Evolving Search Space for Neural Architecture Search

Usage

Install all required dependencies in requirements.txt and replace all ..path/..to in the code to the absolute path to corresponding resources. The LUT (Latency Lookup Table) and meta files of ImageNet we used for searching are placed in resources folder.

Only slurm based distributed training is implemented.

Searching

All experiment files are located in experiment/NSE folder.

To search for NSENET-27 on ImageNet, run

sh search_NSE27.sh <number-of-nodes> <gpu-partition> 

To search for NSENET on ImageNet, run

sh search_NSE_second_space.sh <number-of-nodes> <gpu-partition> 

To search for NSENET-GPU on ImageNet, run

sh search_NSE_GPU.sh <number-of-nodes> <gpu-partition> 

Searched models

Final models (NSENET-27, NSENET, NSENET-GPU) are placed in utils/__init__.py.