Unofficial PyTorch implementation of: ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware.
| Paper | Here | |
|---|---|---|
| Model Parameters (M) | 5.7 | 5.3 |
| Top-1 Test Error (%) | 2.08 | 2.72 |
- python 3
- pyyaml
- torch
- torchvision
- numpy
- graphviz
Optional packages:
- tensorboardX
- adabound
python search.py -n <run_name> -c <config_path> -d <gpu_ids>
By default, checkpoints and genotypes are saved to 'chkpt/'
python augment.py -n <run_name> -c <config_path> -d <gpu_ids> -g <genotype_path>
Run configs are specified in config/default.yaml
- search: arch search run config
- augment: augment run config
- model: model specs.
- genotypes: candidate operations in arch search
- log: logging config