Question about the last operation in the Discriminator

[thunguyenphuoc]

Hello there, As I was going through your code, I was a bit puzzled by the last operation if the discriminator: From 'model.py' file: m = tf.matmul(h, self.W['md']) m = tf.reshape(m, [-1, self.n_kernels, self.dim_per_kernel]) abs_dif = tf.reduce_sum(tf.abs(tf.expand_dims(m, 3) - tf.expand_dims(tf.transpose(m, [1, 2, 0]), 0)), 2) f = tf.reduce_sum(tf.exp(-abs_dif), 2) + self.b['md']

    h = tf.concat(1, [h, f])
    y = tf.matmul(h, self.W['h5']) + self.b['h5']
    return y

Could you explain to me this operation?

Thank you :)

[maxorange]

Hi, This operation is minibatch discrimination proposed in Improved Techniques for Training GANs. It is described in section 3.2.

In the original GAN training, the generator could concentrate on generating just a few examples. All output of the generator are very similar (low diversity).

Minibatch discrimination is a way to remedy this problem. In my experiment, minibatch discrimination helps prevent DCGAN from generating a few examples.

[thunguyenphuoc]

Hello,

Thank you for your fast reply. I guessed so as well. I try to implement 3D GAN for ModelNet10 based on your code but still cannot achieve the results like yours :) Did you find adding noise to the Discriminator also helps?

T

[maxorange]

Adding noise helps training somewhat. I increased the number of training examples by dilating voxel models. I used scipy.ndimage.morphology.binary_dilation for dilation.

Recently, I found that ELU+dropout also works well instead of using (leaky) ReLU+batchnorm.

[thunguyenphuoc]

Hello, Thanks for the tip for data augmentation and the insights in ELU+dropout. Right now, I only use rotation to augment the data.

I ran a couple of more experiments and found that: a/ Virtual batch norm makes it worse (both of training speed) and quality and b/Playing with the stddev of the noise of the Discriminator input helps with both the quality of the result and the diversity of the generated 3d shapes.

Anyhow, thank you for your tips and look forward to more results from you :)

T

[thunguyenphuoc]

One final question (I'm not sure if it is worth opening another issue or not): where did you get your training data from? Did you download the Shapenet data and convert them to binvox files yourself, or there is another publicly available voxelised Shapenet dataset somewhere?

Thank you for your help, T

[maxorange]

Thank you for the insights in virtual batch norm and the noise of the Discriminator input.

I got Shapenet data from their website (here https://www.shapenet.org/). To download the data, registration is required. They provides several subsets of the full Shapenet dataset. I downloaded ShapeNetCore subset which has 57,449 examples and converted them to binvox files myself.

When I converted, I filled the inside of voxel model with voxels. This is called solid voxelization. I trained GANs on solid voxelized dataset, not surface voxelized one.

If you need, feel free to open another issue.