Image Restoration, Face Restoration Papers

GFPGAN 

Paper - Towards Real-World Blind Face Restoration with Generative Facial Prior

Key Notes

• GFP-GAN that leverages rich and diverse priors encapsulated in a pretrained face GAN
• StyleGAN -  generating faithful faces with a high degree of variability
• GFP-GAN with CS-SFT layers achieves a good balance of fidelity and texture faithfulness 
• Generative facial priors (GFP)

Image Restoration Techniques

• Super-resolution
• Denoising
• Deblurring
• Compression removal
• GFP-GAN is comprised of a degradation removal module (U-Net) and a pretrained face GAN (such as StyleGAN2)
• Reconstruction loss that constraints the outputs yˆ close to the ground-truth
• Adversarial loss for restoring realistic textures
• Identity preserving loss.

State-of-the-art face restoration methods

• HiFaceGAN
• DFDNet
• PSFRGAN
• Super-FAN
• DeblurGANv2
• ESRGAN
• Real-ESRGAN
• BasicSR
• FaceXLib
• HandyView
• GFPGAN Inference Demo (Paper Model)
• GFPGAN Inference Demo
• GFPGAN (CVPR 2021)
• More Projects - Link

Keep Reading!!!

Connected Cars - Telematics Use cases - Research Paper

Paper - How much telematics information do insurers need for claim classification?

Key Notes

• By comparing a few classification algorithms, we find that logistic regression with lasso penalty is the most suitable for our problem
• Telematics data become redundant after about 3 months or 4,000 kilometers of observation, at least from a claim classification perspective
• Telematics data fall under the definition of “personal data”, and must therefore be handled according to the relevant legislation
• In this analysis, only collision coverage claims were considered, i.e. the target column given as input to the classification models is the indicator of a collision claim, at-fault or not
• Instead of having the indicator of a claim as the response variable, we would instead have the number of claims, moving us into a counting regression context
• Low claims / Medium / High claims we could bucket and assign premium accordingly

Classical features selected for the analysis

“usage”, “travel habits” and “driving performance”. Claimants (those who have claimed at least once during their observed year) and the non-claimants (those who have not claimed during their observed year)

Summary of Features

Collected Feature

• VIN
• Trip number
• Departure datetime 
• Arrival datetime 
• Distance 
• Maximum speed

Derived Features

• annual_distance
• commute_distance
• conv_count_3_yrs_minor
• gender
• marital_status
• pmt_plan
• veh_age
• veh_use
• years_claim_free
• years_licensed

Derived Features by “usage”, “travel habits” and “driving performance”

• avg_daily_distance
• avg_daily_nb_trips
• med_trip_avg_speed
• med_trip_distance
• med_trip_max_speed
• max_trip_max_speed
• prop_long_trip
• frac_expo_night
• frac_expo_noon
• frac_expo_evening
• frac_expo_peak_morning
• frac_expo_peak_evening
• frac_expo_mon_to_thu
• frac_expo_fri_sat

More Reads

• Evolution of Insurance: A Telematics-Based Personal Auto Insurance Study
• Driver Identification Based on Vehicle Telematics Data using LSTM-Recurrent Neural Network
• Vehicle Telematics Via Exteroceptive Sensors: A Survey
• Secure OTA Software Updates in Connected Vehicles: A survey

Keep Thinking!!!

Can I master all Kubernetes, Computer Vision, Data Algos, NLP ?

A very good read - link

Copying a few lines/summary from it from the perspective that echo's my views

Perception -  I believed that Kubernetes was essential to the DS/ML workflow.

Experience - However, as I learned more about low-level infrastructure, I realized how unreasonable it is to expect data scientists to know about it

Fact / Reality - In theory, you can learn both sets of skills. In practice, the more time you spend on one means the less time you spend on another.

My perspective - We can know few things in-depth and need to master them with multiple experiments. You can master few areas and have a broad understanding of the rest of them. Compile knowledge vs Customize knowledge vs Solve in your own way is different.

Interesting Analogy -  I became a data scientist because I wanted to spend more time with data, not with spinning up AWS instances, writing Dockerfiles, scheduling/scaling clusters, or debugging YAML configuration files.

Recommendations

• Have a separate team to manage production
• Infrastructure abstraction kubeflow, metaflow, google vertex is useful for non-trivial workflows, and multiple models in production.

It's a good thread. 

Keep Going!!!

Vision Fashion Papers

Paper#1 - POG: Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion

Key Notes

• Personalized Out€fit Generation (POG) model = user preferences
• regarding individual items +  outfi€ts with Transformer architecture
• Personalization represents how the recommendations meet users’ personal fashion tastes

• Key features - brand, category, style, pattern
• Fashion Outfit Model (FOM) by learning the compatibilities between each item and all the other items within the outfit
• Personalized Outfit Generation (POG) model, which can generate compatible and personalized outfits based on users’ recent behaviors

• Combination of NLP, Vision, Graph Embedding

• This could be a combination of user-user, item-item and bought sequences of complete pairs

Paper #2 - MMFashion: An Open-Source Toolbox for Visual Fashion Analysis

Key Notes

• Fashion Attribute Prediction, Fashion Recognition and Retrieval, Fashion Landmark Detection, Fashion Parsing and Segmentation and Fashion Compatibility and Recommendation.
• Dataset - DeepFashion, Polyvore
• Clothes Retrieval
• Landmark Detection
• Cloth Detection and Segmentation
• Fashion Compatibility and Recommendation

Paper#3 - c+GAN: Complementary Fashion Item Recommendation

Key Notes

• Bidirectional LSTM model to sequentially predict the next item conditioned on previous ones
• Clustering the intensity field of the images, with K-means clustering results in these dominant clusters
• Combination of Text + Vision Similarity + GAN would be good

More Reads

CRAFT: Complementary Recommendation by Adversarial Feature Transform

Keep Reading. This is just very basic skimming!!!!

