GAN design
Created by: zchen0211
GAN is an important concept to support in our refactoring. Based on the current support of our framework, it is better to implement GANs with following concepts:
- Dependency analysis and sub-graph running;
- Gradient with respect to some certain variables;
- Multiple loss functions design.
Operators and functions required to build GAN:
| Important building blocks | People in Charge | Required |
|---|---|---|
| convolution 2d (done) | Chengduo | Y |
| cudnn conv 2d (missing) | Chengduo | N |
| deconv 2d (missing) | Zhuoyuan, Zhihong | Y |
| cudnn deconv 2d (missing) | Zhuoyuan, Zhihong | N |
| batch norm (missing) | Zhuoyuan, Jiayi | Y |
| cudnn batch norm (missing) | Zhuoyuan, Jiayi | N |
| max-pooling (done) | ? | Y |
| fc (done) | ? | Y |
| softmax loss (done) | ? | Y |
| reshape op (done) | ? | Y |
| Dependency Engine (done) | Jiayi | Y * |
| Python API (done) | Longfei, Jiayi | Y * |
| Executor (done) | Tony | Y * |
| Multi optimizer (woking) | Longfei | Y * |
| Optimizer with any para | ? | Y * |
| Concat op (done) | ? | N (Cond) |
| Repmat op (done) |