API of GAN #4580
API of GAN #4580
Conversation
doc/design/gan_api.md Outdated
| GAN implementation, just a demo. | ||
| ''' | ||
| # pd for short, should be more concise. | ||
| from paddle.v2 as pd |
There was a problem hiding this comment.
use
```python
```
that will give a python syntax highlight.
doc/design/gan_api.md Outdated
| batch_z = np.random.uniform(-1., 1., [batch_size, z_dim]) | ||
| | ||
| if batch_id % 2 == 0: | ||
| sess.run(d_optim, |
There was a problem hiding this comment.
use our interface, pd.eval?
| | ||
| # Conditional-GAN should be a class. | ||
| ### Class member function: the initializer. | ||
| class DCGAN(object): |
There was a problem hiding this comment.
no need to be a class, simply 3 functions should be more simple.
There was a problem hiding this comment.
We may have a discussion later. I refer to several GAN models on github, pretty much all looks like this.
doc/design/gan_api.md Outdated
| if not self.y_dim: | ||
| z = pd.concat(1, [z, y]) | ||
| | ||
| G_h0 = pd.fc(z, self.G_w0, self.G_b0) |
There was a problem hiding this comment.
fc layer support a param_name argument and generate a unique name , just use it like
G_h0 = pd.fc(z, size=y_dim, param_name="G_fc0")call generator twice will not generate different parameters.
and each fc needn't declare parameters somewhere, so is batchnorm.
doc/design/gan_api.md Outdated
| @@ -0,0 +1,193 @@ | |||
| # Design for GAN | |||
| | |||
| GAN (General Adversarial Net) is an important model for unsupervised learning and widely used in many areas. | |||
There was a problem hiding this comment.
maybe add a link to the official definition
doc/design/gan_api.md Outdated
| | ||
| It contains several important machine learning concepts, including building and running subgraphs, dependency tracing, different optimizers in one executor and so forth. | ||
| | ||
| In our GAN design, we wrap it as a user-friendly easily customized python API to design different models. We take the conditional DC-GAN as an example due to its good performance on image generation. |
doc/design/gan_api.md Outdated
| | ||
| GAN (General Adversarial Net) is an important model for unsupervised learning and widely used in many areas. | ||
| | ||
| It contains several important machine learning concepts, including building and running subgraphs, dependency tracing, different optimizers in one executor and so forth. |
There was a problem hiding this comment.
I think contain is not good, maybe It will use ...?
doc/design/gan_api.md Outdated
| | max-pooling (done) | ? | Y | | ||
| | fc (done) | ? | Y | | ||
| | softmax loss (done) | ? | Y | | ||
| | reshape op (done) | ? | Y | |
There was a problem hiding this comment.
I think it is better to create an issue to place this information.
use a - [ ] zzz to create checkboxes to follow up the process.
doc/design/gan_api.md Outdated
| build the whole GAN model, define training loss for both generator and discrimator. | ||
| | ||
| ## Discussion on Engine Functions required to build GAN | ||
| - Trace the ternsor and variable dependency in the engine executor. (Very critical, otherwise GAN can'be be trained correctly) |
5 participants
Based on some popular models in TensorFlow and Torch, we design the logic of GAN API.
It is too complex to realize in an op, so I use python API.