Health vs Career

1. To pursue your interests you do not need titles
2. To make your ideas work, you need to focus
3. Don't wait for the title, prepare yourself as an independent consultant
4. Titles vs contribution are debatable and depends on company, Your ideal role vs what you have gaps address it
5. Every problem can be done in multiple ways, Think like a consultant, observe as a reviewer when you see already someone solved the problem
6. Connect dots of experience and apply for every new use case
7. Health cannot take a pause but a career can take a pause
8. When sabbatical matters, give a break for your body, mind, and soul
9. Career gaps do not mean you stopped learning, It only means you stopped earning
10. Having multiple passive incomes outside your work will assist in long term
11. Compete against yourself, Stay firm on your principles
12. New skills = Old Skills + New changes or features, Apply experience-based differential thinking
13. Overall its just one life, be less regretful
14. Money does not equate experiences learned or earned
15. Money matters but health matters more

Keep Going!!!

For every data / data science use case

Step #1 - Build Reporting

• Consider Dashboards for key metrics, KPIs
• Spot daily business / Trends

Step #2 - Data Exploration to Understand data

• Learn domain, Explore it
• Spot features outside your data
• Build insights from your data
• Scrap external data like local demographics data

Step #3 - Model Considerations @ Lowest level or Enterprise level

• Consider building Global and local model
• Different algos and outputs, Consider ensemble or one model depending on model performance
• Visualize models with interpretation
• Overlay charts for predictions/analysis
• Business knowledge guides model correlations, Constantly validate with business

Step #4 - Model Optimization / Improvements - Keep Learning

• Continuously build optimize models
• Measure Model drift comparison past to present
• Consider keeping add on variables as needed
• Evolve it
• Depending on Deployment scenario quantize / optimize to lite weight models

Step #5 - Be ready to collaborate and take business inputs in regular intervals

• Good design comes from clarity of thinking
• Customer-first approach
• Design for scalability vs get something working

Step #6 - Deployment - Ready for consumption

• Scale with expertise as needed
• Dockerize as possible
• Expose as API endpoints
• Build security
• Deploy and run as needed to minimize costs

Step #7 - Organize code, data, models

• Version data vs model built
• Keep track of patterns of data vs model improvements
• Look at the explainability aspect using LIME / SHAP

Customize the reporting layer built-in step #1 to remap model behavior vs actual data, Leverage it as patterns vs predictions vs actuals.

Keep Iterating!!!

Parenting in 2021

• Observe closely if any behavior changes
• Teach them about good touch / bad touch
• Avoid excess smartphone usage
• Monitor their online activities
• Discuss reported issues/ abuse and make them understand Awareness of surroundings
• Provide guidance, alert them, situational analysis
• Proper sex education and awareness of their hormonal changes
• Today everything is on smartphone both good and bad
• Provide better guidance 

Happy Parenting!!!

Creditworthiness and Features used in Machine Learning models

Some readings on features used.

Paper #1 - Determining Secondary Attributes for Credit Evaluation in P2P Lending

Key Features

Paper #2 - Credit Scoring for Good: Enhancing Financial Inclusion with Smartphone-Based Microlending

Key Notes

1. Socio-Demographic: including bank history, income and debit account behavior.
2. Calling Behavior: aggregated values for number and duration of phone calls made and received on different days and at different times of the day.
3. Link-Based: counts of the number of good and bad credit card holders in each customer’s egonet.
4. Influence Score: the scores each customer obtained after two distinct influence propagation algorithms were applied to the network.

Paper #3 - Machine Learning approach for Credit Scoring

Key Notes

• LIME, Local Interpretable Model-agnostic Explanations, is a novel technique that explains the predictions of any classifier in an interpretable and faithful manner
• SHAP, which stands for (SHapley Additive exPlanation) [23], is a novel approach for model explainability which exploits the idea of Shapley regression value16 to model feature influence scoring.

Paper #4 - A comparative study of forecasting Corporate Credit Ratings using Neural Networks, Support Vector Machines, and Decision Trees

Key Notes

More Reads

Learning Latent Representations of Bank Customers With The Variational Autoencoder

Eliciting Social Knowledge for Creditworthiness Assessment

Keep Exploring!!!

Vision Projects and Inspirations

Vision Projects and Inspirations

• When we have data issues, we can use GAN to generate synthetic images. Code - Generate Artificial Faces with CelebA Progressive GAN Model
• Object Detection from existing trained networks. TensorFlow Hub Object Detection Colab
• Retraining an Image Classifier
• Generating new styles from images. Code - Fast Style Transfer for Arbitrary Styles
• Action Recognition with an Inflated 3D CNN
• Near Duplicate Image Search

Keep bookmarking and Exploring!!!

The Long Journey

• Code from the computer - 2003
• Code from Laptop - 2005
• Code from browser - 2014
• Code in Colab - 2018
• Deploy in GCP - 2020
• Demo in Own Personal Cloud, Sometimes it's better to keep working on ideas than remembering/recollecting them.

Looking at Technology from a Domain perspective or Apply Technology in all applicable businesses. The debate continues!!!

The Journey Continues!!!