Dynamic AutoML: Comprehensive Solution for Diverse Data Tasks

Author

Swaraj Khan P

Published

March 19, 2024

Dynamic AutoML: Comprehensive Solution for Diverse Data Tasks

Dynamic AutoML is a versatile platform designed to streamline various data tasks, including CSV analysis, LSTM modeling, and image classification and detection. Our platform offers advanced features and capabilities to empower developers in handling diverse datasets efficiently.

Dynamic AutoML: A Closer Look

CSV Dataset Analysis and LazyPredict Model

Dynamic Dataset Architecture - The architecture of CSV datasets is dynamically determined by analyzing their structure, including the number of records, columns, and their types (textual, numeric, date). This dynamic approach ensures that our platform can handle a wide range of dataset formats without requiring manual configuration.

Sno X_columns Y_columns Datasets Lazy Predicts (Y / N) LP Model and max accuracy Initial ANN loss Final ANN loss Percentage loss
1 age, anamía, creatine_phospokinase death_event Heart Failure No Null 0.28 0.2 26.60%
2 power_consumed, weather_index, holiday_index, humidity maximum-temperature Electricity Yes AdaBoost - 91.3% Null Null Null
3 tweet label Twitter Tweets Yes XGBoost - 93.1% Null Null Null
4 cap-shape, cap-surface, habitat, veil-color class Mushrooms Yes Null 0.36 0.24 35.10%
5 age, bmi, children, smoker charges Insurance Yes Random Forest - 85.1% Null Null Null
6 age, gender, polyuria, sudden-weight-loss, weakness class Diabetes Yes Decision Tree - 95.1% Null Null Null
7 customer-ID, credit-score IsActiveMember Bank Churning No Null 0.36 0.24 31.70%

LazyPredict Model Implementation - Our LazyPredict model is implemented from scratch to provide developers with a comprehensive tool for model selection and comparison. By automating the process of evaluating multiple models with various configurations, developers can quickly identify the most suitable model for their specific task, saving time and effort.

Image Classification and Detection

Automatic Model Training - Our platform simplifies the process of image classification by automating model training. Developers can upload a zip folder containing images organized into folders as classes. The platform then trains models using this data, enabling accurate classification of new images based on their content.

Sno Dataset Name Accuracy Precision Recall Intersection over Union Dice Coefficient
1 Dogs vs Cats 90% 86% 88% 89% 83%
2 Medical Images 95% 90% 91% 85% 87%
3 Autonomous Driving 85% 84% 87% 86% 87%
4 Satellite Imagery 91% 89% 90% 88% 85%
5 Histopathology 93% 88% 91% 90% 83%
6 Semantic Segmentation Benchmark 85% 89% 86% 86% 88%
7 Lung Nodule Detection 91% 88% 91% 85% 86%
8 Plant Disease Identification 92% 82% 87% 78% 80%

Dynamic Image Segmentation - Using techniques like YOLO, our platform dynamically determines classes from datasets such as COCO128, enabling precise image segmentation. This capability allows developers to identify and isolate specific objects within images, opening up possibilities for applications such as medical image analysis, autonomous vehicles, and more.

LSTM Model Training

Dynamic Architecture Determination - Similar to CSV datasets, the architecture of LSTM models is dynamically determined based on the dataset’s characteristics. This approach ensures that the model architecture is optimized for the specific task and dataset, leading to improved performance and adaptability.

Sno Testing Loss Tabulation X Columns Y Columns Initial Loss Final Loss Decrease in Loss Percentage Decrease
1 Tesla Dataset Date year, Date_month, Open, High, Low Close 1,745,438.13 475.4735 1,745,362.66 99.9728%
2 Traffic Dataset Date year, Date_month Vehicles 952.4327 6.8296 945.6031 99.2838%
3 Air Passengers Date year, Date_month Passengers 96,211.27 51,782.60 44,428.63 46.1366%
4 Panama Electricity datetime year, datetime_month, T2M_toc T2M_san 204.77 6.1987 198.5713 96.9723%
5 Google Train Date year, Open, High, Volume Close 180,635.13 9,386.0029 171,249.13 94.8127%
6 Apple Date year, Date_month, Open, High, Volume Close 27,494.8 3,015.0009 24,479.8 89.0341%
7 Netflix Date year, Date_month, Open, High Close 187,213.42 22,841.346 164,616.19 87.8039%
8 London Bike timestamp_year, timestamp_month, wind_speed hum 5,447.62 3,242.8624 2,204.76 40.4717%
9 Electricity_dah date_year, Date_month, date_day temp 8,542,182.42 1,325,564.88 7,216,617.54 84.4874%
10 LSTM-Multivariate pollution date_day, date_hour, wnd_spd, pressedev, pollution temp 300,186.5 173.9029 126.2786 42.076%

Streamlined Model Training - Our platform streamlines the process of LSTM model training by automatically tuning hyperparameters based on dataset properties. This automation reduces the manual effort required for hyperparameter optimization, allowing developers to focus on model experimentation and refinement.

Usage

  1. Upload Data: Upload CSV files or image datasets to start your analysis and model training.
  2. Dataset Exploration: Explore dataset properties, perform preprocessing tasks, and visualize data distributions.
  3. Model Training: Choose appropriate models (LSTM, image classifiers, etc.) and train them using automated processes.
  4. Model Evaluation: Evaluate model performance using metrics and visualizations provided by the platform.
  5. Deployment: Download trained models for deployment or integrate them directly into your applications.

Access: Access the platform directly here.

Source Code

For more details, visit the source code on GitHub.

How It Helps Developers

Dynamic AutoML offers a range of features designed to streamline the development process and empower developers in handling diverse data tasks