We implemented a Neural Architecture Search technique inspired By DARTS for Semantic Segmentation of Images. Author: BabyGotHack
The algorithm is based on continuous relaxation and gradient descent in the architecture space. It is able to efficiently design high-performance convolutional architectures for Semantic Segmentation (on ISRO's UrbanDataset and Orchards3(junagarh)). Only a single GPU is required.
python >=3.5.5
Pytorch == 0.3.1
torchvision == 0.2.0
- operations.py : Contains all the inter cell operations.
- genotypes.py : Contains all intra cell operations.
- train_search.py : It searchs for genotype(architecture of each cell).
- Model_search.py : it gets imported in train_search.py to find alphas.
- utils.py : It contains all utility methods like mIoU, Accuracy, etc.
- train.py : When a optimal architecture is found out by network, user runs Train.py to train network on that architecture.
- model.py : when alphas are cell architecture are obtained by train_search.py, model.py imports genotype searched and constructs model based on that genotype on every epoch.
- Inference.ipynb : For inferencing the best mIoU model obtained by the network.
To run the application
./run
As of now we have train and tested on UrbanDataset of ISRO.
If you use any part of this code in your research, please cite or just remember the Name BabyGotHack:
@Members{
Shivam Kaushik(rs00188)
Ankush Malik(rs00190)
Rishab Sharma(rs00189)
Harshit Singhal(rs00235)
}