iicd/autodl-projects
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

update CVPR-2019-GDAS re-train NASNet-search-space searched models

Browse Source
This commit is contained in:
D-X-Y
2020-03-06 19:29:07 +11:00
parent 8b6df42f1f
commit 9a83814a46
17 changed files with 278 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
docs/CVPR-2019-GDAS.md
View File

@@ -41,7 +41,16 @@ Please use the following scripts to use GDAS to search as in the original paper:
```
CUDA_VISIBLE_DEVICES=0 bash ./scripts-search/GDAS-search-NASNet-space.sh cifar10 1 -1
```
If you want to train the searched architecture found by the above scripts, you need to add the config of that architecture (will be printed in log) in [genotypes.py](https://github.com/D-X-Y/AutoDL-Projects/blob/master/lib/nas_infer_model/DXYs/genotypes.py).
**After searching***, if you want to re-train the searched architecture found by the above script, you can use the following script:
```
CUDA_VISIBLE_DEVICES=0 bash ./scripts/retrain-searched-net.sh cifar10 gdas-searched \
output/search-cell-darts/GDAS-cifar10-BN1/checkpoint/seed-945-basic.pth 96 -1
```
Note that `gdas-searched` is a string to indicate the name of the saved dir and `output/search-cell-darts/GDAS-cifar10-BN1/checkpoint/seed-945-basic.pth` is the file path that the searching algorithm generated.
The above script does not apply heavy augmentation to train the model, so the accuracy will be lower than the original paper.
If you want to change the default hyper-parameter for re-training, please have a look at `./scripts/retrain-searched-net.sh` and `configs/archs/NAS-*-none.config`.
### Searching on a small search space (NAS-Bench-201)
The GDAS searching codes on a small search space: