this repo is the commented version (only the cnn folder is annotated, and the rnn folder follows similar logic) of code from: here, the implementation for the paper DARTS: Differentiable Architecture Search.
For citations and details on training etc., please refer to here.
To quickly understand the code (take the cnn folder as an example), we can simply divide them into two parts:
- code with '_search' suffix; 2. w/o '_search' suffix.
For part 1 (with suffix), they are used for searching the optimal architecture (alpha), and for part 2 (w/o suffix), they will use the optimal architecture searched by part 1, and ONLY learn the weights under such architecture.