Google Landmark Recognition 2020 Tensorflow | Computer Vision | Kaggle

Datasets from: GLDv2

What I do:

Using Tensorflow backend to first rank all training images by embedding similarity to test images, then performs geometric verification and re-ranking on most similar training images using KD tree

Fine-tune parameters to achieve a bronze medal on leaderboard

Flowers Recognition Tensorflow | Pytorch | Computer Vision | Kaggle

Datasets from: Kaggle

What I do:

Build a deep learning model using densenet169 in Pytorch and the one using efficientnet B7 in Tensorflow to classify 104 classes of flowers

Extracted information in tfrec data, performed data augmentation, then trained 70k datasets with TPU to achieve 97% accuracy

within 50 epochs in top 8% on leaderboard

Obstacles Detection for Self‐Driving Cars Pytorch

What I do:

Built an object detection model for cars that analyzes short videos taken by webcam and detects obstacles like green or red traffic

lights, pedestrians, and other vehicles etc.

Combined with video analysis to create a pipeline, then examined with YOLOv5 and R-CNN algorithms

Trained on 15k real images Datasets with 70% accuracy within 1000 epoches

Machine Learning-based Super-Resolution MRI Tensorflow | Medical Imaging

Datasets from: RAISR

What I do:

Evaluated the accuracy of Deep Learning and Google RAISR algorithms - a ML technique that produces high-quality versions of low-resolution image - in MRI imaging datasets

Explored Enhanced Deep ResNet (EDSR) and Wide Activation Deep ResNet (WDSR) with Tensorflow

The result demonstrated RAISR was yet to be a qualified model for MRI analysis compared to EDSR

Categorical Feature Encoding Challenge II Tensorflow | Kaggle

Datasets from: Kaggle

see code

What I do:

• Using Tensorflow backend, I build a neural network model combined from Dense-NN, batch normalization and dropout for this datasets.
• Since there are already too many columns, I do label encode for the features, and use Kfold validation to evaluate the models.
• As a comparison, I also perform one hot encode (with more focus on the features themselves) and run catboost model on it. After tuning hyper-parameters, it also does a good job in prediction.

NYC Taxi Duration Machine Learning | Kaggle

Datasets from: Kaggle

see code

What I do:

• The large dataset includes about 100k rows. I perform data selection and EDA based on data visiualization drawn to filter out the data that are most related, and do feature engineering (Principal Component Analysis to divide up the map of Manhattan into regions)
• Perform model selection on different regression models (linear, Lasso, Ridge, SVR etc.) and make better use of the features by putting different weights

Scrapy Program: House on Rent and on Sale Python | SQL

What I do:

• Using Python (requests and beautiful soup) and SQL to capture the housing rent and sale on certain websites, given the input from users specifying the name of the city and number of outputs.
• It collects the information and transforms to an Excel spreadsheet that is used to create data visualization.