final classification
- Hyperparameter optimisation to further optimize models
- GridSearchCV from sklearn - allows us to pass a range of values for each parameter, out of which the optimal value is searched and selected
These methods are the industry standard in modern day data-science. I have just made a baseline implementation of the same in the notebook and yet I was able to achieve accuracy upwards of 88%. Over the course of the project, I hope to fine tune my models and methods and achieve much higher accuracies, setting the benchmark for future attempts in the field.
Task 4 - Implementing state of the art Unsupervised and Semi-supervised Clustering for Image/AR Data Classification
- Implement the state of the art Invariant Information Clustering for Unsupervised Image Classification research paper
- This technique is not specialised to computer vision and hence can be used for the AR data extracted by SMART as well
- The method operates on any paired dataset samples. We will use random transforms to obtain a pair from each image or data sample
- By clustering event samples together, this model will also serve as an anomaly detection framework. Moreover, it can run on unannotated data as well.
- Will experiment with unsupervised and semi-supervised learning and compare metrics to see which method proves to be more useful
- The model is trained to maximise mutual information between the class assignments of each paired dataset sample and is has a solid base in information theory
- Will be implemented with Pytorch or TF Keras
Integrating the above mentioned tasks in a new module sunpy.ml, with submodules for supervised and unsupervised learning methods. These methods will have objects of the type of models that I will have implemented, which will have common member functions like fit, predict, etc. I have not had much time to discuss this thoroughly with the mentors so I shall be flexible and accommodate any changes that the mentors may feel appropriate
This constitutes writing unit tests for each sub-module within the project and integration tests to ensure that there are no conflicts within the module itself. I have previous experience with writing tests for my projects. However, I will still refer to good testing practices and test-automation using tools like CircleCI to ensure that code quality remains speckless
Writing down documentation which clearly explains the purpose of the module, the classes, member functions, callbacks and other details along with their usage
- Reading through research papers that will be implemented
- Re-setup local development environment
- Interaction with the community
- Data cleaning, handling missing values
- Trying out different normalization techniques and choosing one that suits best
- Complete data pre-processing
- Exploratory Data Analysis
- Feature generation
- Train models from sklearn, xgboost and lgbm
- Implement ensemble learning
- Write unit tests for the linear model pipeline
- Write integration tests for linear model pipeline
- Publish the preprocessed, feature generated dataset
- Consolidate all the various plots, models and techniques implemented in the past month into a jupyter notebook for the SunPy example gallery
Consists of :
- Notebook for SunPy gallery showcasing EDA
- Code for linear models + ensemble learning pushed to a dev branch under sunpy.ml module
- Implementation of Deep Flare Network (DeFN)
- Allocating 2 weeks here to ensure there is sufficient time for implementing and training the model
- As the model trains, I will use this time to overcome any lag in the proposed schedule
- Writing the essential classes and member functions which the user will interact with directly when using the module
- Write a jupyter notebook showing the use of the neural network for SunPy example gallery
- Write unit tests for the DeFN pipeline
- Write integration tests for the DeFN pipeline
- Will begin SunPy related tasks
- Matching catalogues
- Write lambdas to identify common parameters
- Implement matching algorithm from TOPCAT for the identified common parameters
- Expand hek2vso to be able to accommodate any data archive obtained from Fido and other sources
- Setup uniformity between different data catalogues such that all catalogues have the same attributes, thus increasing their usability
- Other SunPy related tasks that may come up on the way
- Matching catalogues
Consists of :
- Deep Flare Network code pushed to dev branch under sunpy.ml module
- Complete SunPy related tasks by mid-week
- Begin implementation of Unsupervised Clustering Neural Network Model
- Complete implementation and training of Unsupervised Clustering NN Model
- Write unit tests and integration tests for Unsupervised Clustering NN Model
- Test out the model against SDO/HMI data and the generated data and compare accuracy metrics
Consists of:
- Unsupervised Clustering NN code in dev branch under sunpy.ml module
- SunPy related tasks (expansion of hek2vso, catalogue matching etc.)
- Compare the results of models to the ELO ratings and if it supports the existing theories regarding solar activity
- Write notebooks showcasing both neural network implementations, along with external hyperparameter tuning
- Documentation and final touches
Consists of:
- Documentation and final changes in existing code
- Notebooks demonstrating NN implementations to the SunPy example gallery
- Final Evaluation
- Trained linear and non-linear ML based models
- Trained Supervised and Unsupervised Neural Network models
- Unit and Integration tests for seamless integration into the SunPy library and to prevent future breakage of code
- Notebooks explaining how to utilise each component implemented above for the SunPy example gallery
