NAS_PerformanceEstimation

Generalized Latency Performance Estimation for Once-For-All Neural Architecture Search

This implementation is adopted from the source code of Once For All (Cai et al. 2019) and CompOFA (Sahni et al. 2021)

We created a jupyter_notebook: latency_prediction_demo.ipynb which run through all the required steps to show a working version of the code base.

Setup

1. conda env create -f environment.yml
2. conda activate latency_predict_env
3. jupyter notebook
4. Open tutorial/latency_prediction_demo.ipynb
5. Run all

Description of relevant files

1. latency_prediction_demo.ipynb
   - Jupyter notebook Demo with steps to run Dataset Creation, Model Evaluation, and OFA NAS
   - Lists all necessary dependencies for project

2. latency_predictor_driver.py
   - Provides code to create latency datasets. Uses imports from evolution_finder to random sample OFA, CompOFA
   - Code for inference time analysis of latency predictor and measurement

3. latency_NAS_runner.py
   - Similar code to demo jupyter notebook. Prepares and run code to perform NAS

4. evolution_finder.py (modified)
   - this file was modified to include create_latency_dataset()

5. checkpoints/

   • latency_prediction_model/
           - contains all trained latency predictors organized in subfolders of hardware and search spaces

6. latency_predictor/\

   • datasets/
           - contains all datasets created for this project, organized into device subfolders. Also includes GPU generalization datasets
   • model_results/
           - includes images and text files of all experimental results organized in subfolders
   • generalized_dataset_combine.py
           - helper code to add hardware parameters to datasets
   • Iterations.txt
           - Log file from search
   • latency_encoding.py
           - helper code to perform one hot encoding of child architectures
   • latency_finetune.py
           - code for training fine tuned models, case studies, and plotting loss curves
   • latency_predictor.py
           - code for creating latency predictors, data_preprocessing, training using RayTune, testing
   • latency_predictor_generalized.py
           - code for creating generalized latency predictors
   • lookup_table_calculation.py
           - helper code to calculate time taken to create lookup table
   • other images are results from various experiments