This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision | ||
cs501r_f2017:lab7 [2017/10/24 22:36] wingated |
cs501r_f2017:lab7 [2021/06/30 23:42] (current) |
||
---|---|---|---|
Line 20: | Line 20: | ||
**NOTE:** this lab is complex. Please read through **the entire | **NOTE:** this lab is complex. Please read through **the entire | ||
spec** before diving in. | spec** before diving in. | ||
+ | |||
+ | Also note that training on this dataset will likely take some time. Please make sure you start early enough to run the training long enough! | ||
{{ :cs501r_f2017:faces_interpolate.png?direct&200|}} | {{ :cs501r_f2017:faces_interpolate.png?direct&200|}} | ||
Line 51: | Line 53: | ||
One of the advantages of the "Improved WGAN Training" algorithm is that many different kinds of topologies can be used. For this lab, I recommend one of three options: | One of the advantages of the "Improved WGAN Training" algorithm is that many different kinds of topologies can be used. For this lab, I recommend one of three options: | ||
- | * The [[https://arxiv.org/pdf/1511.06434.pdf|DCGAN architecture], see Fig. 1. | + | * The [[https://arxiv.org/pdf/1511.06434.pdf|DCGAN architecture]], see Fig. 1. |
* A [[https://arxiv.org/pdf/1512.03385|ResNet]]. | * A [[https://arxiv.org/pdf/1512.03385|ResNet]]. | ||
- | + | * Our reference implementation used 5 layers: | |
- | Our reference implementation used 5 layers: | + | * A fully connected layer |
- | * A fully connected layer | + | * 4 convolution transposed layers, followed by a relu and batch norm layers (except for the final layer) |
- | * 4 convolution transposed layers, followed by a relu and batch norm layers (except for the final layer) | + | * Followed by a tanh |
- | * A final tanH nonlinearity | + | |
==Part 1: Implement a discriminator network== | ==Part 1: Implement a discriminator network== | ||
Line 63: | Line 64: | ||
Again, you are encouraged to use either a DCGAN-like architecture, or a ResNet. | Again, you are encouraged to use either a DCGAN-like architecture, or a ResNet. | ||
- | Our reference implementation used 4 convolution layers, each followed by a leaky relu (leak 0.2) and batch norm layer, with a sigmoid as the final nonlinearity. | + | Our reference implementation used 4 convolution layers, each followed by a leaky relu (leak 0.2) and batch norm layer (except no batch norm on the first layer). |
+ | |||
+ | Note that the discriminator simply outputs a single scalar value. This value should unconstrained (ie, can be positive or negative), so you should **not** use a relu/sigmoid on the output of your network. | ||
==Part 2: Implement the Improved Wasserstein GAN training algorithm== | ==Part 2: Implement the Improved Wasserstein GAN training algorithm== | ||
- | Gradients: | + | The implementation of the improved Wasserstein GAN training algorithm (hereafter called "WGAN-GP") is fairly straightforward, but involves a few new details about tensorflow: |
- | tf.gradients | + | * **Gradient norm penalty.** First of all, you must compute the gradient of the output of the discriminator with respect to x-hat. To do this, you should use the ''tf.gradients'' function. |
+ | * **Reuse of variables.** Remember that because the discriminator is being called multiple times, you must ensure that you do not create new copies of the variables. Note that ''scope'' objects have a ''reuse_variables()'' function. | ||
+ | * **Trainable variables.** In the algorithm, two different Adam optimizers are created, one for the generator, and one for the discriminator. You must make sure that each optimizer is only training the proper subset of variables! There are multiple ways to accomplish this. For example, you could use scopes, or construct the set of trainable variables by examining their names and seeing if they start with "d_" or "g_": | ||
+ | <code python> | ||
+ | t_vars = tf.trainable_variables() | ||
+ | self.d_vars = [var for var in t_vars if 'd_' in var.name] | ||
+ | self.g_vars = [var for var in t_vars if 'g_' in var.name] | ||
+ | </code> | ||
- | Reuse of variables | + | I didn't try to optimize the hyperparameters; these are the values that I used: |
- | scope.reuse_variables | + | <code python> |
+ | beta1 = 0.5 # 0 | ||
+ | beta2 = 0.999 # 0.9 | ||
+ | lambda = 10 | ||
+ | ncritic = 1 # 5 | ||
+ | alpha = 0.0002 # 0.0001 | ||
+ | m = 64 | ||
- | Trainable variables | + | batch_norm decay=0.9 |
+ | batch_norm epsilon=1e-5 | ||
+ | </code> | ||
- | Two Adam optimizers | + | Changing to number of critic steps from 5 to 1 didn't seem to matter; changing the alpha parameters to 0.0001 didn't seem to matter; but changing beta1 and beta2 to the values suggested in the paper (0.0 and 0.9, respectively) seemed to make things a lot worse. |
==Part 3: Generating the final face images== | ==Part 3: Generating the final face images== | ||
- | Your final deliverable is two images. The first should be a set of randomly generated faces. This is as simple as generating random ''z'' variables, and then running them through your discriminator. | + | Your final deliverable is two images. The first should be a set of randomly generated faces. This is as simple as generating random ''z'' variables, and then running them through your generator. |
For the second image, you must pick two random ''z'' values, then linearly interpolate between them (using about 8-10 steps). Plot the face corresponding to each interpolated ''z'' value. | For the second image, you must pick two random ''z'' values, then linearly interpolate between them (using about 8-10 steps). Plot the face corresponding to each interpolated ''z'' value. | ||
Line 91: | Line 109: | ||
The reference implementation was trained for 8 hours on a GTX 1070. It ran for 25 epochs (ie, scan through all 200,000 images), with batches of size 64 (3125 batches / epoch). | The reference implementation was trained for 8 hours on a GTX 1070. It ran for 25 epochs (ie, scan through all 200,000 images), with batches of size 64 (3125 batches / epoch). | ||
+ | |||
+ | Although, it might work with far fewer (ie, 2) epochs... |