Cameron Olson

Overview

MicroFrame is a lightweight educational library emulating a pandas-like interface for data manipulation, designed for students and beginners in data analysis.

Key Features

• Efficient CSV Reading: Rapidly reads CSV files to create MicroFrame objects, tailored for smaller datasets.
• Data Type Handling: Offers advanced data type inference and explicit setting for better data structure control.
• Flexible MicroFrame Objects: Similar in function to pandas DataFrames, enabling intuitive data manipulation and analysis.
• Clear Tabular Display: Enhanced printing capabilities for well-formatted data presentation.
• Robust Data Manipulation: Includes filtering, column type modification, and data summarization, akin to pandas.
• Advanced Indexing: Efficient data access via advanced indexing methods, including an 'iloc' method.
• Data Conversion Tools: Easy conversion to formats like NumPy arrays using the 'to_numpy' method.
• User-Friendly API: Mirrors pandas API to facilitate an easier transition to real-world data analysis projects.

Technology Used

• Numpy: Fundamental to MicroFrame's backend, providing efficient data structures and performance optimization.
• PyTest: Key tool for testing the functionality and reliability of MicroFrame.
• PyTest.Cov: Used to generate comprehensive code coverage reports, ensuring robustness across the library.
• Hypothesis: Enables thorough property-based testing by generating a wide range of test cases.
• Black: Ensures code consistency and adherence to style standards.
• Sphinx: Transforms code comments into detailed documentation, facilitating easier understanding and use of the library.
• Hatch: Aids in the efficient packaging and distribution of MicroFrame, streamlining deployment processes.
• GitHub: Central hub for source code management, version control, and collaboration.
• GitHub Pages: Hosts the comprehensive documentation for easy access and reference.
• GitHub Actions: Automates continuous integration and continuous deployment, enhancing workflow efficiency.

Explore Further

• GitHub Repository: MicroFrame
• Documentation: MicroFrame Documentation
• PyPI: PyPI Project

Overview

My undergraduate thesis, 'Exploring Audio Diffusion Synthesis,' offers a thorough analysis of Raw Audio Neural Network Generation, exploring the adaptation and application of novel AI technologies to satisfy musicians' requirements and innovate in sound design. The project blends theory and practice, supporting findings with a research paper and a dataset of 20,000+ drum samples. I also created music to showcase the practical and creative possibilities of the AI models in sound design.

Key Points

• Dataset Assembly: Compiled a robust dataset of 20,000+ drum samples, ensuring a comprehensive basis for subsequent neural network training and experimental validation.
• AI Model Development and Application: Engineered and trained AI models in the domain of Audio Diffusion Synthesis, utilizing them to generate music and exemplify their pragmatic and innovative applications in sound design.
• Research Synthesis: Conducted rigorous research into Raw Audio Neural Network Generation, culminating in a scholarly research paper that encapsulates the findings and theoretical underpinnings of the explored technologies.
• Technological Exploration: Investigated and implemented emerging AI technologies, tailoring them to address specific requisites of musicians and to innovate in the field of sound design, with a particular focus on diffusion synthesis using the open-source Dance Diffusion model by Harmonai for a large portion of my research.
• Presentation of Findings: Articulated research findings and practical applications through a multifaceted presentation approach, encompassing a research paper, a showcase video, and a master class, targeting both academic and practitioner audiences at Berklee.

Abstract

This research dives into Audio Diffusion Synthesis, exploring the potential of Neural Networks in sound generation. Specifically focusing on the Dance Diffusion architecture, the paper navigates through system selection, data collection, and model training, demonstrating the effectiveness of DDPMs in generating authentic-sounding audio signals, providing valuable insights into the future of audio synthesis in music production.

Explore Further

• Youtube Video: Audio Diffusion Synthesis Showcase
• Research Paper: Audio Diffusion Synthesis Paper
• Musical Example: Stay With Me

Overview

"Almost Human," an innovative artistic endeavor initiated at Berklee College of Music, delves deep into the intricate interplay between Machine Learning and the Artist. This collective is propelled by the conviction that the future of music resides at the crossroads of cutting-edge technology and human creativity. We aspire to democratize the sheer joy of music creation, ensuring that this exhilarating experience is open and available to everyone.

