About Sahar,
an intermediate data scientist with critical and analytical thinking , more specialized in computer vision tasks like object detection and instance segmentation.
I obtained a Master's degree in image and signal processing in 2021, which included studying Statistics & Probability, Machine learning & Deep learning, computer vision, image processing techniques like Denoising, Segmentation, Morphological Operations.
I started my professional journey in data science as a R&D intern in Deep learning applied to computer vision at IOVision, my main mission was to reproduce Deepfashion2 paper more precisely working on Detection, Segmentation and Re-Identification of Clothing Images.
After 6 months of internship, I was hired by IOVision to start working on their new start-up stile.ai (https://stile.ai/ ). My achieved tasks covered different stages of the project from ideation to prototyping from an AI perspective. In fact the product included features based on: object detection and instance segmentation, recommendation system, virtual try-on, and clothed user animation.
These are my practical skills that I have acquired during my data science journey:
-
Modeling : iterative training mask rcnn and faster rcnn backboned by resnet50+ FPN, from detectron2 model zoo, dealing with overfitting by reducing model complexity, applying early stopping and improving train set quality.
-
Data engineering : data scraping with selenium, background removal, data cleansing, data inspection and EDA and making synthetic data (blending)
-
MLops: model deployment on serverless virtual machine using aws lambda and docker.
-
R&D : tracking trending computer vision papers in CVPR and ICCV workshops that aim to apply AI in fashion industry : clothing detection, virtual try on and novel view synthesis
Passionate with AI, I like to participate in AI competitions on Zindi and Kaggle, and I was winner of 2020 AFD AI solutions for gender based violence.
I like hiking at random forests and attending AI events online.
My Latest Projects
The challenge is to build a classifier that can detect ARGs. Using genetic sequences as input, the classifier should be able to identify if a gene is antibiotic resistant or not.
-
EDA & preprocessing : tokenization, removing rare amino acids
-
encoding gene sequences and preparing dataloaders
-
feature embeddings extratcion using pretrained prot bert
-
reducing embeddings vectors dimensions using pca tsne and other Manifold Learning methods from scikit learn
-
visualize feature distribution
-
training on prot_fat_bert
* Literature review about deep fashion and deep fashion2 data sets
* Exploratory data analysis of the deep fashion2
* Implementation of tensorflow based mask RCNN on deep fashion 2 to recognize and segment clothing items in an image
* Iterative training and regularization
* Evaluation using coco metrics, F1 score and auc
* Balancing the data set by extracting part of it ( only the good data) and reducing the classes from 13 to 8 and then from 8 to 5 classes ( long sleeved shirt, short sleeved shirt, outwear, shorts
and trousers )
* Elaboration of computer vision added value of stile.ai app
* Implementation of pytorch based mask RCNN from detectron2 model zoo on a processed and balanced fashion data set
* Scrapping images from e-commerce fashion websites to test the final model resulted from trainings
* Inference and testing the final model in real context
* Model deployment on local host using docker
* Model deployment on serverless AWS lambda and integration in the stile.ai App as detection engine