d0352276/siamese-nas

This project provides a predictor-based NAS method, and it is namely Siamese-Predictor. The siamese-predictor is constructed with the proposed Estimation Code, which is the prior knowledge about the training procedure.

Siamese-NAS: Using Trained Samples Efficiently to Find Lightweight Neural Architecture by Prior Knowledge

This project provides a predictor-based NAS method, and it is namely Siamese-Predictor. The siamese-predictor is constructed with the proposed Estimation Code, which is the prior knowledge about the training procedure. In our experiments, the proposed siamese-predictor surpasses the current SOTA predictor (BRP-NAS) on NASBench-201. We also propose the search space Tiny-NanoBench for lightweight CNN architecture. This well-designed search space is easier to find better architecture with few FLOPs than NASBench-201.

Paper Link: https://arxiv.org/abs/2210.00546

Requirements

• Python 3
• TensorFlow 2
• OpenCV
• Numpy
• CIFAR-10
• NASBench-201

Training and Evaluation for Paper Results

There are various predictors and search spaces in our paper, and if you want to train the specific predictor, you just change line 140 of file "train_eval.py". For example, you can train the Siamese-Predictor +NSAM +BTS on NASBench-201 (CIFAR-100).

if(__name__=="__main__"):
    model_type="siams +NSAM +BTS"
    data_type="nasbench201_cifar100" 
    ......

If you want to train the naive predictor on Tiny-NanoBench, you just rewrite line 140 to this.

if(__name__=="__main__"):
    model_type="predictor"
    data_type="tiny_nanobench" 
    ......

In default setting, every traing and evaluation procedure will repeat 100 runs, if you want to change it, please change the arg "params_dict["repeats"]" at line 134.

Complete Search Spaces

Attached to this project is already most of the searc space, it contains NASBench-201 (CIFAR-10), NASBench-201 (CIFAR-100) and Tiny-NanoBench. More search spaces are coming soon! Due to the large number of files, you have to unzip it yourself.

More Info for Tiny-NanoBench

We record the relationship between FLOPs and accuracy for NASBench-201. We can find that it consists of several groups at the FLOPs axis. Based on this observation, we set a constraint of about 35 MFLOPs to focus on the smallest FLOPs group (blue group in part (a) of Fig. 4), we call this subset Tiny-NASBench-201. We believe that Tiny-NASBench-201 is not the optimal search space at the tiny FLOPs ( 35M) level, so we proposed the Tiny-NanoBench trained on CIFAR-10.

Results

Fig. 6 shows the experiments on NASBench-201, which is fixed N or K and evaluates these predictors. Part (a) of Fig. 6 is fixed N to 30, then gradually increases K to draw the entire curve. Fig. 7 shows that the siamese-predictor searching on Tiny-NanoBench can find the architecture of higher accuracy than Tiny-NASBench-201 when N is large.

TODOs

• More intuitive API.
• ...