Spring 2021, KAIST CS470 Course Project
A Replication of Neural Predictor for Neural Architecture Search, ECCV'20 with Tensorflow 2 [Report] [Presentation]
File Descriptions
Modules
- neural_predictor.py: implementaion of the neural predictor with its variants (MLP and CNN-based models).
- search_spaces.py: functions for accessing NAS-Bench-101, ProxylessNAS, and NAS-Bench-NLP search spaces.
- input_preprocessing.py: functions for preprocessing input with respect to each search space
- random_search.py: random search methods for NAS-Bench-101 and ProxylessNAS
Experiments
- Neural Predictor.ipynb: main experiments on neural predictor
- Two-stage Predictor.ipynb: experiments of two-stage neural predictor (with classifer)
- E1-NP-1.ipynb: neural predictor replication of Fig.4 in the original paper
- E1-NP-2.ipynb: neural predictor replication of Fig.4 in the original paper
- E1-NP-3.ipynb: neural predictor replication of Fig.4 in the original paper
- E1-NP-4.ipynb: neural predictor replication of Fig.4 in the original paper
- E1-Oracle.ipynb: Oracle replication of Fig.3 & 4 in the original paper
- E1-Random-1.ipynb: Random search replication of Fig.4 in the original paper
- E1-Random-2.ipynb: Random search replication of Fig.4 in the original paper
- Ablation Study-1.ipynb: N vs K ablation study
- Ablation Study-2.ipynb: different architecture ablation study
- Extended Study.ipynb: extend experiments on NAS-Bench-NLP
Directories
- figures: reproduced results from the original paper with a few additional figures
- nasbench: original NAS-Bench-101 search space
- nasbench_nlp: orignal NAS-Bench-NLP search space
- proxylessnas: MobileNetv2-based ProxylessNAS search space
- outputs: saved experimental results from the above experiments