μNet

This directory contains resources related to the µNet line of research.

Below are listed the resources related to each publication in the µNet liner of research. For each publication we provide a colab containing all the code necessary to reproduce the reported experiments. We recommend using the latest published colab to experiment with new ideas and extensions. For questions, requests or feedback feel free to reach out to: agesmundo@google.com.

muNet

1. Publication: "muNet: Evolving Pretrained Deep Neural Networks into Scalable Auto-tuning Multitask Systems" (2022).

2. Colab Notebook that defines the muNet method and allows to reproduce the experiments reported in the publication: muNet.ipynb

3. Analysis Colab Notebook that allows to reproduce plots and analysis reported in the paper: analysis.ipynb

µ2Net

1. Publication: "An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems" (2022).

2. Colab Notebook that defines the μ2Net method and allows to reproduce the experiments reported in the publication: mu2Net.ipynb

3. Analysis Colab Notebook that allows to reproduce plots and analysis reported in the paper: analysis.ipynb

4. The checkpoint of the multitask system produced by the continual-learning experiment described in the μ2Net paper can be loaded and extended by using the mu2Net.ipynb Colab Notebook: Download μ2Net checkpoint (65 GB).

The header of each Colab Notebook reports instructions and license.

µ2Net+

1. Publication: "A Continual Development Methodology for Large-scale Multitask Dynamic ML Systems" (2022)

2. To request publicaion of source code or checkpoints, please reach out to agesmundo@google.com. The method extension described in the µ2Net+ publication are included in the source code of the Vit Agent: mu3Net.ipynb

µ3Net: Multiagent Framework

1. Publication: "A Multiagent Framework for the Asynchronous and Collaborative Extension of Multitask ML Systems" (2022)

2. Colab Notebook that defines the ViT Agent and allows to reproduce the experiments reported in the publication: mu3Net.ipynb

µ4Net: the Multipath Multiagent Multitask Mutant Network

1. Publication: "Multipath Agents for Modular Multitask ML Systems" (2023)

2. Colab Notebook with additional MultiViT Agent allowing to reproduce the experiments reported in the publication: mu4Net.ipynb

3. Download μ4Net checkpoint (61 GB). Containing the µ3Net singlepath models and the multipath model for the imagenet2012 task generated by best experiment reported in the publication.