Towards Generalized Neural Architecture Performance Estimators

This repository accompanies the paper

GENNAPE: Towards Generalized Neural Architecture Performance Estimators
Keith G. Mills, Fred X. Han, Jialin Zhang, Fabian Chudak, Ali Safari Mamaghani, Mohammad Salameh, Wei Lu, Shangling Jui and Di Niu
AAAI-23 Oral Presentation

[Poster][Video][Slides]

Specifically, we provide the following:

• Computation Graph (CG) data caches for all datasets used in the paper.
• Code API for generating, loading and visualizing CGs.
• Code for generating additional NAS-Benchmark and Once-for-All CGs
• Code for generating HiAML, Inception and Two-Path CGs
• k-GNN predictor demo code.

Main Dependencies

python==3.7
torch==1.8.1
torch_geometric==1.7.0
tensorflow==1.15.0
graphviz

See requirements.txt for more specifics

Setting Up The Repository

1. Download and unpack CG_data.zip from the public google drive folder
2. Copy the cache and data folders to the top-level of this directory.
   • cache contains the CG files for the NAS-Bench (101, 201 and 301) as well as OFA (PN, MBv3 and ResNet-50) families.
   • data contains the CG files for the HiAML, Inception and Two-Path families.

Running the Demo k-GNN predictor code

Use the top-level script run_cg_kgnn_predictor.py; notable parameters include:

• -model_name: String for the name of an experiment.
  • Log file will save under logs/gpi_acc_predictor_{model_name}_seed{seed}.txt
  • Model checkpoints will be saved under /saved_models/gpi_acc_predictor_{model_name}_seed{seed}_{best, last}.pt
• -family_train: Training family; default is NB-101
• -family_test: Testing/fine-tuning family. Default is all test families.
  • The flag #50 denotes that 50 samples will be taken-out to fine-tune a model for epochs determined by -fine_tune_epochs
  • If this flag is not specified, e.g., -family_test hiaml is used, no fine-tuning will be performed.
• -e_chk: Allows user to load a specified model checkpoint.
• -seed: Random seed.
• -gnn_type: Default is 'GraphConv' (k-GNN), but can be changed to others, e.g., GCNConv, GINConv, etc.

Visualization of CGs

• API function is gviz_visualize, found in /model_src/comp_graph/tf_comp_graph.py
• We provide a simple script, visualize_cgs.py, to print out N random architectures from a specified family to a target directory.
• The loaded figure shows the node with node features, e.g., kernel sizes, as well as input/output tensor shapes.

Generation of new CGs from TensorFlow protobuf .pb families

• See make_cg.py for an example.
• We provide sample .pb files for EfficientNet-b0 and ResNet18 on the public google drive folder. You can use this script to make CGs and print pictures of them, then compare to what the .pb of the model looks like in Netron.

Generating NAS-Bench and Once-for-All CGs beyond what is provided in our caches

• Our caches consist of architectures we randomly sampled from each search space (except NB-201).
• Some users may want to make CGs of other architectures in these search spaces
• We provide the necessary code to do so in this repo, please see gen_demo_caches.py for details on each family.

Generating HiAML, Inception and Two-Path CGs

• See make_new_family_cgs.py

Miscellaneous:

• The CG API is defined in /model_src/comp_graph/tf_comp_graph.py
• To access CGs directly, /model_src/predictor/gpi_family_data_manager.py handles loading CGs from a cache.

Bibtex

If you find our data or CG API useful, we kindly ask that you cite our paper:

@inproceedings{mills2023gennape,
  title = {GENNAPE: Towards Generalized Neural Architecture Performance Estimators},
  author = {Mills, Keith G. and Han, Fred X. and Zhang, Jialin and Chudak, Fabian and Mamaghani, Ali Safari and Salameh, Mohammad and Lu, Wei and Jui, Shangling and Niu, Di},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2023}
}