Inference Pipeline Adapter (IPA) Extension Project

A CSCE 585 - Machine Learning Systems Project

Team Members

• Misagh Soltani
• Sabah Anis
• Xeerak Muhammad

Table of Contents

• Inference Pipeline Adapter (IPA) Extension Project
  • Team Members
  • Table of Contents
  • Installation
  • Usage
  • Final Deliverables

Installation

1. Follow the instructions in the original ipa repository (https://github.com/reconfigurable-ml-pipeline/ipa/blob/main/infrastructure/automated.md)

2. For measuring energy consumption, install perf tool using the following commands:

sudo apt update
sudo apt install linux-tools-common linux-tools-generic linux-tools-$(uname -r)
perf --version

Usage

1. Run Experiments:

   • Use the experiments/runner/run_revision_perf.sh script to run the experiments.
   • This script internally calls all experiments in experiments/runner/run-revision.sh and uses perf to measure the energy consumption of each experiment.
   • Use the experiments/runner/run_q_learning_perf.sh script to run the Gurobi vs Q-learning experiments.

2. Energy Measurement Logs:

   • The energy consumption logs are saved in the energy_logs directory.
   • The energy consumption logs from the gurobi and Q-learning experiments are save in the energy_comparison directory.

3. Generate Plots:

   • Use the experiments/runner/run_revision_perf_plot.ipynb Jupyter Notebook to generate plots for the energy benchmarks of the workloads.
   • Use the experiments/runner/graph_q_learning.ipynb Jupyter Notebook to generate plots for the energy benchmarks of the Gurobi vs Q-learning workloads.
   • The generated plots are saved in the plots directory.

4. Segment Anything:

   • Set up a conda environment using conda create -n sam python=3.8

   • pip install git+https://github.com/facebookresearch/segment-anything.git

   • cd segment-anything; pip install -e .

   • Download sam weights from and place them in segment-anything/model_checkpoints/: -vit_b: ViT-B SAM model. -vit_h: ViT-H SAM model. -vit_l: ViT-L SAM model.

   • To generate energy consumption results use segment-anything/sam_energy.sh

   • The energy logs will be in segment-anything/sam_energy_log/

   • The segmentation masks will be saved in segment-anything/segmentation_benchmark/{model_type}_results/

   • To graph the energy consumption results you can run the segment_anything/graph_sam_energy_consumption.ipynb file

   • To calculate IoU run segment-anything/metrics.sh

5. How to Run Q-Learning:

To run Q-learning, follow the same instructions as for the Gurobi optimizer. You only need to update the configurations used in the original code. Here are the steps:

1. Locate the YAML configuration file in the folder ipa-ext/data/configs.
2. Update the optimizer parameter:
   • Change the value of the optimizer parameter from gurobi to q_learning.
3. Adjust the function arguments:
   • Modify any function arguments that currently have the value gurobi to q_learning.
4. Moreover, we have provided config files for Q-learning in the ipa-ext/data/configs directory. You can use these files to run Q-learning experiments. The config files with the number 4, 9, 14, and 19 are the Q-learning config files. These config files are used for running the project in the updated scripts and entry point and they can be used in the same way as the other optimizers.

By making these changes, you can use Q-learning as the optimizer in the pipeline.

Final Deliverables

Please find the Final report here. Please find the Final presentation slides here. Please find the Final presentation video here.