metadata

title: Nexa R&D
emoji: 🔬
colourFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.34.2
app_file: App.py
pinned: true
license: apache-2.0
tags:
  - optimization
  - machine-learning
  - research-tool
  - gradio
  - azure-sky
colorFrom: blue

Nexa R&D

Nexa R&D is a visual research platform designed for researchers and industry professionals to compare and evaluate optimisers (e.g., AzureSky, Adam, SGD, AdamW, RMSprop) on analytical benchmark functions (e.g., Himmelblau, Ackley) and machine learning tasks (e.g., MNIST, CIFAR-10). It supports ablation studies, hyperparameter tuning, and side-by-side evaluations through an intuitive Gradio-based interface, optimised for deployment on Hugging Face Spaces.

Features

Modes:
- Benchmark Optimisation: Visualise optimiser trajectories on loss surfaces of mathematical functions.
- ML Task Training: Train and compare optimisers on datasets like MNIST and CIFAR-10.
Optimisers: AzureSky (hybrid SA + Adam), Adam, AdamW, SGD, RMSprop.
Ablation Suite: Configure AzureSky’s Simulated Annealing (SA) with options to enable/disable SA, set initial temperature, and adjust cooling rate.
Interactive UI: Gradio interface with plots, metrics tables, and JSON export for results.
Metrics:
- Benchmark Mode: Distance to global minimum, final loss, convergence rate.
- ML Mode: Train/validation accuracy, generalisation gap, final loss, best epoch.
Deployment: Optimised for Hugging Face Spaces with optional GPU acceleration.

Usage

Configure a Study:

Select Mode: Choose "Benchmark Optimisation" or "ML Task Training" from the Study Configuration tab. Select Optimisers: Pick one or more optimisers (e.g., AzureSky, Adam). Configure Parameters: Benchmark Mode: Select a function (e.g., Himmelblau) and dimensionality (default: 2). ML Task Mode: Select a dataset (e.g., MNIST), epochs (default: 10), batch size (default: 32), and learning rate (default: 0.001).

Ablation Settings (if AzureSky is selected): Enable/disable simulated annealing (default: enabled). Set initial SA temperature (default: 1.0). Set SA cooling rate (default: 0.95).

Run Study:

Click "Run Study" to execute the experiment. View results in the "Results" tab, including: Plots (loss surfaces for benchmarks or accuracy/loss curves for ML tasks). Metrics table summarising performance. Detailed JSON metrics.

Export Results:

Click "Export Results as JSON" to download a results.json file containing metrics, paths, and histories.

Ablation Suite The ablation suite enables detailed analysis of the AzureSky optimiser’s components:

Simulated Annealing (SA): Toggle SA on/off to assess its impact on optimisation. Initial Temperature: Adjust the starting temperature for SA (higher values increase exploration). Cooling Rate: Control the rate at which SA cools (values closer to 1 result in slower cooling, preserving exploration).

To use:

Select AzureSky in the optimisers list. Open the "AzureSky Ablation Settings" accordion in the Gradio UI. Adjust SA parameters and run the study to compare results with other optimisers or configurations.

Example To compare AzureSky (with SA) and Adam on the Himmelblau function:

Select "Benchmark Optimisation" in the Study Configuration tab. Choose "Himmelblau" and the optimisers "AzureSky" and "Adam". Set dimensionality to 2. In the AzureSky Ablation Settings, enable SA, set temperature = 1.0, and set cooling rate = 0.95. Click "Run Study". View the 3D loss surface plot and metrics table in the Results tab.

Testing Checklist

Optimisers: Verify convergence on benchmark functions. Benchmarks: Confirm global minima and surface plots are accurate. ML Tasks: Check epoch stability and output formats. UI: Test mode switching, input validation, and result display. Ablation: Validate AzureSky behaviour with/without SA and different temperature/cooling settings. Export: Ensure JSON exports include all metrics and results.

Future Enhancements

Support for user-defined benchmark functions via file uploads. Additional ML datasets (e.g., Fashion-MNIST). API access for scripted experiments. Extended ablation options for other optimisers.

For issues or contributions, contact the maintainers.