• Lyrics: Generated with GPT-4 and a fine tuned Llama2 on lyrics, Llama2 generating the name "I'm a Little Too Young to Be This Old" and GPT-4 generating lyrics from that starting place.
• MIDI Generation: Employed a Transformer XL Architecture trained on the Lakh midi dataset.
• Vocals: Generated with Kits.ai vocal style transfer models fine-tuned with custom vocal data from splice.com. Original vocal performance by Faith Manning.
• Sound Design & Sample Generation: All generated with Harmonai’s Dance Diffusion.
  • Drum Samples: Leveraged a 20,000 drum sample dataset from my Undergraduate Thesis Project.
  • Bird Samples: Recordings from the serene parks of Seattle, Washington.
  • Instrument Samples: Multiple models refined on instrument data from splice.com, creating diverse sonic textures.
  • Glitch Samples: A collection of 300 unique glitch samples sourced from splice.com, adding a contemporary edge.
  • Bass Growls:Utilizing 200 bass growls from splice.com, adding intensity in the drop.

AI Song Contest Entry

The piece encapsulates the vision of AI as a bridge in the vast landscape of music. It stands as a testament to AI's capability to serve beginners, empower those with disabilities, and inspire even the most seasoned musicians. This composition is not just a melody but a representation of the inclusive and innovative power of AI in music.

Explore Further

Overview

"Catch-A-Waveform" is a project primarily focused on audio generation and manipulation, with capabilities to generate audio from a single short example. This is a full Google Colab build of a modified Catch-A-Waveform repository from the Dadabots' Zack Zukowski.

Key Features

• Google Colaboratory Integration:
  • Utilize Cloud GPU resources for faster model training.
• Google Drive Mounting:
  • Conveniently access and manage project data stored in Google Drive.
• No Code Setup:
  • Clone the project's GitHub repository effortlessly, requiring minimal coding knowledge.
• Model Management:
  • Model Checkpointing: Easily save and restore model checkpoints, ensuring seamless training progress tracking.
• Enhanced Catch-A-Waveform Capabilities:
  • Training Run Modes:
    • Introduces new training run modes, including resume and transfer, for improved model training flexibility.
  • Generating Run Mode:
    • Introduces a new generating run mode called "infinite" for generating audio continuously.
  • Audio Sampling:
    • Implements a new audio sampling method called "--scale_crop" to efficiently fit maximum audio data into memory at different scales.
  • Model Building:
    • Allows the creation of models with skip connections in the 1D dilated convolution stacks, enhancing model complexity and performance.
  • Hyperparameters:
    • Adds new hyperparameters tailored for complex music generation, expanding the range from 250Hz to 44.1kHz.
  • Conditioning Options:
    • Offers options to condition audio generation on any input audio file using "--condition_file" and customize the frequency scale with "--condition_fs."
  • Lite Training:
    • Includes an experimental "lite training" mode with precision-reduced optimizers to conserve memory resources during training.

Explore Further:

• Original Project: Catch-A-Waveform
• Notebook: Google Colab Notebook

Overview

The YouTube Playlist Length Calculator is a Python script designed to calculate the total duration of all videos within a specified YouTube playlist. This can be particularly useful for assessing the total watch time of course material, tutorial series, or any collection of videos grouped into a YouTube playlist.

Key Points

• YouTube Data API v3: The script utilizes the YouTube Data API v3 to fetch data about videos and playlists.
• Python: The project is implemented in Python and makes use of various libraries and modules.
• API Key Management: The script has functionality to manage (save, update, and use) YouTube API keys.
• Progress Bar: Utilizes the tqdm library to display a progress bar during data retrieval and processing.
• Error Handling: Implements error handling for various issues like HTTP errors, connection issues, timeout errors, and more.
• Command-Line Usage: The script is designed to be used from the command line with various arguments for user convenience.

Explore Further

• GitHub Repository: YouTubePlaylistCalculator

Overview

The Socket Morse Chat App is a communication application developed in Python that allows users to send and receive messages in Morse code through a client-server architecture. The application translates text messages into Morse code and plays the corresponding audio (dots and dashes) to the users.

Key Points

• Client-Server Architecture: The application operates on a client-server model where multiple clients can connect to a server and communicate with each other.
• Socket Programming: Utilizes Python's socket programming to handle the communication between the server and clients.
• Multithreading: Implements threading to manage multiple clients and handle sending and receiving messages simultaneously.
• Morse Code Translation and Playback: Translates text messages into Morse code and plays the corresponding sounds using audio files.

Explore Further

• Github Repository: SocketMorseChatApp